Ten Years To The Singularity

Ten Years
To The Singularity
If We Really Really Try
and other Essays
on AGI and its Implications
Ben Goertzel
Copyright 2014 Ben Goertzel

All rights reserved.
ISBN:
ISBN-14:
Table&of&Contents&
Preface .............................................................................................................. 7!
Next Steps For And Beyond Humanity .................................................. 10!
The Top Priority for Mankind ........................................................................ 15!
Why AGI ? ....................................................................................................... 20!
Why the Time is Ripe for AGI ....................................................................... 24!
Why So Little Work on AGI? ......................................................................... 34!
Ten Years to the Singularity If We Really Really Try ................................. 40!
How Might a Positive Singularity Get Launched In 10 Years From Now? .... 43!
Does it Take a Manhattan Project? .............................................................. 48!
Keeping it Positive ........................................................................................ 50!
Ten Years to the Singularity? ....................................................................... 51!
How Long Till AGI? ........................................................................................ 54!
Two Earlier Surveys of Expert Opinion on the Timing of Human-Level AI ... 54!
The Questions Asked of the Experts ............................................................ 56!
What the Experts Said About The Timing of Human-Level AI ...................... 57!
What Kind of Human-Level AI Will Come First? ........................................... 60!
What Kind of Technical Approach Will First Achieve Human-Level AI? ....... 61!
Impacts of AGI .............................................................................................. 62!
Conclusions of the Survey ............................................................................ 63!
What Is General Intelligence? ....................................................................... 65!
Defining General Intelligence in Theory and Practice ................................... 70!
Why an Intelligence Explosion is Probable ................................................ 76!
Why an Intelligence Explosion is Probable ................................................... 79!
The One Clear Prerequisite for an Intelligence Explosion ............................ 79!
Distinguishing the Explosion from the Build-Up ............................................ 80!
What Constitutes an Explosion? ................................................................ 81!
Defining Intelligence (Or Not) ....................................................................... 83!
Key Properties of the Intelligence Explosion ................................................ 83!
Objection 1: Economic Growth Rate and Investment Availability ................. 86!
Objection 2: Inherent Slowness of Experiments and Environmental
Interaction .................................................................................................. 88!
Objection 3: Software Complexity ................................................................. 91!
Objection 4: Hardware Requirements........................................................... 92!
Objection 5: Bandwidth ................................................................................. 93!
Objection 6: Lightspeed Lags ....................................................................... 95!
Objection 7: Human-Level Intelligence May Require Quantum (Or More

Exotic) Computing ......................................................................................96!
The Path from AGI to Intelligence Explosion Seems Clear ...........................97!
Ten Common Objections to AGI And Why They Dont Scare Me ..........99!
Approaches to AGI .......................................................................................108!
AGI Via Narrow AI? .....................................................................................110!
AGI Via Brain Emulation? ............................................................................112!
AGI Via Artificial Evolution ...........................................................................117!
Good Old-Fashioned AI ...............................................................................120!
GOFAI Grows Up: SOAR ............................................................................124!
Millions and Millions of Rules: Cyc ..............................................................127!
The Modern, Direct Approach to AGI ..........................................................133!
Unified Versus Integrative ...........................................................................138!
Emergent Versus Explicit ............................................................................142!
Theory Versus Experimentation in AGI Development .................................146!
Can Digital Computers Really Be Intelligent? ..............................................149!
Will AGIs Be Conscious? .............................................................................152!
Deep Blue Reflections ..................................................................................160!
Today Jeopardy! Tomorrow the World?....................................................166!
What Does Watson Mean for AI? ................................................................168!
Under Watsons Hood .................................................................................168!
Wolfram on Watson .....................................................................................172!
Kurzweil on Watson .....................................................................................173!
Today Jeopardy!, Tomorrow the World? .....................................................175!
Chatbots Versus Cognition Bots ................................................................177!
2010 Loebner Winner: Suzette! ...................................................................179!
ELIZA...........................................................................................................183!
AliceBot .......................................................................................................185!
MegaHal ......................................................................................................189!
Building a Better Ramona............................................................................198!
Toward Intelligent Dialogue Systems: The AGI Preschool Project ..............201!
Is a Statistical Turing Bot Possible? ............................................................209!
The AGI Roadmap Workshop ......................................................................213!
Laird and Wrays Requirements for Human-Level AGI Systems .................218!
Laird and Wrays Requirements for Environments for Human-Level AGI
Systems ....................................................................................................218!
The Breadth of Human Competencies ........................................................222!
Scenarios for Assessing AGI .......................................................................225!
Other AGI Scenarios................................................................................... 233!

From Scenarios to Tasks and Metrics ........................................................ 235!
Mind Uploading ............................................................................................ 242!
Mind Uploading Via Brain Emulation .......................................................... 245!
Is Your Upload You?................................................................................... 247!
Mind Uploading and AGI............................................................................. 248!
AGI Against Aging ....................................................................................... 250!
The Biology of Cell Senescence ................................................................. 253!
Long Life Through Calorie Restriction? ...................................................... 256!
The Paleo Lifestyle ..................................................................................... 257!
Aubrey de Grey and the Seven Pillars of Senescence ............................... 258!
Radical Life Extension: Whats the Real Bottleneck? ................................. 262!
AI Against Aging: Tales from the Research Front ...................................... 265!
Unraveling the Mystery of the Methuselah Flies ......................................... 274!
AI That Reads Biological Research Articles ............................................... 276!
The Holistic Biobase ................................................................................... 278!
AGI and Magical Thinking ........................................................................... 285!
Plenty More Room at the Bottom ............................................................... 294!
Physics at the Femto Scale ........................................................................ 297!
Degenerate Matter as a Possible Substrate for Femtotech ........................ 300!
Bolonkin's Fantastic Femtotech Designs .................................................... 301!
Might Dynamic Stabilization Work on Degenerate Matter? ........................ 305!
Gell-Mann Gives Femtotech A Definite Maybe ........................................... 307!
Strange Matter and Other Strange Matters ................................................ 309!
The Future of Femtotech ............................................................................ 312!
The Singularity Institute's Scary Idea (and Why I Don't Buy It) .............. 315!
SIAI's Scary Idea (Which I Don't Agree With) ............................................. 318!
What is the Argument for the Scary Idea? .................................................. 320!
Is Provably Safe or "Friendly" AGI A Feasible Idea? .................................. 322!
Practical Implications .................................................................................. 324!
Conclusion .................................................................................................. 326!
Does Humanity Need an AI Nanny? ........................................................... 329!
The AI Nanny in Detail ................................................................................ 330!
The Argument for an AI Nanny ................................................................... 332!
Complaints and Responses ........................................................................ 334!
AGI, Consciousness, Spirituality, Life, the Universe, and Everything ... 341!
Physical Monism ......................................................................................... 345!
Informational Monism ................................................................................. 346!
Quantum Informational Monism ..................................................................349!

Experiential Vitalism ....................................................................................351!
Cognitive Vitalism ........................................................................................354!
Panpsychism ...............................................................................................356!
Spiritual Monism ..........................................................................................359!
The Implicate Order .....................................................................................367!
Conclusion ...................................................................................................370!
Preface
Ive been thinking about Artificial General Intelligence (AGI) and
the future of human and transhuman intelligence since I was 4 or
5 years old and first encountered such ideas in SF novels. As
time went on my thinking veered from the purely speculative and
futuristic to the practical and engineering-oriented; in the last two
decades the pursuit of real-world AGI systems in the here and
now has been the central theme of my career (most recently
largely via the OpenCog project). But Ive never taken my eye
off the broader potential AGI holds, and its connection to various
other aspects of current and future life i.e. to almost anything
you can think of!
I thoroughly believe that the transition from human intelligence to
transhuman artificial intelligence is going to be far huger than
any transformation in the history of humanity more akin to the
transition from bacteria to humans, or from rocks to bacteria, if
one wants to draw historical analogies.
In the shorter term,
there will be plenty of dramatic implications for human society
and psychology. And in the longer term, the implications will go
way, way beyond anything we can imagine. Both the shorterterm and longer-term implications are amazing and exciting to
think about.
During the period 2009-2011, when I had a bit more time for
writing than I do now, I wrote a series of essays for H+ Magazine
the online zine I co-edit -- on the future of AGI, the Singularity
and related topics. While these dont exactly flow together like
chapters of a single-focused book, they do preoccupy with a
common set of themes and present a coherent point of view. So
I have gathered them here together in a single volume, along
with a handful of non-H+-Mag essays from the same period a
couple that were published elsewhere, and a couple that
languished on my hard drive unpublished till now.
Ten Years to the Singularity If We Really Really Try
Worth noting, perhaps, is that one thing this book DOESNT

contain is a detailed overview of my own work on AGI, which Ive
described in a variety of technical works before, and which is
currently centered on the OpenCog open source AGI platform.
Im currently (mid-2014) cooking a non-technical book called
Faster Than You Think which will cover this ground. Various
chapters here in Ten Years do touch on aspects of my own
work, but only here and there in passing not in a systematic
way. What this book DOES give, though, is a broad overview of
the way I think about AGI in terms of its relationship with other
futurist technologies, and the amazing future that AGI and these
technologies together are likely to bring to future humans and the
future successors of humanity.
None of the essays collected here are technical. Some are very
lightweight, at the level of a newspaper article. Some are more
in-depth than that, at the level of a fairly intense pop-sci book or
an article in a popular science magazine.
It should all be
readable by anyone with a couple years of university education
in any field.
Some of the material here is dated in some respects relative
to mid-2014 when Im writing this preface -- but I havent made
any effort to update it. By the time youre reading this, 2014 may
be long gone as well. The essays collected here make some
points that transcend their time of writing; and they also may
have some value as a record of how their topics were being
thought about during the particular period 2009-2011. AGI is
advancing rapidly; and thinking about the Singularity and the
future of humanity and transhumanity is also advancing rapidly.
Humanitys thoughts about the (conceptually or calendrically) faroff future, written at a certain time, are generally at least as
informative about the state of humanity at the time of writing, as
they are about the far-off future.
Friedrich Nietzsche, in a draft preface for a book he planned to
write with the title The Will to Power (different from the collection
Preface
of notes that was released posthumously under that title),

described the book as
A book for thinking, nothing else; it belongs to those for whom
thinking is a delight, nothing else.
The thoughts in this book are offered in a similar spirit. There is
nothing definitive here.
Its all necessarily quite preliminary,
because we are still in the early stages of creating AGI; and even
though we may be fairly close to a Technological Singularity in
terms of calendar years, we have loads of technological,
scientific and conceptual progress to pass through before we get
there. All the matters discussed here are going to seem quite
different when we get to Singularity Eve. But thinking through
these topics, now as I did in writing these essays; and as I
invite you to do while reading them -- is a critical part of the
process of getting there.
-The cover art for the book was done by Zarko Paunovic.
1
Next Steps For And Beyond
Humanity
Some introductory musings
We humans have come such a long way in such a short time!
Yet, were just at the beginning of our evolution, with
extraordinarily dramatic transformations soon to come.
According to our best understanding, the universe has existed
for around 14 billion years, Earth for around 4 billion, multicellular
life for around 2 billion years, and humans with roughly modernstyle brains and bodies only couple million years or so.
Human civilization dates back only 10,000 year or so, and a lot
more recently on most of the planet. In our relatively short time of
existence, we humans have invented/discovered an amazing
variety of stuff like language, mathematics, science, religion,
fashion, sports, social networking, romance, art, music,
corporations, computers, spacecraft well, you get the idea.
The pace of our invention and discovery has been accelerating
fairly consistently. There was a lot more new creation in the
human world from 1000-2000 A.D. than from 0-1000 A.D.; and a
lot more from 0-1000 A.D. than from 1000 B.C.-0 A.D., and so
forth.
Certainly history has had its ups and downs, and the rate of
human progress along various axes is difficult to quantify. But
qualitatively, the accelerating rate of progress is plain to see. Ive
even felt it in my own lifetime; The pace of new gadgets available
and the pace of scientific discoveries is now is incredible
compared to when I was a kid in the 1970s.
10
Next Steps For And Beyond Humanity
However, all the change weve seen so far is trivial compared to

whats coming. So far, the advances the human race has made,
and the transformations its undergone, have all been within the
scope of the human. Weve created new methods for
communicating, new physical tools for doing stuff, and various
techniques for modifying our minds and bodies. Our current
lifestyles and thought-systems are extremely different from those
of our hunter-gatherer forebears; indeed, the life and mind-state
of a young, wired, fervently multitasking Silicon Valley or Hong
Kong technology entrepreneur today bears scant resemblance to
anything that existed 50 or 100 years ago. But, as dramatic as all
this has been, there have been certain limitations. So far, the
human body and brain have remained mostly invariant, and have
provided constraint, structure and guidance to the unfolding
changes.
The nature of the next step in humanitys progress has been
apparent to visionaries for some time, and is lately coming to the
attention of a larger swath of the population. We have created
tools to carry out much of the practical work previously done by
human bodies. Next we will create tools to carry out the work
currently done by human minds. We will create powerful robots
and artificially intelligent software programs not merely narrow
AI programs carrying out specific tasks, but AGIs, artificial
general intelligences capable of coping with unpredictable
situations in intelligent and creative ways. We will do this for the
same reason we created hand-axes, hammers, factories, cars,
antibiotics and computers because we seek to make our lives
easier, more entertaining, and more interesting. Nations and
corporations will underwrite AGI research and development in
order to gain economic advantage over competitors this
happens to a limited extent now, but will become far more
dramatic once AGI technology has advanced a little further. The
result will be something unprecedented in human history; At a
certain point, we will no longer be the most generally intelligent
creatures on the planet.
11
What will be the next step be beyond this? Where will the
acceleration of technology ultimately lead us? Of course its
impossible for us to say, at this point. By analogy, I like to
imagine the first humans to create a language with complex
sentences, sitting around the campfire speculating about where
the wild new invention of language is ultimately going to lead
them. They might have some interesting insights, but would they
foresee mathematics, Dostoevsky, hip-hop, PhotoShop,
supersymmetry, remote-piloted cruise missiles, World of
Warcraft or the Internet?
However, basic logic lets us draw a few conclusions about the
nature of a world where powerful AGIs exist. One is that, if
humans can create AGIs more intelligent than themselves, most
likely these first-generation AGIs will be able to create AGIs with
yet greater intelligence, and so on. These second-generation
AGIs will be able to create yet smarter AGIs. This is what
mathematician I.J. Good, back in the 1960s, called the
intelligence explosion. The dramatic potential consequences of
this sort of intelligence explosion led science fiction writer Vernor
Vinge, in the 1980s, to speak of a coming technological
Singularity.
Mathematically, one interpretation of the term Singularity would
be a point at which some curve or surface changes infinitely fast.
Of course, the rate of technological change is unlikely to actually
approach infinity. Though the physical constraints we now
perceive may well be surmountable, there are probably other
limitations beyond our current understanding. But even so, the
intelligence explosion may well bring a Singularity in the sense
of a dramatic, incredibly hard to predict qualitative change in the
world we live in.
A lot has been written about the Singularity and accelerating
technological change in recent years. Famed inventor Ray
Kurzweil, in particular, has done a huge amount to bring the
Singularity meme to the worlds attention, via his 2006 book The
Singularity Is Near and his numerous speeches and articles. (I
12
Next Steps For And Beyond Humanity
was pleased to have a brief appearance in the Ray Kurzweil

documentary Transcendent Man if you havent seen it, check it
out online!) I wont try to repeat here what has already been said
by Ray, and others before him such as Damien Broderick in his
excellent 1997 book The Spike -- not as catchy a term as
Singularity, but the same basic idea. I encourage you to check
out Kurzweil and Brodericks books, if you havent already, for an
in-depth analysis of the human races accelerating progress over
the last decades, centuries and millennia.
Not everyone believes the Singularity is near, of course. On
Kurzweils kurzweilai.net website, you can also find counterarguments by Kurzweil to many of the complaints made by
detractors of the Singularity idea. Some thinkers whom I much
respect think the Singularity idea is a bit overstated
transhumanist thinker Max More, for example, argues that were
more likely to see a fairly gradual Surge than a Singularity.
Philosopher David Chalmers has laid out a rigorous analytical
argument explaining why a Singularity is very likely to happen,
following the basic logic of I.J. Goods intelligence explosion
idea. But Chalmers is a bit more conservative than Kurzweil in
the particulars. Kurzweil projects a Singularity sometime around
2045, but the Singularity hypothesis as Chalmers analyzes it has
to do with a Singularity sometime within the next few hundred
years, characterized by an explosion from human-level AGI to
massively superintelligent AGI during a period of decades. I
respect the care and rigor of Chalmers analytical approach, but
in the end I suspect Kurzweil is probably close to the mark. I
even think we could see a Singularity well before Kurzweils
projected 2045, if the right alignment of scientific progress and
economic interest comes about. This is one of the things Im
eager to tell you about in this book.
My own particular focus, as a scientist, is on the AGI aspect. The
Singularity (or Surge, or whatever) is going to involve a variety of
different technologies, including genetic engineering, nanotech,
novel computing hardware and maybe quantum computing,
robotics, brain computer interfacing, and a lot more. No doubt
13
there will be new categories of technology that we cant now

envision. However, in the intelligence explosion perspective,
AGI plays a special role its the main technology catalyzing the
next wave of radical change, taking us from the state of humans
with advanced tools but old fashioned bodies and brains to a
new condition that includes radically posthuman features.
The main focus of my own research is on the quest to create AGI
with general intelligence at the human level, and with the
capability to take the next step and create new AGIs with even
greater general intelligence. Furthermore, I want to do this in a
way thats going to have beneficial effects, for human beings in
their current and future form and for the new intelligences we
create. In addition to this, I also work on applications of simpler
AI technologies to practical problems, such as the genetics of
human longevity, and the analysis of financial markets. And Ive
put a fair bit of energy into organizing the AGI research
community, such as launching a series of annual AGI
conferences, and forming a professional research association
called the AGI Society. Ive dabbled with leadership roles in the
futurist, transhumanist community, including a period as
Chairman of Humanity+, which is the only international broadbased nonprofit organization devoted to technology and the
future of humanity. All these experiences have given me a
unique perspective on the present and future of the AGI field,
which is what Ill try to convey to you in these pages.
Unless youre a very rare reader, many of the ideas Ill discuss in
these pages will strike you as radical and out there. But I dont
mind coming across that way. I dont aim to shock, but I also
dont aim to shape my message to suit currently culturally
prevalent ideas. In this book I just call it like I see it.
14
2
The Top Priority for Mankind
In early 2009 I was contacted by some folks associated with the
World Economic Forum best known for their annual conference
in Davos, Switzerland to write an article for distribution at the
Summer Davos World Economic Forum that summer in Dalian,
China. The attendees at Davos are the worlds political and
business movers and shakers top politicians and CEOs and
philanthropists and so forth. My contribution was to be included
(in Chinese translation) in a special issue of Green Herald
Magazine, comprised of articles on the theme of Mankinds Top
Priority.
I like to get right to the point, so the title of my article was:
Mankinds Top Priority Should Be the Creation of Beneficent AI
With Greater than Human Intelligence. The Summer Davos
Forum is also called the Annual Meeting of the New
Champions, so I thought this was particularly appropriate. The
New Champions phrase was presumably chosen to refer to the
leaders of China, India and other emerging markets. But I
wanted to question the assumption that the new champions
leading humanity onwards would always continue to be humans.
Here is a section of what I wrote for the Davos participants:
What should mankinds top priority be going forwards?
This may seem to be a hard question, because there are so
many different problems in the world, and also so many
different opportunities for positive development.
But the answer Ill propose is a very simple one. Our top
priority should be the creation of beneficent artificial minds
15
with greater than human intelligence, which can then work

together with us to solve the other hard problems and
explore various positive avenues.
The pursuit of Artificial General Intelligence (AGI) is currently
a small niche of the research and development world most
of the AI research done today focuses on the creation of
highly specialized narrow AI systems solving isolated
problems. But my contention is that AGI should receive
dramatically more emphasis and should in fact be the human
races top priority going forward. The reasoning behind this
contention is simple: whatever the hard problem at hand, a
greater than human intelligence with the proper motivational
system is going to solve it better than humans can.
This may seem a farfetched notion, but an increasing
number of technologists believe that greater than human AI
could arrive sooner than most people think, quite plausibly
within the next 2-3 decades. Toward this end each year a
number of researchers have gathered in an international
conference on Artificial General Intelligence. If these
optimistic predictions are accurate, clearly this has extremely
important implications for the way we think about ourselves
and our future.
The Benefits Advanced AGI Will Bring
The potential benefits to be delivered by advanced AGI are
difficult to overestimate. Once we are able to create artificial
minds vastly more intelligent than our own, there are no
clear limits to what we may be able to achieve, working
together with our creations. The possibilities quickly start to
sound science fictional but in the era of space travel, the
Internet, genetic engineering, nanotech and quantum
computing, the boundary between science fiction and reality
progressively blurs.
16
Acceleration of Biomedical Research

According to contemporary biology, the human body is in
large part a complex piece of molecular machinery, and
once we have understood its operation sufficiently well, there
is no reason in principle we shouldnt be able to repair its
problems. The main limiting factor here is the human minds
ability to understand the biological data already collected,
and to design new instrumentation able to gather more
accurate data. The advent of AGI biologists should lead to
radical improvements in medicine, including potentially an
end to most or all human disease.
We can already see a start in this direction in the form of
recent work applying AI to bioinformatics (such as the work
of my own firm Biomind LLC) and biomedical text processing
(see the MEDSCAN product by Ariadne Genomics, for
instance). For example, in my own work, we used AI
technology to discover the roots of Parkinsons Disease in
mutations of mitochondrial DNA in brain cells work which is
now being commercialized as a diagnostic test, and being
used to guide research toward a cure. When diseases like
cancer and aging are finally cured, the odds are high that
either narrow AI or AGI will play a major role and there can
be little doubt that AGI would get us there faster.
Discovery and Advancement of Alternative Energy
Technologies
Affordable and sustainable energy is one of the most
daunting problems facing humanity today. However, the
brute fact is that, at the present time, by far the most
affordable energy sources are fossil fuels (with limited roles
for nuclear, geothermal, hydroelectric and other energy
sources in particular situations). Technologies are known
that are in-principle capable of providing vastly more efficient
energy generation without consuming nonrenewable
resources but the process of refining these technologies to
17
the point of economic practicality has proved a slow one,

both because of funding limitations, and because of the
fundamental complexities of the physical processes
involved. It is easy to imagine that an artificial mind with
modestly greater than human intelligence could discover the
electrochemical basis for more effective solar panels, for
example thus revolutionizing the energy field.
Practical Drexlerian Nanotechnology
Back in the 1980s, in his path-breaking book Nanosystems,
Eric Drexler proposed the creation of molecular assemblers
machines capable of piecing together structures out of
molecules, into essentially any configuration desired Much
like a Lego or Erector set, but using real molecules. Today,
nanotechnology is a thriving field of engineering, but most of
the current work avoids the deeper aspects of nanotech that
Drexler envisioned, and focusing instead on such valuable
but narrower problems as using nanotech to create stronger
materials, more efficient conductors, or fabrics with useful
properties.
There is little doubt that Drexlerian nanotech will come; and
firms such as Drexlers NanoRex, with their nanoscale
Computer-Aided Design software, are pushing hard in that
direction. But, one of the major challenges keeping us from
getting there is that human intuition is simply not well-tuned
for the nanoworld. AGI systems need not suffer from this
limitation!
Human Cognitive Enhancement
If smarter-than-human AGI systems can revolutionize the
world, why not smarter-than-human humans? Why not use
advanced technology to make human brains better?
Indeed a number of neurobiologists are thinking hard about
this problem. And at the present time, the worst bottleneck in
18
this research direction seems to be the understanding of

how the human brain works in the first place. In order to
connect external devices to neurons in a useful way, we
need to understand what the signals passed along those
neurons mean; and currently our understanding of human
brain dynamics is extremely crude. The solution to this
problem has two aspects: the creation of better brainimaging devices for noninvasively capturing highly
spatiotemporally accurate neurological data, and the
effective analysis of the data emanating from these brainimaging devices. It is very clear that advanced AI technology
would accelerate our understanding in both these areas.
Making smart AIs may be the most effective path toward
making smarter people.
Acceleration of AGI Research
Last but not least, one of the most promising application
areas for a human-level AGI is AGI itself. As every software
engineer knows, the design and implementation of complex
software is a process that constantly pushes against the
limitations of the human brain such as our limited shortterm memory capacity, which doesnt allow us to
simultaneously manage the states of more than a few dozen
variables or software objects. There seems little doubt that a
human-level AGI, once trained in computer science, would
be able to analyze and refine its own underlying algorithms
with a far greater effectiveness than any human being
My screed on the amazing potential of AGI to solve all the
worlds problems surely struck the Davos crowd as a little far out
there, compared to their usual fare of international politics,
journalism, law, philanthropy and so forth. However, Davos is
supposed to be about thinking big, and I wanted to encourage
the participants to think really big. The creation of AGI with
general intelligence at the human level or beyond will have a
more dramatic, important and fascinating impact than anything
on the program at Davos that year, or any other year so far.
19
3
Why AGI?
Before going any further, I suppose Id better deal with the AGI
issue. Everyone knows what AI is, so why do I insist on using
the less familiar term AGI?
This tweak of terminology points at deeper issues.
The basic reason for talking about AGI is: Modern AI has a split
personality.
In movies and SF novels, AI refers to robots or computer
programs with a high level of intelligence and autonomy R2D2
or C3PO or Hal 9000 or the Terminator or whatever.
However, in university computer science departments and
industry research labs, AI usually refers to something far more
prosaic mainly the creation of highly specialized software
carrying out particular functions that, when humans do them, are
generally considered to require intelligence. The machine at the
paint store that mixes your colors for you and mixes them in with
the white paint, for instance, is an AI at work that has become so
pervasive, we dont even think about the AI in use when were at
the paint store. The term AI needs to be used exclusively when
we are referring to these very specific types of intelligence
programs, for they are a far cry from an AGI; an actual thinking
machine, capable of discerning information coming in
simultaneously from a number of varied and specialized sources.
So far, the only AGI in existence is the human brain. To many,
the promise of building a software-based thinking machine
requires the reverse engineering of our real brains, in an attempt
to set up some kind of real time brain simulation, before we start
feeding questions into it. I do not believe this is the proper
20
Why AGI?
approach. Reverse engineering the brain is a process that will

take 100 years on its own to complete. We should be able to
write software that thinks far before that time, as long as were
not hung up on its processes mimicking that of our real brains.
A new university student, head full of science fiction, might
imagine the community of AI scientists around the world is
working day and night trying to create computers smarter than
people trying to work toward computers capable of holding
intelligent conversations, outsmarting Nobel Prize winners,
writing beautiful poetry, proving amazing new math theorems.
But a glance at the tables of contents of the leading academic AI
journals or conferences would rapidly dash this students hope,
for instead they would find dry formal analyses of very particular
aspects of intelligence, or applications of AI technology to
perform simple puzzles or very particular commercial problems.
AI research as it currently exists focuses almost entirely on what
Kurzweil calls narrow AI highly specialized problem-solving
programs, constituting at most small aspects of intelligence and
involving little or no spontaneity or creativity.
This split personality was not there at the inception of the AI field,
back in the 1950s and 60s; its something that evolved over time.
The founders of AI were thinking about AI in the ambitious
sense. Listen to Nils Nilsson, one of the early leaders of the AI
field back in the 1960s, who reiterated his vision of AI forcefully
in a 2005 article titled Human-Level Artificial Intelligence? Be
Serious!
I claim achieving real Human-Level artificial intelligence
would necessarily imply that most of the tasks that humans
perform for pay could be automated. Rather than work
toward this goal of automation by building special-purpose
systems, I argue for the development of general-purpose,
educable systems that can learn and be taught to perform
any of the thousands of jobs that humans can perform.
Joining others who have made similar proposals, I advocate
beginning with a system that has minimal, although
21
extensive, built-in capabilities. These would have to include

the ability to improve through learning along with many other
abilities.
Hear, hear! This is AI in the grand sense AGI as I conceive it.
The early AI pioneers had some brave, exciting ideas, and some
rather deep insights. But then, as their grand goals proved
unexpectedly tough to achieve, the field drifted away from its
original goals.
Make no mistake, this focus on narrow problem-solving methods
has yielded some exciting fruit. Practical, real-world narrow-AI
work has yielded far more results than those outside the AI field
realize. It has led to a variety of valuable technologies, pervading
nearly every area of society, e.g.
22
the language processing AI underlying search engines

like Google
the planning and scheduling AI used extensively in

government and industry, including many modern
military operations
grandmaster-beating chess programs
industrial robots
video game AI characters, playing a central role in

games played by
hundreds of millions of people
AI financial trading systems
AI data mining systems helping businesses, scientists

and others with historical analysis and decision making
(aka business intelligence)
Why AGI?
However, these narrow-AI programs have not progressed that

far toward the original goal of the AI field, the construction of
what I call AGI.
Only in the last decade has a significant subset of the field began
to shift back toward the fields original focus, on creating AI
systems with general intelligence at the human level and
ultimately beyond. The term AGI arose as part of an effort to
accelerate and deepen this shift of focus.
Terminology is not a terribly deep topic compared to the nature
of intelligence or the construction of thinking machines, yet it has
more influence on the development of science and engineering
than one might think. Black holes have attracted a lot more
research, as well as media attention, than would have been the
case if theyd been called gravitational sinks (a more literal and
transparent name). The field of artificial life flourished
dramatically for a while, in part because of the funky animations
it led to, and in part because of the sexy name. Chaos theory
sounds way whizzier than nonlinear dynamics, so the former is
commonly used even when the nonlinear dynamics in question
arent technically chaotic.
On the other hand, some fields like bioinformatics and
functional genomics have extremely dry and boring names that
conceal the dramatic wonder and importance of the subject
matter. Data mining was fashionable for a while because of its
exciting sound, but then it fell out of favor due to a period in
which the technology was used crudely and led to a lot of
meaningless results, and so now its more suave to refer to
applied machine learning. And of course, both data mining
and machine learning comprise topics typically taught in the
AI courses in university but its sometimes considered
preferable in industry not to refer to them as AI because the
latter carries too much science-fictional baggage Even though
98% of AI research done in academia and industry is quite
prosaic with no direct relation to AGI. Even scientists can be
23
influenced very substantially by the names used to refer to

things.
The strong points of the term AGI are the obvious connection
with the recognized term AI, and the connection with the g
factor of general intelligence, well known in the psychology field.
(It's what IQ tests are supposed to measure.) There are also
some weak sides to the term though; Actually three weak points:
artificial, general, and intelligence!
Artificial isnt really appropriate since AGI isnt just about
building systems to be our tools or artifices. I guess its about
using artifice to build general intelligences though!
General is problematic, because no real-world intelligence will
ever be totally general. Every real-world system will have some
limitations, and will be better at solving some kinds of problems
than others.
Intelligence, finally, is a rather poorly defined term, and, except
in highly abstract mathematical contexts, fairly divorced from
real-world systems. Nobody has a totally clear idea what it
means.
However, in spite of these shortcomings, I actually like the term
AGI well enough. It seems to be catching on reasonably well,
both in the scientific community and in the futurist media. So
Long Live AGI!
Why the Time is Ripe for AGI

But why am I so confident that building a powerful AGI is really
feasible?
Doesnt it worry me that the bulk of my fellow humans, and even
the bulk of professional AI researchers, hold a different opinion
from me on the feasibility of creating AGI in the near term?
24
Why AGI?
In the big picture, it definitely doesnt bother me at all. The vast

majority of major innovations in the history of humanity have
been pooh-poohed by average people and experts alike. Of
course, the vast majority of visionaries moving forward according
to their wild-eyed ideas, while ignoring the naysayers, have
proved them wrong after all.History provides a huge diversity of
examples and situations, leaving one with no realistic choice but
to evaluate each new situation on its own merits. AGI is
something Ive thought about for a long time, including reflecting
on the fact that many others in the past have had misplaced
confidence about the proximity of human-level AGI, and, by this
point, I have a lot of trust in my own intuition on the matter.
Put simply there are five big reasons why I think advanced AGI is
just around the corner:
1. Computers and computer networks are
powerful now, and getting more and more so.
very
2. Computer science has advanced greatly, providing all

manner of wonderful algorithms, many wrapped up in
easily accessible code libraries (like STL, Boost and
GSL, to name three libraries I use in my own work), and
many already engineered so as to be capable of running
on distributed networks of multiprocessor machines.
3. Robots
and
virtual
worlds
have
matured
significantly, making it feasible to interface AI systems
with complex environments at relatively low cost.
4. Cognitive science has advanced tremendously, so
that we now have a pretty strong basic understanding of
what the different parts of the human mind are, what
sorts of things they do, and how they work together
(even though we still dont understand that much about
the internal dynamics of the various parts of the mind,
nor how theyre implemented in the brain)
5. The
Internet
provides
fantastic
tools
for
collaboration both in terms of sharing of ideas among
25
groups of people (email lists, wikis, Web repositories of

research papers), and in terms of collaborative software
code creation (e.g. version control systems used by
open source projects)
Putting these five factors together, one has a situation where a
distributed, heterogenous group of passionate experts can work
together, implementing collections of advanced algorithms on
powerful computers, within software architectures modeled on
human cognitive science. This is, I strongly suspect, how AGI is
going to get created.
There has been quite amazing progress in all in the last two
decades in all five of these areas. Theres no comparison
between the situation now, and the situation back in the 1960s
and 1970s, or the late 1950s when the AI field was founded.
My first AI programs were written around 1980, on an Atari 400
computer with 8KB of RAM and a single 1-2Mhz processor.
Presently, Im running AI software on a machine with 8 3GHz
processors and 96GB of RAM; plus an Nvidia GPU machine with
4 graphics cards, each having several hundred processors
operating in parallel. These are not super-expensive
supercomputers, theyre both a couple years old and had cost in
the ten thousand dollar range. My humble Macbook has 4GB of
RAM and two 2GHz processors. These quantitative changes
make a qualitative difference the larger amounts of RAM
available in current computers make it possible to create AI
programs that maintain a lot of knowledge in a readily-available
state, which is critical if one wants to try for advanced AGI. The
faster processors available today make it possible to utilize this
RAM effectively.
The process of programming is also totally different now than
when I got started. Back then you had to write all your own code
(except the programming language compiler and device drivers,
although I had to write some drivers too now and then).
Nowadays, most programming is a matter of hooking together
26
Why AGI?
bits and pieces that have each been written by a different

person. Hard-core programmers frequently become frustrated
that they dont need to write their own algorithms anymore, as
best-of-breed algorithms are provided inside code libraries, with
nice interfaces to allow them to be rapidly plugged into your own
code. The implementations of the algorithms are regularly
updated to keep pace with algorithmic and hardware
improvements.
For example, recently, we have been modifying our OpenCog
proto-AGI system to try to make it run better on machines with
multiple processors. This has required a fair bit of work, but not
as much as anticipated, because many of the data structures
used are library functions that were already updated for efficient
multiprocessor performance.
Back in the 80s, I did some amateur game programming, and it
involved writing code that drew colors on individual pixels on the
screen. Early AI programs involved computer-simulated blocks
worlds, which were extremely simplistic and allowed very few
kinds of blocks manipulation. Now, in our OpenCog project, we
are still using a sort of blocks world but its implemented in
the Unity3D game engine, using an open-source Unity plug-in
that makes Unity behave somewhat like the popular blocksbuilding game Minecraft.
In 1996, the first robot I built looked like an overturned salad
bowl on three wheels.
It had a bump sensor, a sonar
rangefinder, and was a bit like an early-version Roomba that
didnt vacuum. Presently, in OpenCog, were working with Nao
1
robots .
Nao Robots are short plastic humanoids that walk around and
talk and actually see through their camera eyes. Well soon be
working with Hanson Robokind as well, who creates similar
http://www.youtube.com/watch?v=2STTNYNF4lk
27
humanoids with various technical advantages over the current

Nao bots, including emotionally expressive faces enabled by
innovative artificial skin. Both Nao and Robokind robots are
currently priced around $15,000. If you have a bigger budget,
you can buy the PR2 for a mid-six-figure price tag, which lacks
the humanoid appearance, but is very nearly capable of serving
as a wheeled home service robot. All these robots can be
simulated fairly effectively in freely available robot simulation
software.
When I was in college in the early 80s, and poked around in the
psychology section in the library, I found was an odd mixture of
the nebulous and unscientific with the extremely dry, boring
and narrow-focused. Granted, there were the psychotherapists
and other clinical psychologists, who were thinking about the
whole mind, but in a quite impressionistic way that seemed more
related to art or literature than science.
Freud and Jung and Maslow were fantastic to read, but it was
impossible to take them seriously as guides for building AI.
Clearly, they were just presenting an individual view on mind and
life, much in the manner of philosophers like Plato or Nietzsche,
or the medieval Buddhist psychologists. Sure, they were
developing therapeutic methods that some people found helpful,
but multiple, mutually contradictory religions have done that too.
And ability to help is not a measure of underlying theoretical
accuracy.
On the other hand, laboratory psychology experiments (whether
on rats or humans) seemed to focus on very tiny issues like word
priming or visual illusions, shying away from all the really
interesting parts of the mind. As I later understood when I
worked in a psychology department for a while (at the University
of Western Australia, in the 1990s), this narrowness of focus was
partly due to taste and partly due to practicality, as designing
rigorous psychology lab experiments is really hard. Advances in
neuroimaging have since helped experimental psychology a bit,
but not as broadly as one might think, because of the lack of any
28
Why AGI?
non-invasive brain imaging technology that can measure the

fine-grained temporal dynamics of numerous specific regions of
a brain simultaneously.
I was frustrated that nobody seemed to be trying to put all the
pieces together to try to really understand how the human mind
worked, in a manner both holistic and scientific, using all the data
available from all the different relevant disciplines: psychology,
neuroscience, computer modeling, linguistics, philosophy, etc.
Well guess what? This happened; The interdisciplinary discipline
of Cognitive Science emerged! In my academic days, I helped
found two cognitive science degree programs: Waikato
University in Hamilton, New Zealand, and the University of
Western Australia in Perth.
Cognitive science certainly has not resolved all the questions
about the mind, and, as a discipline, its success has been mixed.
University cognitive science programs have all too often been
absorbed into psychology departments, losing much of their
cross-disciplinary flair and becoming little more than cognitive
psychology. However, the cognitive science movement has
made tremendous progress toward a unified understanding of
the human mind.
And finally, all of these innovations computing hardware,
algorithms, cognitive science and robotics have taken their
current form in large part due to the Internet. Cloud computing,
open-source software, downloadable code libraries, crossdisciplinary collaboration between scientists at different
universities and in different nations, online virtual worlds and
games all this and more is enabled by Internet technology. Its
not any one human nor any small team whos building the
infrastructure to support AGI its the emerging Global Brain.
Putting together advances in computing hardware, robots and
virtual worlds, algorithms and cognitive science, one arrives at a
systematic and viable approach to advanced AGI. You start with
29
a diagram explaining how the human mind works, and showing

what are the main processes and how they Interact. Next, you
look at the assemblage of existing algorithms and data
structures, and figure out a set of algorithms and data structures
that will do everything specified in the cognitive science diagram.
Finally, you implement these algorithms and data structures in a
way that can operate and is able to scale to run on modern
networks of multiprocessor machines. Since this is a damn big
job, you dont do it yourself, but you rely on a team of people
communicating via the Internet, and drawing on expertise from
others outside your team, in the form of Internet
communications, online research papers and so forth.
This is what were doing now with the OpenCog team. Its what
a number of other research teams are doing too. According to
my best understanding, its how powerful AGI will finally be
created.
Of course this isnt the only possible path. Its possible that
detailed brain emulation will get there first. Ill focus on the
integrative cognitive and computer science based approach
here, because its the one Im following in most of my own work,
and its the one that I think has the greatest chance of rapid
success. But ultimately, if either approach succeeds it will enable
the other one. An AGI built via integrative cognitive and
computer science will be a huge help in unraveling the mysteries
of the brain; and an AGI built via emulating the brain would
enable all sorts of experimentation that cant be done on
biological brains, thus leading us to various less brain-like AGI
architectures incorporating various computer science advances.
Naturally, each of the areas whose advance Ive cited above also
has its limitations. I think were just barely at the point now where
these various supporting disciplines have advanced enough to
permit advanced AGI work. 10 years ago, creating advanced
AGI might have been possible, but it would have been much
more difficult to implement. 20 years ago it would have been a
miracle. 10 years from now it will probably be a lot easier, due to
30
Why AGI?
better tools embodying richer understanding. 20 or 30 years from

now, it may be a high school class project.
Computers are fast and RAM is abundant, but programming
software for multiprocessor and distributed systems is still a pain.
In a decade, this will be tremendously easier, due to more
advanced software libraries and algorithms.
Algorithm and data structure libraries are great these days, but
where AI-specific algorithms and structures are concerned, one
often has to roll ones own. MATLAB is good for prototyping
neural networks, but to build a scalable neural net with a novel
architecture, you have to do your own coding. There are logic
rule engines you can download and use to prototype inference
ideas, but none of them scales up well, and all have fairly rigid
restrictions. Whatever AI paradigms you prefer, chances are the
available code libraries are not nearly up to the level of the best
libraries for standard computer science algorithms and data
structures. This is going to change, and it will eventually make
implementation of fairly complex AI systems at least as simple as
implementing database-back-ended websites is today.
Its amazing that we can experiment with humanoid robots now
in a university research lab without massive funding but these
robots, cool as they are, have a lot of limitations. They cant walk
outdoors without falling (Big Dog and some other robots can, but
theyre not humanoid and dont have good hands for
manipulating things, and have other limitations). Their hands
dont work that well. The Naos camera eyes are not so great, at
time of writing, though this may be remedied by the time you
read these words.
And the video game and virtual worlds at our disposal today lack
the richness of the real world by a long shot. No fabrics, no
richly dynamic fluids, no crap, no spit, no peanut butter no
oceans! When a game character picks up an object, this involves
a preprogrammed invisible socket on the characters hand
interacting with a preprogrammed invisible socket on the object
31
its not a free and flexible interaction like when a person picks
up an object with their hand, or when a dog picks up a stick with
its mouth. Robot simulation software doesnt have this limitation
but its slow and not very scalable. Nobody has yet built a
massive multiplayer robot simulator. But one thing is sure they
will.
Cognitive science, as exciting as it is and as fast as it has
progressed, is still a lot stronger on structures than dynamics. In
1991 I joined a fantastically fun group of researchers called the
Society for Chaos Theory in Psychology, devoted to exploring
implications of nonlinear dynamics for intelligence. The group still
exists, and a portion of it, centered around Fred Abrahams lively
Winter Chaos Conference in Vermont, maintains the
revolutionary spirit the group had in the early 90s. One of the
core ideas of this group is that most of the important features of
intelligence are not static structural things, but rather dynamic
attractors complex emergent patterns of system dynamics that
are neither stable nor repetitive nor random, but display more
complex temporal structures. More and more data supporting
this view has emerged, yet the mainstream of cognitive science
hasnt yet incorporated these ideas into its thinking. The boxes
and links in Figure 1 above arent discrete brain regions with
bundles of wires between them in large part, theyre
dynamically assembled, continually self-organizing networks of
neurons that maintain themselves via complex nonlinear
chaotic dynamics.
Much of the reason cognitive science hasnt yet embraced
nonlinear dynamics is that its really hard to measure. We dont
have good enough neuroimaging tools to measure the formation
and interaction of chaotic attractors in the brain, and dynamical
phenomena are pretty hard to study in psychology lab
experiments. But as neuroimaging improves, our understanding
of dynamics in the brain will improve, and cognitive science will
get richer and deeper.
32
Why AGI?
The Internet, as it stands, is nowhere near as great as it could be

at fostering deep intellectual, scientific and engineering
collaboration. Its hard to find the signal among the noise
sometimes, and even with cheap multi-person videoconferencing
like we have now, face-to-face meetings still have higher
semantic and emotional bandwidth. Francis Heylighen at the
Free University of Brussels Well talk to him later - is
developing a host of new technologies aimed at accelerating the
transformation of the Internet into a more richly intelligent Global
Brain. This seems to be happening quite rapidly and effectively,
and AGI research will benefit from this Global Brain, along with
so many other pursuits.
As all these allied areas advance, AGI research will get easier
and easier. So, if I wanted to make my job easier, I would just
wait for the technology infrastructure to mature, and then start
working on AGI some number of years from now. But of course,
if I did that, somebody else might get there first!
Why do I care if someone else gets there first? Of course I have
a fair dose of pride and ambition. It would be fun to be part of the
first team to build AGI! Just as Edmund Hilary wanted to be the
first to ascend Everest, although he also got a lot of pleasure just
from the pure challenge of the climb itself. However, I have
concerns that if AGI is developed too late, the risk of a
dangerous outcome for humanity is greater.
If we can develop advanced AGI soon, then the chance of a
young AGI somehow spiraling out of human control, or being
rapidly deployed by evil humans for massive destruction, seems
fairly low. For a new AGI to be used in destructive ways anytime
soon (now, or in the near future), it would require the use of a lot
of complex, slow-moving infrastructure, and require the
participation and coordination of a lot of people. On the other
hand, once there is a lot more advanced technology of various
sorts available, it could well be possible for a young AGI to wreak
a lot of damage very quickly. For these reasons, I think its a
better idea to bring our first baby non-human AGI into the world
33
when the toys it will be interacting with are a bit less potent. It
would be a little like giving a toddler a loaded gun, but without
the risk of the toddler shooting itself.
Why So Little Work on AGI?

So now youre thinking Well, if all thats all true, and were
poised for an AGI revolution, then how come AGI isnt the focus
of a trillion dollar industry right now? How come AGI isnt the
largest, best-funded department at MIT and Caltech?
However, if youre old enough to have been there, or are familiar
enough with the history of the internet to be able to imagine
Lets go back to 1990. How much thought and research was
going into the Internet and the Web? Hardly any compared to
all the energy and resources on the planet at the time. Yet the
world was poised for it and obviously so, in hindsight. The
world has a way of missing the obvious. But the cool thing is that
when the obvious rears its head, wakes up the world, and bites it
on its ass, the world tends to come around very quickly.
For these reasons, the people of the world will be more than
ready for AGI, when it finally happens. All the sci fi movies and
television shows of the past 100 years, and the exponential rate
of scientific advancement will have prepared them adequately to
put it to good use.
In the case of AGI, there are various historical and practical
factors encouraging most of the world to miss the obvious. Peter
Voss, an AI theorist, entrepreneur and futurist, has summarized
the situation well. His observation, back in 2002, was that, of all
the scientists and engineers working in the AI field,
1. 80% dont believe in the concept of General Intelligence
(but instead, in a large collection of specific skills &
knowledge).
34
Why AGI?
2. Of those that do, 80% dont believe its possibleeither

ever, or for a long, long time.
3. Of those that do, 80% work on domain-specific AI
projects for reasons of commercial or academic politics
(results are a lot quicker).
4. Of those left, 80% have the wrong conceptual
framework.
5. And nearly all of the people operating under basically
correct conceptual premises lack the resources to
adequately realize their ideas.
I think Peters argument is basically on-target. Of course, the
80% numbers are crude approximations, and most of the
concepts involved are fuzzy in various ways. But an interesting
observation is that, whatever the percentages actually are, most
of them have decreased considerably since 2002. Today,
compared to in 2002, many more AI researchers believe AGI is a
feasible goal to work towards, and that it might arise within their
lifetimes. The researchers participating in the AGI conference
series have mostly bypassed the first 2 items on Peters list. And
while funding for AGI research is still very difficult compared to
some other research areas, the situation has definitely improved
in the last 10 years.
From my personal point of view as an AGI researchers, the most
troubling of Peters five points is the fifth one. Most AI
researchers who believe AGI is feasible and would like to be
spending their lives working on it are still working on highly
domain-specific AI projects with much of their time, because
thats where the funding is. Even scientists need to eat, and AGI
research requires computers and programmers and so forth. Im
among the worlds biggest AGI advocates, and I myself spend
about half my time on AGI research and the other half on narrow
AI projects that bring me revenue to pay my mortgage and put
my kids through college.
35
Narrow AI gets plenty of money, in forms like Googles and

Microsofts expenditure on AI-based Web search and ad
placement, and the militarys expenditure on AI-based
intelligence analysis and unmanned vehicle control. But AGI is
relatively minimally funded, compared to these shorter-term,
narrower-scope AI applications.
Indeed, from a practical business perspective, at this stage AGI
is research with at best a medium term payoff its not going to
make anyones profits higher next quarter. One can potentially
chart paths that transition from Narrow AI to AGI, and this may
be a viable way to get to advanced AGI, but its certainly not the
fastest or easiest way and its different than what would
happen if society were to explicitly fund AGI research in a big
way.
The relatively paltry funding of AGI isnt just due to its
speculative nature society is currently willing to fund a variety
of speculative science and engineering projects. Billion dollar
particle accelerators, space exploration, the sequencing of
human and animal genomes, stem cell research, and so forth. If
these sorts of projects merit Big Science level funding, why is
AGI research left out? After all, the potential benefits are
obviously tremendous. There are potential dangers too, to be
sure but there are also clear potential dangers of particle
physics research (discovering better bombs is arguably a
hazardous pursuit), and that doesnt stop us.
Any social phenomenon has multiple intertwined causes, but the
main reason for the AGI fields relatively paltry funding stream is
probably negative momentum from the failures of the original
generation of AI researchers. The AI gurus of the 1960s were
claiming they could create human-level AI within a decade or
less. They were wrong they lacked the needed hardware, their
software tools were primitive, and their conceptual understanding
of intelligence was too crude. Just because they were wrong
then, doesnt mean the current AGI field is similarly wrong but
the guilt by association lingers.
36
Why AGI?
There is an interesting analogy here, related to my observation

above that almost nobody was hyping or funding the Web in
1990. Think about the early visionaries who foresaw the Web
Vannevar Bush in the 1950s, Ted Nelson in the 1960s, and
others. They understood the potential computer technology held
to give rise to something like todays Web and Ted Nelson
even tried to get something Web-like built, back before 1970.
But the technology just wasnt there to support his vision. In
principle it might have been do-able doing the technology of that
era, but it would have been insanely difficult whereas by the
time the Web came about in the 1990s, it seemed almost a
natural consequence of the technological infrastructure existing
at that time. Similarly, in the 1960s, even if someone had come
up with a workable design for human-level AGI, it would have
been extraordinarily difficult to get it implemented and working
using the hardware and software tools available. But now, with
cloud computing, multiprocessor machines with terabytes of
RAM, powerful algorithm libraries and debuggers, and a far more
mature theory of cognitive science, we are in a whole different
position. The conceptual and technological ecosystem is poised
for AGI, in the same sense that it was poised for the Web in the
1990s. And just as the Web spread faster than almost anybody
foresaw so will AGI, once it gets started.
Shifting to a different historical analogy, I think about the future of
AGI as falling into two phases before and after the AGI
Sputnik event.
When the Russians launched Sputnik, this sent a message to
the world: Wow! Going into space is not only a possibility; its a
dramatic reality! The time for humanity to explore space is now!
The consequence was the space race, and the rise of modern
space technology.
Similarly, at some point, some AGI research team is going to
produce a computer program or robot that does something that
makes the world wake up and say: Wow! Genuinely smart AI is
not just a possibility; its a dramatic reality! The time for humanity
37
to create smart machines is now! At that point, government and

industry will put themselves fully behind the creation of advanced
AGI and progress will accelerate tremendously. The potential
of AGI to benefit humanity, and every nation and corporation, is
evidently humongous and all thats needed to shift things
dramatically from the current funding regime to something
overwhelmingly
different,
is
one
single
crystal-clear
demonstration that human-level AGI is feasible in the near term.
We dont have that demonstration right now, but I and a
number of other AGI researchers believe we know how to do
it And Im betting it will happen within the lifetimes of most
people reading this book.
In the late 60s and early 70s, in the era of the Apollo moon
missions, I and every other little American kid wanted to grow up
and be an astronaut. After the AGI Sputnik event, kids will want
to grow up and be AGI developers or AGIs!
I believe my OpenCog AGI approach has what it takes to get us
to an AGI Sputnik event in the form, perhaps, of a video game
character or a humanoid robot that holds meaningful
conversations about the things in its environment. Imagine
talking to a robot that really gives you sense that it knows what
its talking about that it understand what its doing, and knows
who it is and who you are. That will be a damn strange feeling
and a wonderful one. And everyone who gets that feeling will
understand that humanity is about to take the next, huge, step.
Ill talk a bit about OpenCog in these pages but my main point
here isnt to sell my own particular approach to AGI; rather, to
talk about AGI in general, and the broader implications that the
advent of AGI will have for humanity and beyond. Many of my
colleagues have their own different perspective on the optimal
technical approach to create AGI. My main goal here is, first, to
get across the points that AGI is probably coming fairly soon,
and its going to be a huge change and quite possibly a fantastic
one for all of us; and second, to explore some of the things that
AGI has to teach us about the nature of mind and intelligence.
38
Why AGI?
AGI is coming probably faster than you think and its going to
be really, really interesting
39
4
Ten Years to the Singularity If We Really
Really Try
Weve discussed the Vinge-ean, Kurzweil-ian argument that
human-level AGI may be upon us shortly. By extrapolating
various key technology trends into the near future, in the context
of the overall dramatic technological growth the human race has
seen in the past centuries and millennia, it seems quite plausible
that superintelligent artificial minds will be here much faster than
most people think.
This sort of objective, extrapolative view of the future has its
strengths, and is well worth pursuing. But I think its also valuable
to take a more subjective and psychological view, and think
about AGI and the Singularity in terms of the power of the human
spirit; What we really want for ourselves, and what we can
achieve if we really put our minds to it.
I presented this sort of perspective on the timeline to Singularity
and advanced AGI at the TransVision 2006 futurist conference,
in a talk called Ten Years to a Positive Singularity (If We Really,
Really Try). The conference was in Helsinki, Finland, and I
wasnt able to attend in person so I delivered the talk by video
2
if youre curious you can find it online .
The basic point of the talk was that if society put the kind of
money and effort into creating a positive Singularity that we put
into things like wars or television shows, then some pretty
amazing things might happen. To quote a revised version of the
talk, given to a different audience just after the financial crisis of
Fall 2008:
http://www.youtube.com/watch?v=BelOkx4Jxyg
40
Look at the US governments response to the recent

financial crisis suddenly theyre able to materialize a trillion
dollars here, a trillion dollars there. What if those trillions of
dollars were being spent on AI, robotics, life extension,
nanotechnology and quantum computing? It sounds
outlandish in the context of how things are done now but
its totally plausible.
If we made a positive Singularity a real focus of our society, I
think a ten year time-frame or less would be eminently
possible.
Ten years from now would be 2020. Ten years from 2007,
when that talk was originally given, would have been 2017,
only 7 years from now. Either of these is a long time before
Kurzweils putative 2045 prediction. Whence the gap?
When he cites 2045, Kurzweil is making a guess of the
most likely date for the Singularity. The ten more years
prediction is a guess of how fast things could happen with an
amply-funded, concerted effort toward a beneficial
Singularity. So the two predictions have different intentions.
We consider Kurzweils 2045 as a reasonable extrapolation
of current trends, but we also think the Singularity could
come a lot sooner, or a lot later, than that.
How could it come a lot later? Some extreme possibilities
are easy to foresee. What if terrorists nuke the major cities of
world? What if anti-technology religious fanatics take over
the worlds governments? But less extreme outcomes could
also occur, with similar outcomes. Human history could just
take a different direction than massive technological
advance, and be focused on warfare, or religion, or
something else.
Or, though we reckon this less likely, it is also possible we
could hit up against tough scientific obstacles that we cant
41
foresee right now. Intelligence could prove more difficult for

the human brain to puzzle out, whether via analyzing
neuroscience data, or engineering intelligent systems.
Moores Law and its cousins could slow down due to
physical barriers, designing software for multicore
architectures could prove problematically difficult; the pace
of improvement in brain scanners could slow down.
How, on the other hand, could it take a lot less time? If the
right people focus their attention on the right things.
The Ten Years to the Singularity talk began with a well-known
motivational story, about a guy named George Dantzig (no
relation to the heavy metal singer Glenn Danzig!). Back in 1939,
Dantzig was studying for his PhD in statistics at the University of
California, Berkeley. He arrived late for class one day and found
two problems written on the board. He thought they were the
homework assignment, so he wrote them down, then went home
and solved them. He thought they were particularly hard, and it
took him a while. But he solved them, and delivered the solutions
to the teachers office the next day. Turns out, the teacher had
put those problems on the board as examples of "unsolvable"
statistics problems; Two of the greatest unsolved problems of
mathematical statistics in the world, in fact. Six weeks later,
Dantzig's professor told him that hed prepared one of his two
"homework" proofs for publication. Eventually, Dantzig would use
his solutions to those problems for his PhD thesis.
Heres what Dantzig said about the situation: If I had known that
the problems were not homework, but were in fact two famous
unsolved problems in statistics, I probably would not have
thought positively, would have become discouraged, and would
never have solved them.
Dantzig solved these problems because he thought they were
solvable; He thought that other people had already solved them.
42
He was just doing them as homework, thinking everyone else

in his class was going to solve them too.
Theres a lot of power in expecting to win. Athletic coaches know
about the power of streaks. If a team is on a roll, they go into
each game expecting to win, and their confidence helps them
see more opportunities to win. Small mistakes are just shrugged
away by the confident team, but if a team is on a losing streak,
they go into each game expecting to screw up, somehow. A
single mistake can put them in a bad mood for the whole game,
and one mistake can pile on top of another more easily.
To take another example, lets look at the Manhattan Project.
America thought they needed to create nuclear weapons before
the Germans did. They assumed it was possible, and felt a huge
burning pressure to get there first. Unfortunately, what they were
working on so hard, with so much brilliance, was an ingenious
method for killing a lot of people. But, whatever you think of the
outcome, theres no doubt the pace of innovation in science and
technology in that project was incredible, and it all might have
never happened if the scientists involved didnt already believe
that Germany was ahead of them, and that somehow their
inventing the ability to kill thousands, first, would save humanity.
How Might a Positive Singularity Get Launched In

10 Years From Now?
This way of thinking leads to a somewhat different way of
thinking about the timing of the Singularity. What if, rather than
thinking about it as a predictive exercise (an exercise in objective
studying whats going to happen in the world, as if we were
outsiders to the world). What if we thought about it the way an
athlete thought about a game when going into it, or the way the
Manhattan Project scientists thought at the start of the project, or
the way Dantzig thought about his difficult homework problems?
43
What if we knew it was possible to create a positive Singularity in

ten years? What if we assumed we were going to win, as a
provisional but reasonable hypothesis?
What if we thought everyone else in the class knew how to do it
already?
What if we were worried the bad guys were going to get there
first?
Under this assumption, how then would we go about trying to
create a positive Singularity?
Following this train of thought, even just a little ways, will lead
you along the chain of reasoning that led us to write this book.
One conclusion that seems fairly evident when taking this
perspective is that AI is the natural area of focus.
Look at the futurist technologies at play these days:
nanotechnology
biotechnology
robotics
AI
and ask, which ones have the most likelihood of bringing us a

positive Singularity within the next ten years?
Nano and bio and robotics are all advancing fast, but they all
require a lot of hard engineering work.
AI requires a lot of hard work too, but its a softer kind of hard
work. Creating AI relies only on human intelligence, not on
painstaking and time-consuming experimentation with physical
substances and biological organisms.
And how can we get to AI? There are two big possibilities:
44
Copy the human brain, or
Come up with something cleverer
Both approaches seem viable, but the first approach has a

problem. Copying the human brain requires far more
understanding of the brain than we have now. Will biologists get
there in ten years from now. Probably not. Definitely not in five
years.
So were left with the other choice, come up with something
cleverer. Figure out how to make a thinking machine, using all
the sources of knowledge at our disposal: Computer science and
cognitive science and philosophy of mind and mathematics and
cognitive neuroscience and so forth.
But if this is feasible to do in the near term, which is what were
suggesting, then why dont have AIs smarter than people right
now? Of course, its a lot of work to make a thinking machine,
but making cars and rockets and televisions is also a lot of work,
and society has managed to deal with those problems.
The main reason we dont have real AI right now is that almost
no one has seriously worked on the problem. And (here is where
things get even more controversial!) most of the people that have
worked on the problem have thought about it in the wrong way.
Some people have thought about AI in terms of copying the
brain, but, as I mentioned earlier, that means you have to wait
until the neuroscientists have finished figuring out the brain.
Trying to make AI based on our current, badly limited
understanding of the brain is a clear recipe for failure. We have
no understanding yet of how the brain represents or manipulates
abstraction. Neural network AI is fun to play with, but its hardly
surprising it hasnt led to human-level AI yet. Neural nets are
based on extrapolating a very limited understanding of a few
very narrow aspects of brain function.
45
The AI scientists who havent thought about copying the brain,

have mostly made another mistake; Theyve thought like
computer scientists. Computer science is like mathematics; Its
all about elegance and simplicity. You want to find beautiful,
formal solutions. You want to find a single, elegant principle. A
single structure. A single mechanism that explains a whole lot of
different things. A lot of modern theoretical physics is in this vein.
The physicists are looking for a single, unifying equation
underlying every force in the universe. Well, most computer
scientists working on AI are looking for a single algorithm or data
structure underlying every aspect of intelligence.
But thats not the way minds work. The elegance of mathematics
is misleading. The human mind is a mess, and not just because
evolution creates messy stuff. The human mind is a mess
because intelligence, when it has to cope with limited computing
resources, is necessarily messy and heterogenous.
Intelligence does include a powerful, elegant, general problemsolving component, and some people have more of it than
others. Some people I meet seem to have almost none of it at
all.
But intelligence also includes a whole bunch of specialized
problem-solving components dealing with things like: vision,
socialization, learning physical actions, recognizing patterns in
events over time, and so forth. This kind of specialization is
necessary if youre trying to achieve intelligence with limited
computational resources.
Marvin Minsky has introduced the metaphor of a society. He
says a mind needs to be a kind of society, with different agents
carrying out different kinds of intelligent actions and all
interacting with each other.
But a mind isnt really like a society. It needs to be more tightly
integrated than that. All the different parts of the mind, parts
which are specialized for recognizing and creating different kinds
46
of patterns, need to operate very tightly together, communicating

in a common language, sharing information, and synchronizing
their activities.
And then comes the most critical part; The whole thing needs to
turn inwards on itself. Reflection. Introspection. These are two of
the most critical kinds of specialized intelligence that we have in
the human brain, and both rely critically on our general
intelligence ability. A mind, if it wants to be really intelligent, has
to be able to recognize patterns in itself, just like it recognizes
patterns in the world, and it has to be able to modify and improve
itself based on what it sees in itself. This is what self is all
about.
This relates to what the philosopher Thomas Metzinger calls the
phenomenal self . All humans carry around inside our minds a
phenomenal self. An illusion of a holistic being. A whole
person. An internal self that somehow emerges from the mess
of information and dynamics inside our brains. This illusion is
critical to what we are. The process of constructing this illusion
is essential to the dynamics of intelligence.
Brain theorists havent understood the way the self emerges
from the brain yet, because brain mapping isnt advanced
enough.
Computer scientists havent understood the self, because it isnt
about computer science. Its about the emergent dynamics that
happen when you put a whole bunch of general and specialized
pattern recognition agents together; A bunch of agents created in
a way that they can really cooperate, and when you include in
the mix agents oriented toward recognizing patterns in the
society as a whole.
The specific algorithms and representations inside the pattern
recognition agents algorithms dealing with reasoning, or
seeing, or learning actions, or whatever these algorithms are
what computer science focuses on. Theyre important, but
47
theyre not really the essence of intelligence. The essence of

intelligence lies in getting the parts to all work together in a way
that gives rise to the phenomenal self. That is, its about wiring
together a collection of structures and processes into a holistic
system in a way that lets the whole system recognize significant
patterns in itself. With very rare exceptions, this has simply not
been the focus of AI researchers.
When talking about AI in these pages, Ill use the word patterns
a lot. This is inspired in part by my book The Hidden Pattern,
which tries to get across the viewpoint that everything in the
universe is made of patterns. This is not a terribly controversial
perspective Kurzweil has also described himself as a
patternist. In the patternist perspective, everything you see
around you, everything you think, everything you remember;
Thats a pattern!
Following a long line of other thinkers in psychology and
computer science, we conceive intelligence as the ability to
achieve complex goals in complex environments. Even
complexity itself has to do with patterns. Something is complex
if it has a lot of patterns in it.
A mind is a collection of patterns for effectively recognizing
patterns. Most importantly, a mind needs to recognize patterns
about what actions are most likely to achieve its goals.
The phenomenal self is a big pattern, and what makes a mind
really intelligent is its ability to continually recognize this pattern;
The phenomenal self in itself.
Does it Take a Manhattan Project?

One of the more interesting findings from the How Long Till
Human-Level AI survey we discussed above was about funding,
and the likely best uses of hypothetical massive funding to
promote AGI progress.
48
In the survey, we used the Manhattan Project as one of our

analogies, just as I did in part of the discussion above -- but in
fact, it may be that we dont need a Manhattan Project scale
effort to get Singularity-enabling AI. The bulk of AGI researchers
surveyed at AGI-09 felt that, rather than a single monolithic
project, the best use of massive funding to promote AGI would
be to fund a heterogenous pool of different projects, working in
different but overlapping directions.
In part, this reflects the reality that most of the respondents to the
survey probably thought they had an inkling (or a detailed
understanding) of a viable path to AGI, and feared that an AGI
Manhattan Project would proceed down the wrong path instead
of their correct path. But it also reflects the realities of software
development. Most breakthrough software has come about
through a small group of very brilliant people working together
very tightly and informally. Large teams work better for hardware
engineering than software engineering.
It seems most likely that the core breakthrough enabling AGI will
come from a single, highly dedicated AGI software team. After
this breakthrough is done, a large group of software and
hardware engineers will probably be useful for taking the next
step, but thats a different story.
What this suggests is that, quite possibly, all we need right now
to get Singularity-enabling AGI is to get funding to a dozen or so
of the right people. This would enable them to work on the right
AGI project full time for a decade or so, or maybe even less.
Its worth emphasizing that my general argument for the potential
imminence of AGI does not depend on my perspective on any
particular route to AGI being feasible. Unsurprisingly, Im a big
fan of the OpenCog project, of which Im one of the founders and
leaders. Ill tell you more about this a little later on. But you dont
need to buy my argument for OpenCog as the most likely path to
AGI, in order to agree with my argument for creating a positive
Singularity by funding a constellation of dedicated AGI teams.
49
Even if OpenCog were the wrong path, there could still be a lot
of sense in a broader bet that funding 100 dedicated AGI teams
to work on their own independent ideas will result in one of them
making the breakthrough. Whats shocking, given the amount of
money and energy going into other sorts of technology
development, is that this isnt happening right now. (Or maybe it
is, by the time you are reading this!!)
Keeping it Positive
Ive talked more about AI than about the Singularity or
positiveness. Let me get back to those.
It should be obvious that if you can create an AI vastly smarter
than humans, then pretty much anything is possible.
Or at least, once we reach that stage, theres no way for us, with
our puny human brains, to really predict what is or isnt possible.
Once the AI has its own self, and has superhuman level
intelligence, its going to start learning and figuring things out on
its own.
But what about the positive part? How do we know this AI wont
annihilate us all? Why wont it just decide were a bad use of
mass-energy, and re-purpose our component particles for
something more important?
Theres no guarantee of this not happening, of course.
Just like theres no guarantee that some terrorist wont nuke your
house tonight, or that you wont wake up tomorrow morning to
find the whole life you think youve experienced has been a long
strange dream. Guarantees and real life dont match up very
well. (Sorry to break the news.)
However, there are ways to make bad outcomes unlikely, based
on a rational analysis of AI technology and the human context in
which its being developed.
50
The goal systems of humans are pretty unpredictable, but a

software mind can potentially be different, because the goal
system of an AI can be more clearly and consistently defined.
Humans have all sort of mixed-up goals, but there seems no
clear reason why one cant create an AI with a more crisply
defined goal of helping humans and other sentient beings, as
well as being good to itself. We will return to the problem of
defining this sort of goal more precisely later, but to make a long
story short, one approach (among many) is to set the AI the goal
of figuring out what is most common among the various requests
that various humans may make of it.
One risk of course is that, after it grows up a bit, the AGI
changes its goals, even though you programmed it not to. Every
programmer knows you cant always predict the outcome of your
own code. But there are plenty of preliminary experiments we
can do to understand the likelihood of this happening. And there
are specific AGI designs, such as the GOLEM design well
discuss below, that have been architected with a specific view
toward avoiding this kind of pathology. This is a matter best
addressed by a combination of experimental science and
mathematical theory, rather than armchair philosophical
speculation.
Ten Years to the Singularity?

How long till human-level or superhuman AGI? How long till the
Singularity? None of us knows. Ray Kurzweil, and others, have
made some valuable projections and predictions. But you and I
are not standing outside of history analyzing its progress; we are
co-creating it. Ultimately the answer to this question is highly
uncertain, and, among many other factors, it depends on what
we do.
To quote the closing words of the Ten Years to The Singularity
TransVision talk:
A positive Singularity in 10 years?
51
Am I sure its possible? Of course not.

But I do think its plausible.
And I know this: If we assume it isnt possible, it wont be.
And if we assume it is possible and act intelligently on this
basis it really might be. Thats the message I want to get
across to you today.
There may be many ways to create a positive Singularity in
ten years. The way Ive described to you the AI route is
the one that seems clearest to me. There are six billion
people in the world so theres certainly room to try out many
paths in parallel.
But unfortunately the human race isnt paying much attention
to this sort of thing. Incredibly little effort and incredibly little
funding goes into pushing toward a positive Singularity. Im
sure the total global budget for Singularity-focused research
is less than the budget for chocolate candy let alone beer
... Or TV... Or weapons systems!
I find the prospect of a positive Singularity incredibly exciting
and I find it even more exciting that it really, possibly could
come about in the next ten years. But its only going to
happen quickly if enough of the right people take the right
attitude and assume its possible, and push for it as hard
as they can.
Remember the story of Dantzig and the unsolved problems
of statistics. Maybe the Singularity is like that. Maybe
superhuman AI is like that. If we dont think about these
problems as impossibly hard quite possibly theyll turn out
to be solvable, even by mere stupid humans like us.
This is the attitude Ive taken with my work on OpenCog. Its
the attitude Aubrey de Grey has taken with his work on life
52
extension. The more people adopt this sort of attitude, the

faster the progress well make.
We humans are funny creatures. Weve developed all this
science and technology but basically were still funny little
monkeylike creatures from the African savannah. Were
obsessed with fighting and reproduction and eating and
various apelike things. But if we really try, we can create
amazing things new minds, new embodiments, new
universes, and new things we cant even imagine.
53
5
How Long Till AGI?
with Ted Goertzel and Seth Baum
This chapter originated as an H+ Magazine article. A longer,
more technical version appeared in the journal Technological
Forecasting and Social Change.
What do other experts think about Kurzweils projection of an AIpowered Singularity around 2045?
Opinions are all over the map but one interesting data point is
a survey that Seth Baum, my father Ted Goertzel and I
conducted at the AGI-09 conference, a gathering of Artificial
General Intelligence researchers in Washington DC in March
2009, on the specific topic of How Long Till Human Level AI?
Rather than a shallow survey of a large number of people, this
was an in-depth survey of a small number of experts what is
known in the trade as an expert elicitation. The sample of
experts was avowedly biased most (though not all) researchers
who would bother to attend an AGI conference are relatively
optimistic about the near-term feasibility of AGI. But even given
this bias, its very interesting to sample the perspective of AGIinterested experts and see where there opinions fall on various
aspects of the question when will human-level AI be here?
Two Earlier Surveys of Expert Opinion on the

Timing of Human-Level AI
We know of two previous studies exploring expert opinion on the
future of artificial general intelligence. In 2006, a seven-question
poll was taken of participants at the AI@50 conference. Four of
the seven questions are particularly relevant. Asked when will
54
How Long Till AGI?
computers be able to simulate every aspect of human

intelligence?, 41% said More than 50 years and 41% said
Never. Seventy-nine percent said an accurate model of thinking
is impossible without further discoveries about brain functioning.
Thus approximately 80% of the participants at the conference
were AI timing pessimists and 20% timing optimists,
percentages that may reflect the wider pattern in the field. Sixty
percent of the participants strongly agreed that AI should take a
multidisciplinary approach, incorporating stats, machine learning,
linguistics, computer science, cognitive psychology, philosophy,
and biology. And finally, 71% said that statistical/probabilistic
methods are most accurate in representing how the brain works.
The other survey, taken in 2007 by futurist entrepreneur Bruce
Klein, was an online survey that garnered 888 responses, asking
one question: When will AI surpass human-level intelligence?
Most of those who chose to respond to this survey were AI
optimists who believed that human-level artificial intelligence
would be achieved during the next half century. The distribution
of responses is shown below:
The AI@50 and Klein studies are interesting because they show
that significant numbers of experts and interested persons
believe that AGI with intelligence at least equaling the human
level will exist within upcoming decades. Our own study probes
more deeply into the thinking of people with substantial AGI
expertise, and comes up with similar findings, but giving more
details about the experts underlying beliefs.
55
The Questions Asked of the Experts

The first set of questions in our survey elicited experts beliefs
about when AI would reach each of four milestones: passing the
3
Turing test, performing Nobel quality work, passing third grade,
and becoming superhuman. These specific milestones were
selected to span a variety of levels of advanced general
intelligence.
For each milestone, two question versions were asked with
and without massive additional funding making for eight total
milestone questions. These two versions explore the possibility
that when the milestones are reached depends on the resources
available to researchers. The amount of additional funding listed,
$100 billion per year, is obviously more than could be used for
AGI research; the intent with this figure was to ensure that
money would not be a scarce resource in this hypothetical AGI
development scenario.
(Statistics geeks may be interested that, for each of the eight
milestone questions, we asked the respondents to give us
estimates representing 10%, 25%, 75% and 90% confidence
intervals, as well as their best estimate dates. In the full write-up
of our report we will present this data, but in this article well stick
with the best estimates, for sake of simplicity.)
Our next set of questions covered four topics. Three questions
asked what embodiment the first AGIs would have: physical or
While taking the survey, several participants expressed concern that

the Turing test milestone is ambiguous, due to the numerous variations
of the test. In response to this, several participants specified the Turing
test variant their responses are based on. At the time of survey
distribution, a verbal suggestion was given to consider the one hour
rather than the five minute Turing test as some potential participants
felt the latter could too easily be gamed by narrow-AI chatbots without
significant general intelligence.
56
How Long Till AGI?
virtual robot bodies or a minimal text- or voice-only embodiment.

Eight questions asked what AI software paradigm the first AGIs
would be based on: formal neural networks, probability theory,
uncertain logic, evolutionary learning, a large hand-coded
knowledge-base, mathematical theory, nonlinear dynamical
systems, or an integrative design combining multiple paradigms.
Three questions asked the likelihood of a strongly negative-tohumanity outcome if the first AGIs were created by: an opensource project, the US military, or a private for-profit software
company. Two true/false questions asked if quantum computing
or hypercomputing would be required for AGI. Two yes/no
questions asked if AGIs emulating the human brain conceptually
or near-exactly would be conscious in the sense that humans
are. Finally, 14 questions asked experts to evaluate their
expertise in: cognitive science, neural networks, probability
theory, uncertain logic, expert systems, theories of ethics,
evolutionary learning, quantum theory, quantum gravity theory,
robotics, virtual worlds, software engineering, computer
hardware design, and cognitive neuroscience.
We posed these questions to 21 participants, with a broad range
of backgrounds and experience, all with significant prior thinking
about AGI. Eleven are in academia, including six Ph.D. students,
four faculty members and one visiting scholar, all in AI or allied
fields. Three lead research at independent AI research
organizations, and three do the same at information technology
organizations. Two are researchers at major corporations. One
holds a high-level administrative position at a relevant non-profit
organization. One is a patent attorney. All but four participants
reported being actively engaged in conducting AI research.
What the Experts Said About The Timing of

Human-Level AI
Unsurprisingly, our results showed that the majority of the
experts who participated in our study were AGI timing optimists
but there was a significant minority of timing pessimists.
Although its worth noting that all the experts in our study, even
57
the most pessimistic ones, gave at least a 10% chance of some

AI milestones being achieved within a few decades.
The range of best-guess time estimates for the AGI milestones
without massive additional funding is summarized below:
Distributions of best guess estimates for when AI would achieve

four milestones without additional funding: the Turing Test
(horizontal lines), third grade (white), Nobel-quality work (black),
and superhuman capability (grey).
Another interesting result concerns the role of funding the

predicted impact on the AGI field of a hypothetical multibillion
dollar funding infusion. Most experts estimated that a massive
funding increase of this nature would cause the AI milestones to
be reached sooner. However, for many of these experts, the
difference in milestone timing with and without massive funding
was quite small just a few years. Furthermore, several experts
estimated that massive funding would actually cause the AI
milestones to be reached later. One reason given for this is that
with so much funding, many scholars would focus on making
money and administration instead of on research. Another is
that massive funding increases corruption in a field and its
oppression of dissenting views in the long term. Of those who
thought funding would make little difference, a common reason
was that AGI progress requires theoretical breakthroughs from
just a few dedicated, capable researchers, something that does
not depend on massive funding. Another common reason was
that the funding would not be wisely targeted. Several noted that
funding could be distributed better if there was a better
58
How Long Till AGI?
understanding of what paradigms could produce AGI, but such

an understanding is either not presently available or not likely to
be understood by those who would distribute funds.
Because of the lack of agreement on a single paradigm, several
experts recommended that modest amounts of funding should
be distributed to a variety of groups following different
approaches, instead of large amounts of funding being given to a
Manhattan Project type crash program following one approach.
Several also observed that well-funded efforts guided by a single
paradigm had failed in the past, including the Japanese Fifth
Generation Computer Systems project. On this, one said, AGI
requires more theoretical study than real investment. Another
said, I believe the development of AGIs to be more of a tool and
evolutionary problem than simply a funding problem. AGIs will
be built upon tools that have been developed from previous
tools. This evolution in tools will take time. Even with a crash
project and massive funding, these tools will still need time to
develop and mature. Given that these experts are precisely
those who would benefit most from increased funding, their
skeptical views of the impact of hypothetical massive funding are
very likely sincere.
Distributions of best guess estimates for when AI would achieve

four milestones with additional funding: the Turing Test (horizontal
lines), third grade (white), Nobel-quality work (black), and
superhuman capability (grey).
59
There was substantial agreement among the more optimistic

participants in the study. Most thought there was a substantial
likelihood that the various milestones would be passed sometime
between 2020 and 2040. 16 experts gave a date before 2050 as
their best estimate for when the Turing test would be passed. 13
experts gave a date between 2020 and 2060 as their best
estimate for when superhuman AI would be achieved. (The
others all estimated later dates, ranging from 2100 to never.) The
opinions of the optimists were similar to and even slightly more
optimistic than Kurzweils well-known projections.
As noted above, optimism in this sense is about the timing of AI
milestones. It does not imply a belief that achieving AGI would
be a good thing. To the contrary, one can be optimistic that AGI
will happen soon yet believe that AGI would have negative
outcomes. Indeed, several experts reported this pair of beliefs.
Results about the likelihood of negative outcomes are discussed
further below.
What Kind of Human-Level AI Will Come First?

To our surprise, there was not much agreement among experts
on the order in which the four milestones (Turing test; third
grade; Nobel; superhuman) would be achieved. There was
consensus that the superhuman milestone would be achieved
either last or at the same time as other milestones. However,
there was significant divergence regarding the order of the other
three milestones. One expert argued that the Nobel milestone
would be easier precisely because it is more sophisticated: to
pass the Turing test, an AI must skillfully hide such mental
superiorities. Another argued that a Turing test-passing AI
needs the same types of intelligence as a Nobel AI but
additionally needs to fake a lot of human idiosyncrasies
(irrationality, imperfection, emotions). Finally, one expert noted
that the third grade AI might come first because passing a third
grade exam might be achieved by advances in natural language
processing, without actually creating an AI as intelligent as a
third-grade child. This diversity of views on milestone order
60
How Long Till AGI?
suggests a rich, multidimensional understanding of intelligence.

It may be that a range of milestone orderings are possible,
depending on how AI development proceeds.
The milestone order results highlight the fact that many experts
do not consider it likely that the first human-level AGI systems
will closely mimic human intelligence. Analogy to human
intelligence would suggest that achieving an AGI capable of
Nobel level science would take much longer than achieving an
AGI capable of conducting a social conversation. However, as
discussed above, an AGI would not necessarily mimic human
intelligence. This could enable it to achieve the intelligence
milestones in other orders.
What Kind of Technical Approach Will First

Achieve Human-Level AI?
Our initial survey asked about eight technical approaches used
in constructing AI systems: formal neural networks, probability
theory, uncertain logic, evolutionary learning, a large hand-coded
knowledge-base, mathematical theory, nonlinear dynamical
systems, and integrative designs combining multiple paradigms.
For each, it requested a point estimate of the odds that the
approach would be critical in the creation of human-level AGI.
They survey also asked about the odds that physical robotics,
virtual agent control, or minimal text- or voice-based
embodiments would play a critical role.
The question about robotics vs. other forms of embodiment
received a range of responses. There were responses of .9, .94,
.89 and .6 for physical robot embodiment, but the mean
response was only 27%. The handful of participants who felt
robotics was crucial were all relatively optimistic. The preliminary
impression one obtains is that a few researchers are highly
bullish on robotics as the correct path to AGI in the relatively
near term, whereas the rest feel robotics is probably not
necessary for AGI.
61
Impacts of AGI
In science fiction, intelligent computers frequently become
dangerous competitors to humanity, sometimes even seeking to
exterminate humanity as an inferior life form. And indeed, based
on our current state of knowledge, its hard to discount this as a
real possibility, alongside much more benevolent potential
outcomes. We asked experts to estimate the probability of a
negative-to-humanity outcome occurring if an AGI passes the
Turing test. Our question was broken into three parts, for each of
three possible development scenarios: if the first AGI that can
pass the Turing test is created by an open source project, the
United States military, or a private company focused on
commercial profit.
This set of questions marked another instance in which the
experts wildly lacked consensus. Four experts estimated a less
than 20% chance of a negative outcome, regardless of the
development scenario. Four experts estimated a greater than
60% chance of a negative outcome, regardless of the
development scenario. Only four experts gave the same
estimate for all three development scenarios; the rest of the
experts reported different estimates of which development
scenarios were more likely to bring a negative outcome. Several
experts were more concerned about the risk from AGI itself,
whereas others were more concerned that AGI could be misused
by humans who controlled it.
Some interesting insights can be found in the experts orderings
of the riskiness of the development scenarios. Of the 11 experts
who gave different dangerousness estimates for each of the
three scenarios, 10 gave the private company scenario the
middle value. Of these 10, 6 gave the US military scenario the
highest value and 4 gave it the lowest value. Thus the open
source scenario and the US military scenario tend to be
perceived opposites in terms of danger but experts are divided
on which is the safe one of the pair! Experts who estimated that
the US military scenario is relatively safe noted that the US
62
How Long Till AGI?
military faces strong moral constraints, has experience handling

security issues, and is very reluctant to develop technologies that
may backfire (such as biological weapons), whereas open
source development lacks these features and cannot easily
prevent the deliberate insertion of malicious code. In contrast,
experts who estimated that the open source scenario is relatively
safe praised the transparency of open source development and
its capacity to bring more minds to the appropriate problems, and
felt the military has a tendency to be destructive.
Several experts noted potential impacts of AGI other than the
catastrophic. One predicted that in thirty years, it is likely that
virtually all the intellectual work that is done by trained human
beings such as doctors, lawyers, scientists, or programmers, can
be done by computers for pennies an hour. It is also likely that
with AGI the cost of capable robots will drop, drastically
decreasing the value of physical labor. Thus, AGI is likely to
eliminate almost all of todays decently paying jobs. This would
be disruptive, but not necessarily bad. Another expert thought
that, societies could accept and promote the idea that AGI is
mankind's greatest invention, providing great wealth, great
health, and early access to a long and pleasant retirement for
everyone. Indeed, the experts comments suggested that the
potential for this sort of positive outcome is a core motivator for
much AGI research.
Conclusions of the Survey

In the broadest of terms, our results concur with those of the two
previous studies mentioned above. All three studies suggest that
significant numbers of interested, informed individuals believe it
is likely that AGI at the human level or beyond will occur around
the middle of this century, and plausibly even sooner. Due to the
greater depth of the questions, our survey also revealed some
interesting additional information, such as the disagreement
among experts over the likely order of AGI milestones and the
relative safety of different AGI development scenarios. The
experts suggestions regarding funding (with several advocating
63
modest funding increases distributed among a broad range of

research groups) are also potentially valuable.
64
6
What Is General Intelligence?
AGI researchers Shane Legg and Marcus Hutter wrote a paper
in 2007 collecting 70+ different definitions of intelligence from the
scientific literature The ten definitions that most reflect a similar
perspective to Legg and Hutters perspective, and ours, are as
follows:
1. It seems to us that in intelligence there is a fundamental
faculty, the alteration or the lack of which, is of the
utmost importance for practical life. This faculty is
judgment, otherwise called good sense, practical sense,
initiative, the faculty of adapting oneself to
circumstances. A. Binet
2. The capacity to learn or to profit by experience.
3. Ability to adapt oneself adequately to relatively new
situations in life. R. Pinter
4. A person possesses intelligence insofar as he has
learned, or can learn, to adjust himself to his
environment. S. S. Colvin
5. We shall use the term intelligence to mean the ability
of an organism to solve new problems W. V. Bingham
6. A global concept that involves an individuals ability to
act purposefully, think rationally, and deal effectively with
the environment. D. Wechsler
7. Individuals differ from one another in their ability to
understand complex ideas, to adapt effectively to the
environment, to learn from experience, to engage in
various forms of reasoning, to overcome obstacles by
taking thought. American Psychological Association
8. I prefer to refer to it as successful intelligence. And the
reason is that the emphasis is on the use of your
intelligence to achieve success in your life. So I define it
65
as your skill in achieving whatever it is you want to attain

in your life within your sociocultural context meaning
that people have different goals for themselves, and for
some its to get very good grades in school and to do
well on tests, and for others it might be to become a very
good basketball player or actress or musician. R. J.
Sternberg
9. Intelligence is part of the internal environment that
shows through at the interface between person and
external environment as a function of cognitive task
demands. R. E. Snow
10. Certain set of cognitive capacities that enable an
individual to adapt and thrive in any given environment
they find themselves in, and those cognitive capacities
include things like memory and retrieval, and problem
solving and so forth. Theres a cluster of cognitive
abilities that lead to successful adaptation to a wide
range of environments. D. K. Simonton
All these variants roughly say the same sort of thing, but different
researchers have different slants. The most common feature of
these definitions is an emphasis on learning, adaptability, and
interaction with some external environment. Intelligence does not
exist separately from the ability to interact with an environment,
but rather is developed through it. In the view represented here,
there is no magical higher order process that establishes
intelligence independent of experience. A truly intelligent system
is able to observe the environment, identify the problem,
formulate goals, and then devise, implement, and test a solution.
Only being able to reason about a previously defined domain of
situations, problems, and solutions such as a chess computer
does is not such a powerful kind of intelligence, in this sense.
As Legg and Hutter put it: intelligence is not the ability to deal
with a fully known environment, but rather the ability to deal with
some range of possibilities which cannot be wholly anticipated.
In other words, the definition we formulated at the outset: The
66
ability to autonomously achieve complex goals in complex,

changing environments.
Note that nothing in these definitions makes any statements
about emotions. Intelligence does not require the ability to love,
anger, or be bored. Also, the ability to observe a situation is
intentionally stated vaguely. No specific sense vision, audition,
etc. is specifically required in order to be generally intelligent. A
human-like body is not required, nor is any kind of robot body nor
virtual body, etc. This kind of sensorimotor apparatus may be
very valuable for developing and maintaining some particular
sorts of intelligence, including some of value and relevant to
humans, but its not critical to intelligence generally conceived.
Potentially one could have an intelligence whose only world
was what it saw through a Web browser, for example. This would
not be a very human-like intelligence, but it might be human-level
and perhaps even superhuman in some ways.
In general, the connection between the abstract nature of
intelligence and the particularities of human intelligence is
something thats understood only broadly and conceptually, at
present. A full understanding of this connection probably awaits
much fuller development of both AGI and cognitive neuroscience
though Ive done some thinking in this direction, and will share
some of my thoughts below.
The lack of a broadly agreed-upon definition for the I in AI has
long been the source of sardonic humor among AI researchers
and criticism from those outside the field But I have always
thought this criticism largely misplaced. After all, biology lacks a
crisp definition of life, and physics lacks crisp definitions of
space, time or cause. So what? Sometimes foundational
concepts like these elude precise definition, but serve to spawn
other more crisply definable and measurable concepts.
One humorous definition goes: Intelligence is whatever humans
can do that computers cant do yet. There is some truth to this
quip: for instance, most people would say that playing chess
67
requires intelligence. Yet, now that we see Deep Blue playing

grandmaster-beating chess via some simple mathematical
algorithms (rather than through any human-like intuition or
strategy), there is a temptation to reclassify chess-playing as
something on the borderline of intelligence, rather than
something definitively involving intelligence.
A less sarcastic slant on the same basic observation would be to
say that AI research has helped us to clarify our understanding
of what intelligence really is. The fact that various activities (like
chess, navigation, medical diagnosis, or Web search),
apparently requiring considerable general intelligence in
humans, can be carried out by simple specialized algorithms
this fact was not at all obvious from the get-go, and constitutes
a significant discovery on the part of the AI field. And the fact that
similar simple algorithms cannot be used to integrate a broad
understanding of real-world context into the solution of particular
problems, also was not initially obvious, and has been
discovered by the AI field during the last 5 decades. So, as well
as creating a lot of very useful specialized AI programs, the
corpus of AI work done so far has taught us a great deal about
the nature of intelligence and its relationship with computer
technology. Among many other things, we have learned that a
key focus area for AGI research should be integration of
contextual knowledge into problem-solving, and that focusing on
solving narrowly-defined problems is not the right way to
approach general-purpose AI.
A few prescient researchers limned some of these lessons, at
least in sketchy form, even in the early days of the field. In Alan
Turings classic paper from the 1950s, the computing pioneer
proposed what is now called the Turing test a way of gauging
whether one has created an AI with human-level intelligence.
The Turing test basically says: Write a computer program that
can simulate a human in a text-based conversational
interchangeand any human should be willing to consider it
intelligent. The Turing test certainly has its weaknesses but it
does place the focus on something broad-based and requiring
68
conceptual understanding, rather than a specific task or problem.

Turing understood that the crux of human intelligence lies, not in
the specialized tasks that certain clever humans do so much
better than others, but rather in the commonsense processes of
reasoning, learning, communicating and doing that we all carry
out every day.
Indeed, if an AI can chat with educated humans for a couple
hours, and fool them into think its human, then Im comfortable
saying that AI really does have human-level intelligence. That is:
I do buy the Turing Test as a sufficient condition for general
intelligence. However, I dont buy it as a necessary condition. It
seems quite possible that a radically non-human intelligent
computer system could develop, as unable to imitate humans as
we are to imitate dogs or anteaters or sharks, but still just as
intelligent as humans, or even more so.
Also, partial success at tricking humans into believing an AI is
human, seems a particularly poor measure of partial progress
toward human level AI a point that Ill enlarge on later on, in
reviewing some conversations with chat bots that try to imitate
human conversation without any real understanding. So this kind
of trickery does not seem a good way to make a graded AI-Q
score, with current AIs near the bottom and human-level general
intelligence further up. But even outside the scope of Turing Test
style conversation, this kind of graded AI-Q test seems extremely
difficult to come by, given our present level of understanding of
AGI.
Overall, I think I have a fairly good theoretical idea of what
intelligence is all about beginning from the basic idea of
achieving complex goals in complex environments, mentioned
above and Ill say more about that in a moment. But neither I
nor anyone else has a really good IQ test for would-be AGIs. The
creation of such a test might be an interesting task, but I suspect
it cant be effectively approached until we have a population of
fairly similar, fairly advanced AGI systems which we can then
study, to obtain data to guide the creation of AGI intelligence
69
metrics. Human IQ tests will probably guide the development of

such tests only loosely. After all, they work moderately well
within a single culture, but much worse across the spectrum of
human cultures so how much worse will they work across
species, or across different types of computer programs, which
may well be as different as different species of animals?
Defining General Intelligence in Theory and

Practice
The same pair of researchers who collected the 70+ definitions
of intelligence mentioned above Shane Legg and Marcus
Hutter also wrote a paper presenting a formal mathematical
theory of general intelligence. The equations are moderately
hairy but the basic idea is pretty simple. They model an
environment as something that interacts with an intelligent
system, giving it percepts and rewards and accepting actions
from it. And they say that a systems intelligence is its ability at
getting rewarded, averaged over all possible environments. The
trick is that the average is weighted by complexity so that
getting rewarded in a complex environment is weighted more
than getting rewarded in a simple one. The trickier trick is that
theres no objective measure of complexity so that their
4
definition is relative to which complexity measure you use.
These are fairly obvious concepts, and a lot of researchers said
similar things previously (for instance, I did in my 1993 book The
Structure of Intelligence, and Ray Solomonoff did back in the
1960s). But Legg and Hutter crossed the ts and dotted the is
and, leveraging Hutters earlier work, they proved the existence
Formally speaking relative to an assumed Universal Turing Machine.

Also, for the math geeks in the audience: their definition assumes an
exponential decaying complexity measure over the space of
environments, which causes the most complex environments to be
weighted far more than the vast bulk of other ones, a property that may
not be desirable.
70
of a computer program displaying very high intelligence

according to their definition. The only catch is that this program,
called AIXI, requires an infinite amount of computing power to
run. There are also some close approximations to AIXI that can
use a finite amount of computing power but still are totally
impractical to build in reality. Marcus Hutter and a number of
other researchers are now working on more practical
approximations to AIXI an effort that so far has led to some
interesting algorithms and theories, but no pragmatic
breakthroughs in terms of building advanced AGI systems.
I wrote a paper in 2010 that generalizes their definition a bit
extending the kind of complexity weighting involved, optionally
incorporating the amount of computational effort required, and
looking at systems that try to achieve goals represented in their
minds rather than trying to get reward from the environment. I
think these extensions have some technical value, but they dont
radically change the picture. They share the main weakness of
Legg and Hutters approach which is being very mathematical,
and somewhat disconnected from the practicalities of real-world
intelligent systems.
Human beings and other animals are simply not very intelligent
according to this kind of abstract definition. If you put us in an
arbitrary, highly complex environment chosen from the scope
of all mathematically possible environments then we probably
wont do anything very smart, and wont lead a very rewarding
existence. In principle, perhaps, we could cope with any sort of
complex environment you stick us in if we have enough time
and resources. For example, we could potentially build an AGI to
figure out how to cope with the environment and help us out! But
in reality, given finite amounts of time and space, we would
screw up royally in any environment not fairly similar to the ones
that weve evolved to deal with.
In other words, were not so intelligent in terms of average
behavior over all mathematically possible environments rather,
were fairly intelligent in the context of environments of the
71
general type that weve evolved for. Were land animals and
were pretty good at managing situations involving piecing solid
objects together but put us underwater where its all about
subtle interactions between fluid flows, and we become a lot
stupider. 30% of our brains are devoted to vision processing, so
if you put us in a dark place where sonar is the best way to get
around, we become a lot dumber. Were so terrible at
mathematical calculations, that we focus on areas of math where
we can use analogies to things like vision (e.g. geometry) and
language (e.g. logic, algebra) that we are better evolved to think
about. We evolved to handle social situations involving the <300
people in a Stone Age tribe, so we get easily bollixed by the
larger, more diffuse social networks that modern society
presents us with. And so forth.
All this may make you wonder just how general is our
intelligence after all!! But still, theres something to the fact that,
in principle, given enough time and resources, we have the
flexibility of mind to solve just about any problem. And theres no
doubt that our intelligence is dramatically more general than that
of Deep Blue, or a mouse.
Harking back to the formal theory of general intelligence, one
can quantify the generality of an intelligent system by the
entropy (the breadth or spread) of the class of environments in
which its good at achieving goals. A system thats really good at
achieving goals in a narrow class of environments, maybe smart,
but doesnt have so much generality of intelligence. A system
thats moderately good at achieving goals in a broad variety of
environment, may be said to have more generality of
intelligence. Humans have both more intelligence, and more
generality of intelligence, than Deep Blue or a mouse! A mouse
has way more generality of intelligence than Deep Blue, even
though mice are lousy chess players.
You might also question all this emphasis on rewards and goals.
People do all sorts of things, some of which seem to have little to
do with achieving goals, and some of which seem to have little to
72
do with achieving near or long term rewards. Were complex selforganizing systems displaying all manner of different behaviors.
Goals are things we sometimes try to follow, or ascribe to each
other descriptively, rather than things that rigorously govern all
our behaviors. And while sometimes we seek short or long term
rewards, we also engage in a lot of activities that seem to have
little reward at all, for ourselves or anybody. But, if you want to
measure or define a systems intelligence, it still seems the best
way to do this is relative to some collection of goals. Even if not
everything a system does is intelligent, you can still measure its
intelligence via the complexity of the goals it can achieve in
complex environments. Otherwise, its hard to see how to
distinguish intelligence from more general self-organizing
complexity, coupled with the environment which is also
interesting and important, but a broader sort of concept.
So, when we talk about building an artificial general intelligence
in practice, what were mostly talking about is building a system
that in principle could achieve nearly any goal in any
environment given sufficient time and resources; and that in
practice is good at achieving complex goals similar to those
goals needed for the survival and flourishing of a humanlike organism in a human society, in environments like the
ones humans lived in historically or live in presently. Thats
not very compact or elegant, but its what AGI in reality is mainly
about. Eventually the AGI field may branch beyond this, and get
concerned with building broader kinds of AGI systems displaying
varieties of intelligence further detached from human history and
perhaps even incomprehensible to the human mind.
Viewed in this way, the practical AGI task has a lot to do with the
specifics of human bodies, societies and environments which
is somewhat messy and not very mathematical, but seems to be
the reality of the matter. The ideal theory of AGI, in my view,
would be something like a recipe for producing intelligent
systems from environments. You would feed the theory a
description of the environment and goals that an intelligent
system needs to deal with and the theory would spit out a
73
description of what kind of system could be intelligent with

respect to that set of environment and goals, given limited
computing resources. Unfortunately we dont have this kind of
theory right now Ive spent some time trying to formulate it, but
I dont feel particularly close to success at the moment. So we
need to proceed in this direction via a more intuitive approach:
given an intuitive, cross-disciplinary understanding of the
environment and goals pertinent to a human-level, roughly
human-like AGI, integrate all the knowledge we have from
various sources to figure out how to craft an AGI capable of
dealing with these environments and goals.
As a single, important example, consider some of the main types
of memory that cognitive psychologists have identified as being
critical to human intelligence, and the operation of the human
brain:
Declarative memory, for facts and beliefs
Procedural memory, for practical, routinized knowledge

of how to do stuff (the way we can know how to serve a
tennis ball, or how to seduce a girl, or how to prove a
theorem, even though we cant accurately describe in
language exactly what we do to accomplish these
things)
Sensorimotor memory, recalling what we perceive and

enact
Episodic memory, our life-history memory of the stories

weve been through
Why are these types of memory important for human intelligence

and human-like AGI? Because of the kinds of goals and
environments humans have. Our bodies imply an obvious need
for sensorimotor memory. Our social lives imply an obvious need
for episodic memory. Our facility for and habit of linguistic
communication implies a need for declarative memory. Our
ability to learn via practice and communicate via example implies
a need for procedural memory. The particularities of our
74
embodiment and environment leads fairly clearly to the need for

the various types of memory that the human brain possesses.
And then each of these types of memory must be served by
appropriate learning mechanisms a train of thought that has
significant implications for AGI design. But the main point for
now is simply that the specifics of an intelligent systems
cognitive architecture, connect fairly closely to the specifics of
the environment and goals that the system is intended to
primarily deal with. This is how the abstract notion of intelligence
as achieving complex goals in complex environments connects
with the reality of real-world intelligent systems doing real things.
75
7
Why an Intelligence Explosion is
Probable
The path from here to vastly superhuman AGI, Ive suggested,
can meaningfully be broken down into two phases: Phase 1, the
creation of human-level AGI; Phase 2, the path from human-level
AGI into the unknown realm beyond. The second of these
phases is what I.J. Good referred to as the intelligence
explosion, in his famous 1965 quote:
Let an ultra-intelligent machine be defined as a machine that
can far surpass all the intellectual activities of any man
however clever. Since the design of machines is one of
these intellectual activities, an ultra-intelligent machine could
design even better machines; there would then
unquestionably be an intelligence explosion, and the
intelligence of man would be left far behind. Thus the first
ultra-intelligent machine is the last invention that man need
ever make.
Some would argue against the wisdom of isolating this kind of
intelligence explosion as an event unto itself, preferring to
emphasize the continuity of such a potential explosion, not only
with the creation of superhuman AGI, but also with the larger,
longer explosion of intelligence on Earth, beginning with the
origins of life (or maybe even before). But while I think this
broad-scope view is valuable, I also think its important to
recognize the singular nature of the intelligence explosion a la
I.J. Good. Engineering is different from evolution, even though it
was invented by evolved beings in the indirect pursuit of their
evolutionary goals; and the recursive self-improvement of
engineered minds, if it does indeed occur, will have different
characteristics than the natural selection driven evolution of
intelligence that preceded it.
76
Why an Intelligence Explosion is Probable
My own research work is aimed at carrying off Phase 1, the

creation of human-level AGI, in a way that will not only lead to
beneficial results straight away, but also bias the odds toward a
positive version of Phase 2, Goods intelligence explosion. To
me, its always seemed fairly intuitively obvious that, once
human-level AGI is achieved, a Good-style intelligence explosion
will follow. Whether this explosion will take months (seems
unlikely), years (seems probable), decades (maybe), or centuries
(IMO, almost surely not) is less obvious. But the potential of a
digitally-implemented, engineered human-level AGI to analyze,
refine and improve itself, in the a manner leading to exponential
improvement in its general intelligence, seems plain. The human
brain is hard to study and hard to modify, because it wasnt
created with analysis and improvement in mind. Engineered
systems can be instrumented differently, leading to much more
straightforward routes to improvement. No doubt the path from
human-level AGI to dramatically superhuman AGI will involve
many challenges, but, these will surely be qualitatively different
ones from those involved with understanding and improving
human brains, or with achieving human-level AGI in the first
place.
But as obvious as Goods view seems to me and some others,
unsurprisingly, not all futurist thinkers agree. Skeptics often cite
limiting factors that could stop an intelligence explosion from
happening, even after the achievement of human-level AGI; and
in a 2010 post on the Extropy email discussion list, the futurist
Anders Sandberg articulated some of those possible limiting
factors, in a particularly clear way:
One of the things that struck me during our Winter
Intelligence workshop on intelligence explosions was how
confident some people were about the speed of recursive
self-improvement of AIs, brain emulation collectives or
economies. Some thought it was going to be fast in
comparison to societal adaptation and development
timescales (creating a winner takes all situation), some
thought it would be slow enough for multiple superintelligent
77
agents to emerge. This issue is at the root of many key

questions about the singularity (One superintelligence or
many? How much does friendliness matter?).
It would be interesting to hear this list's take on it: what do
you think is the key limiting factor for how fast intelligence
can amplify itself?
1. Economic growth rate
2. Investment availability
3. Gathering of empirical information (experimentation,
interacting with an environment)
4. Software complexity
5. Hardware demands vs. available hardware
6. Bandwidth
7. Lightspeed lags
Clearly many more can be suggested. But which bottlenecks
are the most limiting, and how can this be ascertained?
Following Anders post presenting these issues, the AGI
researcher Richard Loosemore posted a detailed reply
explaining why he thought none of these were such serious
objections. I liked Richards reply so I asked him if I could clean
it up a bit and turn it into an article for H+ Magazine he
graciously agreed, and then it turned into more of a collaborative
writing project, resulting in a jointly authored article with the title
Why An Intelligence Explosion is Probable, elaborating
considerably on Richards original email.
And so, here is the article that Richard and I wrote
78

by Richard Loosemore and Ben Goertzel
[This chapter originally appeared in the form of an H+ Magazine
article, which began with a list of Anders Sandbergs objections
to the intelligence explosion idea, as cited just above.]
We are grateful to Sandberg for presenting this list of questions
because it makes it especially straightforward for us to provide a
clear refutation, in this article, of the case against the viability of
an intelligence explosion. We intend to explain why these
bottlenecks are unlikely to be significant issues, and thus why, as
I.J. Good predicted, an intelligence explosion is indeed a very
likely outcome.
The One Clear Prerequisite for an Intelligence

Explosion
To begin, we need to delimit the scope and background
assumptions of our argument. In particular, it is important to
specify what kind of intelligent system would be capable of
generating an intelligence explosion.
According to our interpretation, there is one absolute prerequisite
for an explosion to occur, and that is that an artificial general
intelligence (AGI) must become smart enough to understand its
own design. In fact, by choosing to label it an artificial general
intelligence we have already said, implicitly, that it will be
capable of self-understanding, since the definition of an AGI is
that it has a broad set of intellectual capabilities that include all
the forms of intelligence that we humans possessand at least
some humans, at that point, would be able to understand AGI
design.
But even among humans there are variations in skill level and
knowledge, so the AGI that triggers the explosion must have a
79
sufficiently advanced intelligence that it can think analytically and

imaginatively about how to manipulate and improve the design of
intelligent systems. It is possible that not all humans are able to
do this, so an AGI that met the bare minimum requirements for
AGI-hoodsay, a system smart enough to be a general
household factotumwould not necessarily have the ability to
work in an AGI research laboratory. Without an advanced AGI of
the latter sort, there would be no explosion, just growth as usual,
because the rate-limiting step would still be the depth and speed
at which humans can think.
The sort of fully-capable AGI were referring to might be called a
seed AGI, but we prefer to use the less dramatic phrase selfunderstanding, human-level AGI. This term, though accurate, is
still rather cumbersome, so we will sometimes use the phrase
the first real AGI or just the first AGI to denote the same idea.
In effect, we are taking the position that for something to be a
proper artificial general intelligence it has to be capable of
competing with the best that the human intellect can achieve,
rather than being limited to a bare minimum. So the first AGI
would be capable of initiating an intelligence explosion.
Distinguishing the Explosion from the Build-Up

Given that the essential prerequisite for an explosion to begin
would be the availability of the first self-understanding, humanlevel AGI, does it make sense to talk about the period leading up
to that arrivalthe period during which that first real AGI was
being developed and trainedas part of the intelligence
explosion proper? We would argue that this is not appropriate,
and that the true start of the explosion period should be
considered to be the moment when a sufficiently well qualified
AGI turns up for work at an AGI research laboratory. This may
be different from the way some others use the term, but it seems
consistent with I.J. Goods original usage. So our concern here is
to argue for the high probability of an intelligence explosion,
80
given the assumption that a self-understanding, human-level AGI

has been created.
By enforcing this distinction, we are trying to avoid possible
confusion with the parallel (and extensive!) debate about
whether a self-understanding, human-level AGI can be built at
all. Questions about whether an AGI with seed level capability
can plausibly be constructed, or how long it might take to arrive,
are of course quite different. A spectrum of opinions on this
issue, from a survey of AGI researchers at a 2009 AGI
5
conference, were presented in a 2010 H+ magazine article . In
that survey, of an admittedly biased sample, a majority felt that
an AGI with this capability could be achieved by the middle of
this century, though a substantial plurality felt it was likely to
happen much further out. While we have no shortage of our own
thoughts and arguments on this matter, we will leave them aside
for the purpose of the present paper.
What Constitutes an Explosion?

How big and how long and how fast would the explosion have to
be to count as an explosion?
Goods original notion had more to do with the explosions
beginning than its end, or its extent, or the speed of its middle or
later phases. His point was that in a short space of time a
human-level AGI would probably explode into a significantly
transhuman AGI, but he did not try to argue that subsequent
improvements would continue without limit. We, like Good, are
primarily interested in the explosion from human-level AGI to an
AGI with, very loosely speaking, a level of general intelligence 23 orders of magnitude greater than the human level (say, 100H
or 1,000H, using 1H to denote human-level general intelligence).
This is not because we are necessarily skeptical of the explosion
http://hplusmagazine.com/2010/02/05/how-long-tillhuman-level-ai/
81
continuing beyond such a point, but rather because pursuing the

notion beyond that seems a stretch of humanitys current
intellectual framework.
Our reasoning, here, is that if an AGI were to increase its
capacity to carry out scientific and technological research, to
such a degree that it was discovering new knowledge and
inventions at a rate 100 or 1,000 times the rate at which humans
now do those things, we would find that kind of world
unimaginably more intense than any future in which humans
were doing the inventing. In a 1,000H world, AGI scientists could
go from high-school knowledge of physics to the invention of
relativity in a single day (assuming, for the moment, that the
factor of 1,000 was all in the speed of thoughtan assumption
we will examine in more detail later). That kind of scenario is
dramatically different from a world of purely human inventiveness
no matter how far humans might improve themselves in the
future, without AGI, its seems unlikely there will ever be a time
when a future Einstein would wake up one morning with a childs
knowledge of science and then go on to conceive the theory of
relativity by the following dayso it seems safe to call that an
intelligence explosion.
This still leaves the question of how fast it has to arrive, to be
considered explosive. Would it be enough for the first AGI to go
from 1H to 1,000H in the course of a century, or does it have to
happen much quicker, to qualify?
Perhaps there is no need to rush to judgment on this point. Even
a century-long climb up to the 1,000H level would mean that the
world would be very different for the rest of history. The simplest
position to take, we suggest, is that if the human species can get
to the point where it is creating new types of intelligence that are
themselves creating intelligences of greater power, then this is
something new in the world (because at the moment all we can
do is create human babies of power 1H), so even if this process
happened rather slowly, it would still be an explosion of sorts. It
82
might not be a Big Bang, but it would at least be a period of

Inflation, and both could eventually lead to a 1,000H world.
Defining Intelligence (Or Not)

To talk about an intelligence explosion, one has to know what
one means by intelligence as well as by explosion. So its
worth reflecting that there are currently no measures of general
intelligence that are precise, objectively defined and broadly
extensible beyond the human scope.
However, since intelligence explosion is a qualitative concept,
we believe the commonsense qualitative understanding of
intelligence suffices. We can address Sandbergs potential
bottlenecks in some detail without needing a precise measure,
and we believe that little is lost by avoiding the issue. We will say
that an intelligence explosion is something with the potential to
create AGI systems as far beyond humans as humans are
beyond mice or cockroaches, but we will not try to pin down
exactly how far away the mice and cockroaches really are.
Key Properties of the Intelligence Explosion

Before we get into a detailed analysis of the specific factors on
Sandbergs list, some general comments are in order.
Inherent Uncertainty. Although we can try our best to understand
how an intelligence explosion might happen, the truth is that
there are too many interactions between the factors for any kind
of reliable conclusion to be reached. This is a complex-system
interaction in which even the tiniest, least-anticipated factor may
turn out to be either the rate-limiting step or the spark that starts
the fire. So there is an irreducible uncertainty involved here, and
we should be wary of promoting conclusions that seem too firm.
General versus Special Arguments. There are two ways to
address the question of whether or not an intelligence explosion
83
is likely to occur. One is based on quite general considerations.

The other involves looking at specific pathways to AGI. An AGI
researcher (such as either of the authors) might believe they
understand a great deal of the technical work that needs to be
done to create an intelligence explosion, so they may be
confident of the plausibility of the idea for that reason alone. We
will restrict ourselves here to the first kind of argument, which is
easier to make in a relatively non-controversial way, and leave
aside any factors that might arise from our own understanding
about how to build an AGI.
The Bruce Wayne Scenario. When the first self-understanding,
human-level AGI system is built, it is unlikely to be the creation of
a lone inventor working in a shed at the bottom of the garden,
who manages to produce the finished product without telling
anyone. Very few of the lone inventor (or Bruce Wayne)
scenarios seem plausible. As communication technology
advances and causes cultural shifts, technological progress is
increasingly tied to rapid communication of information between
various parties. It is unlikely that a single inventor would be able
to dramatically outpace multi-person teams working on similar
projects; and also unlikely that a multi-person team would
successfully keep such a difficult and time-consuming project
secret, given the nature of modern technology culture.
Unrecognized Invention. It also seems quite implausible that the
invention of a human-level, self-understanding AGI would be
followed by a period in which the invention just sits on a shelf
with nobody bothering to pick it up. The AGI situation would
probably not resemble the early reception of inventions like the
telephone or phonograph, where the full potential of the invention
was largely unrecognized. We live in an era in which practicallydemonstrated technological advances are broadly and
enthusiastically communicated, and receive ample investment of
dollars and expertise. AGI receives relatively little funding now,
for a combination of reasons, but it is implausible to expect this
situation to continue in the scenario where highly technically
capable human-level AGI systems exist. This pertains directly to
84
the economic objections on Sandbergs list, as we will elaborate

below.
Hardware Requirements. When the first human-level AGI is
developed, it will either require a supercomputer-level of
hardware resources, or it will be achievable with much less. This
is an important dichotomy to consider, because world-class
supercomputer hardware is not something that can quickly be
duplicated on a large scale. We could make perhaps hundreds of
such machines, with a massive effort, but probably not a million
of them in a couple of years.
Smarter versus Faster.
There are two possible types of
intelligence speedup: one due to faster operation of an intelligent
system (clock speed increase) and one due to an improvement
in the type of mechanisms that implement the thought processes
(depth of thought increase). Obviously both could occur at once
(and there may be significant synergies), but the latter is
ostensibly more difficult to achieve, and may be subject to
fundamental limits that we do not understand. Speeding up the
hardware, on the other hand, is something that has been going
on for a long time and is more mundane and reliable. Notice that
both routes lead to greater intelligence, because even a human
level of thinking and creativity would be more effective if it were
happening a thousand times faster than it does now.
It seems quite possible that the general class of AGI systems
can be architected to take better advantage of improved
hardware than would be the case with intelligent systems very
narrowly imitative of the human brain. But even if this is not the
case, brute hardware speedup can still yield dramatic intelligent
improvement.
Public Perception. The way an intelligence explosion presents
itself to human society will depend strongly on the rate of the
explosion in the period shortly after the development of the first
self-understanding human-level AGI. For instance, if the first
such AGI takes five years to double its intelligence, this is a
85
very different matter than if it takes two months. A five-year time

frame could easily arise, for example, if the first AGI required an
extremely expensive supercomputer based on unusual
hardware, and the owners of this hardware were to move slowly.
On the other hand, a two-month time frame could more easily
arise if the initial AGI were created using open source software
and commodity hardware, so that a doubling of intelligence only
required addition of more hardware and a modest number of
software changes. In the former case, there would be more time
for governments, corporations and individuals to adapt to the
reality of the intelligence explosion before it reached dramatically
transhuman levels of intelligence. In the latter case, the
intelligence explosion would strike the human race more
suddenly. But this potentially large difference in human
perception of the events would correspond to a fairly minor
difference in terms of the underlying processes driving the
intelligence explosion.
So now, finally, with all the preliminaries behind us, we will
move on to deal with the specific factors on Sandbergs list, one
by one, explaining in simple terms why each is not actually likely
to be a significant bottleneck. There is much more that could be
said about each of these, but our aim here is to lay out the main
points in a compact way.
Objection 1: Economic Growth Rate and

Investment Availability
The arrival, or imminent arrival, of human-level, selfunderstanding AGI systems would clearly have dramatic
implications for the world economy. It seems inevitable that
these dramatic implications would be sufficient to offset any
factors related to the economic growth rate at the time that AGI
began to appear. Assuming the continued existence of
technologically advanced nations with operational technology
R&D sectors, if self-understanding human-level AGI is created,
then it will almost surely receive significant investment. Japans
economic growth rate, for example, is at the present time
86
somewhat stagnant, but there can be no doubt that if any kind of

powerful AGI were demonstrated, significant Japanese
government and corporate funding would be put into its further
development.
And even if it were not for the normal economic pressure to
exploit the technology, international competitiveness would
undoubtedly play a strong role. If a working AGI prototype were
to approach the level at which an explosion seemed possible,
governments around the world would recognize that this was a
critically important technology, and no effort would be spared to
produce the first fully-functional AGI before the other side does.
Entire national economies might well be sublimated to the goal of
developing the first superintelligent machine, in the manner of
Project Apollo in the 1960s. Far from influencing the intelligence
explosion, economic growth rate would be defined by the various
AGI projects taking place around the world.
Furthermore, it seems likely that once a human-level AGI has
been achieved, it will have a substantialand immediate
practical impact on multiple industries. If an AGI could
understand its own design, it could also understand and improve
other computer software, and so have a revolutionary impact on
the software industry. Since the majority of financial trading on
the US markets is now driven by program trading systems, it is
likely that such AGI technology would rapidly become
indispensable to the finance industry (typically an early adopter
of any software or AI innovations). Military and espionage
establishments would very likely also find a host of practical
applications for such technology. So, following the achievement
of self-understanding, human-level AGI, and complementing the
allocation of substantial research funding aimed at outpacing the
competition in achieving ever-smarter AGI, there is a great
likelihood of funding aimed at practical AGI applications, which
would indirectly drive core AGI research along.
The details of how this development frenzy would play out are
open to debate, but we can at least be sure that the economic
87
growth rate and investment climate in the AGI development

period would quickly become irrelevant.
However, there is one interesting question left open by these
considerations. At the time of writing, AGI investment around the
world is noticeably weak, compared with other classes of
scientific and technological investment. Is it possible that this
situation will continue indefinitely, causing so little progress to be
made that no viable prototype systems are built, and no
investors ever believe that a real AGI is feasible?
This is hard to gauge, but as AGI researchers ourselves, our
(clearly biased) opinion is that a permanent winter scenario is
too unstable to be believable. Because of premature claims
made by AI researchers in the past, a barrier to investment
clearly exists in the minds of todays investors and funding
agencies, but the climate already seems to be changing. And
even if this apparent thaw turns out to be illusory, we still find it
hard to believe that there will not eventually be an AGI
investment episode comparable to the one that kicked the
internet into high gear in the late 1990s. Furthermore, due to
technology advanced in allied fields (computer science,
programming language, simulation environments, robotics,
computer hardware, neuroscience, cognitive psychology, etc.),
the amount of effort required to implement advanced AGI
designs is steadily decreasing so that as time goes on, the
amount of investment required to get AGI to the explosionenabling level will keep growing less and less.
Objection 2: Inherent Slowness of Experiments

and Environmental Interaction
This possible limiting factor stems from the fact that any AGI
capable of starting the intelligence explosion would need to do
some experimentation and interaction with the environment in
order to improve itself. For example, if it wanted to re-implement
itself on faster hardware (most probably the quickest route to an
88
intelligence increase) it would have to set up a hardware

research laboratory and gather new scientific data by doing
experiments, some of which might proceed slowly due to
limitations of experimental technology.
The key question here is this: how much of the research can be
sped up by throwing large amounts of intelligence at it? This is
closely related to the problem of parallelizing a process (which is
to say: You cannot make a baby nine times quicker by asking
nine women to be pregnant for one month). Certain algorithmic
problems are not easily solved more rapidly simply by adding
more processing power, and in much the same way there might
be certain crucial physical experiments that cannot be hastened
by doing a parallel set of shorter experiments.
This is not a factor that we can understand fully ahead of time,
because some experiments that look as though they require
fundamentally slow physical processeslike waiting for a silicon
crystal to grow, so we can study a chip fabrication mechanism
may actually be dependent on the intelligence of the
experimenter, in ways that we cannot anticipate. It could be that
instead of waiting for the chips to grow at their own speed, the
AGI could do some clever micro-experiments that yield the same
information faster.
The increasing amount of work being done on nanoscale
engineering would seem to reinforce this pointmany processes
that are relatively slow today could be done radically faster using
nanoscale solutions. And it is certainly feasible that advanced
AGI could accelerate nanotechnology research, thus initiating a
virtuous cycle where AGI and nanotech research respectively
push each other forward (as foreseen by nanotech pioneer Josh
6
Hall ). As current physics theory does not even rule out more
http://hplusmagazine.com/2011/01/24/towardintelligent-nano-factories-and-fogs/
89

7
outlandish possibilities like femtotechnology , it certainly does

not suggest the existence of absolute physical limits on
experimentation speed existing anywhere near the realm of
contemporary science.
Clearly, there is significant uncertainty in regards to this aspect
of future AGI development. One observation, however, seems to
cut through much of the uncertainty. Of all the ingredients that
determine how fast empirical scientific research can be carried
out, we know that in todays world the intelligence and thinking
speed of the scientists themselves must be one of the most
important. Anyone involved with science and technology R&D
would probably agree that in our present state of technological
sophistication, advanced research projects are strongly limited
by the availability and cost of intelligent and experienced
scientists.
But if research labs around the world have stopped throwing
more scientists at problems they want to solve, because the
latter are unobtainable or too expensive, would it be likely that
those research labs are also, quite independently, at the limit for
the physical rate at which experiments can be carried out? It
seems hard to believe that both of these limits would have been
reached at the same time, because they do not seem to be
independently optimizable. If the two factors of experiment speed
and scientist availability could be independantly optimized, this
would mean that even in a situation where there was a shortage
of scientists, we could still be sure that we had discovered all of
the fastest possible experimental techniques, with no room for
inventing new, ingenious techniques that get over the physicalexperiment-speed limits. In fact, however, we have every reason
to believe that if we were to double the number of scientists on
the planet at the moment, some of them would discover new
ways to conduct experiments, exceeding some of the current
http://hplusmagazine.com/2011/01/10/theres-plentymore-room-bottom-beyond-nanotech-femtotech
90
speed limits. If that were not true, it would mean that we had
quite coincidentally reached the limits of science talent and
physical speed of data collecting at the same timea
coincidence that we do not find plausible.
This picture of the current situation seems consistent with
anecdotal reports: Companies complain that research staff are
expensive and in short supply; they do not complain that nature
is just too slow. It seems generally accepted, in practice, that
with the addition of more researchers to an area of inquiry,
methods of speeding up and otherwise improving processes can
be found.
So based on the actual practice of science and engineering
today (as well as known physical theory), it seems most likely
that any experiment speed limits lie further up the road, out of
sight. We have not reached them yet, and we lack any solid
basis for speculation about exactly where they might be.
Overall, it seems we do not have concrete reasons to believe
that this will be a fundamental limit that stops the intelligence
explosion from taking an AGI from H (human-level general
intelligence) to (say) 1,000 H. Increases in speed within that
range (for computer hardware, for example) are already
expected, even without large numbers of AGI systems helping
out, so it would seem that physical limits, by themselves, would
be very unlikely to stop an explosion from 1H to 1,000 H.
Objection 3: Software Complexity

This factor is about the complexity of the software that an AGI
must develop in order to explode its intelligence. The premise
behind this supposed bottleneck is that even an AGI with selfknowledge finds it hard to cope with the fabulous complexity of
the problem of improving its own software.
This seems implausible as a limiting factor, because the AGI
could always leave the software alone and develop faster
91
hardware. So long as the AGI can find a substrate that gives it a

thousand-fold increase in clock speed, we have the possibility for
a significant intelligence explosion.
Arguing that software complexity will stop the first selfunderstanding, human-level AGI from being built is a different
matter. It may stop an intelligence explosion from happening by
stopping the precursor events, but we take that to be a different
type of question. As we explained earlier, one premise of the
present analysis is that an AGI can actually be built. It would take
more space than is available here to properly address that
question.
It furthermore seems likely that, if an AGI system is able to
comprehend its own software as well as a human being can, it
will be able to improve that software significantly beyond what
humans have been able to do. This is because in many ways,
digital computer infrastructure is more suitable to software
development than the human brains wetware. And AGI software
may be able to interface directly with programming language
interpreters, formal verification systems and other programmingrelated software, in ways that the human brain cannot. In that
way the software complexity issues faced by human
programmers would be significantly mitigated for human-level
AGI systems. However, this is not a 100% critical point for our
arguments, because even if software complexity remains a
severe difficulty for a self-understanding, human-level AGI
system, we can always fall back to arguments based on clock
speed.
Objection 4: Hardware Requirements

We have already mentioned that much depends on whether the
first AGI requires a large, world-class supercomputer, or whether
it can be done on something much smaller.
This is something that could limit the initial speed of the
explosion, because one of the critical factors would be the
92
number of copies of the first AGI that can be created. Why would
this be critical? Because the ability to copy the intelligence of a
fully developed, experienced AGI is one of the most significant
mechanisms at the core of an intelligence explosion. We cannot
do this copying of adult, skilled humans, so human geniuses
have to be rebuilt from scratch every generation. But if one AGI
were to learn to be a world expert in some important field, it
could be cloned any number of times to yield an instant
community of collaborating experts.
However, if the first AGI had to be implemented on a
supercomputer, that would make it hard to replicate the AGI on a
huge scale, and the intelligence explosion would be slowed
down because the replication rate would play a strong role in
determining the intelligence-production rate.
However, as time went on, the rate of replication would grow, as
hardware costs declined. This would mean that the rate of arrival
of high-grade intelligence would increase in the years following
the start of this process. That intelligence would then be used to
improve the design of the AGIs (at the very least, increasing the
rate of new-and-faster-hardware production), which would have
a positive feedback effect on the intelligence production rate.
So if there was a supercomputer-hardware requirement for the
first AGI, we would see this as something that would only
dampen the initial stages of the explosion. Positive feedback
after that would eventually lead to an explosion anyway.
If, on the other hand, the initial hardware requirements turn out to
be modest (as they could very well be), the explosion would
come out of the gate at full speed.
Objection 5: Bandwidth
In addition to the aforementioned cloning of adult AGIs, which
would allow the multiplication of knowledge in ways not currently
available in humans, there is also the fact that AGIs could
93
communicate with one another using high-bandwidth channels.

This is inter-AGI bandwidth, and it is one of the two types of
bandwidth factors that could affect the intelligence explosion.
Quite apart from the communication speed between AGI
systems, there might also be bandwidth limits inside a single
AGI, which could make it difficult to augment the intelligence of a
single system. This is intra-AGI bandwidth.
The first oneinter-AGI bandwidthis unlikely to have a strong
impact on an intelligence explosion because there are so many
research issues that can be split into separately-addressable
components. Bandwidth between the AGIs would only become
apparent if we started to notice AGIs sitting around with no work
to do on the intelligence amplification project, because they had
reached an unavoidable stopping point and were waiting for
other AGIs to get a free channel to talk to them. Given the
number of different aspects of intelligence and computation that
could be improved, this idea seems profoundly unlikely.
Intra-AGI bandwidth is another matter. One example of a
situation in which internal bandwidth could be a limiting factor
would be if the AGIs working memory capacity were dependent
on the need for total connectivityeverything connected to
everything elsein a critical component of the system. If this
case, we might find that we could not boost working memory
very much in an AGI because the bandwidth requirements would
increase explosively. This kind of restriction on the design of
working memory might have a significant effect on the systems
depth of thought.
However, notice that such factors may not inhibit the initial phase
of an explosion, because the clock speed, not the depth of
thought, of the AGI may be improvable by several orders of
magnitude before bandwidth limits kick in. The main element of
the reasoning behind this is the observation that neural signal
speed is so slow. If a brain-like AGI system (not necessarily a
whole brain emulation, but just something that replicated the
94
high-level functionality of the brain) could be built using

components that kept the same type of processing demands,
and the same signal speed as neurons, then we would be
looking at a human-level AGI in which information packets were
being exchanged once every millisecond. In that kind of system
there would then be plenty of room to develop faster signal
speeds and increase the intelligence of the system. The
processing elements would also have to go faster, if they were
not idling, but the point is that the bandwidth would not be the
critical problem.
Objection 6: Lightspeed Lags

Here we need to consider the limits imposed by special relativity
on the speed of information transmission in the physical
universe. However, its implications in the context of AGI are not
much different than those of bandwidth limits.
Lightspeed lags could be a significant problem if the components
of the machine were physically so far apart that massive
amounts of data (by assumption) were delivered with a
significant delay. But they seem unlikely to be a problem in the
initial few orders of magnitude of the explosion. Again, this
argument derives from what we know about the brain. We know
that the brain's hardware was chosen due to biochemical
constraints.
We are carbon-based, not silicon-and-copperbased, so there are no electronic chips in the head, only pipes
filled with fluid and slow molecular gates in the walls of the pipes.
But if nature was forced to use the pipes-and-ion-channels
approach, that leaves us with plenty of scope for speeding things
up using silicon and copper (and this is quite apart from all the
other more exotic computing substrates that are now on the
horizon). If we were simply to make a transition membrane
depolarization waves to silicon and copper, and if this produced
a 1,000x speedup (a conservative estimate, given the intrinsic
difference between the two forms of signalling), this would be an
explosion worthy of the name.
95
The main circumstance under which this reasoning would break

down would be if, for some reason, the brain is limited on two
fronts simultaneously: both by the carbon implementation and by
the fact that other implementations of the same basic design are
limited by disruptive light-speed delays. This would mean that all
non-carbon-implementations of the brain take us up close to the
lightspeed limit before we get much of a speedup over the brain.
This would require a coincidence of limiting factors (two limiting
factors just happening to kick in at exactly the same level), that
we find quite implausible, because it would imply a rather bizarre
situation in which evolution tried both the biological neuron
design, and a silicon implementation of the same design, and
after doing a side-by-side comparison of performance, chose the
one that pushed the efficiency of all the information transmission
mechanisms up to their end stops.
Objection 7: Human-Level Intelligence May

Require Quantum (Or More Exotic) Computing
Finally we consider an objection not on Sandbergs list, but
raised from time to time in the popular and even scientific
literature. The working assumption of the vast majority of the
contemporary AGI field is that human-level intelligence can
eventually be implemented on digital computers, but the laws of
physics as currently understood imply that, to simulate certain
physical systems without dramatic slowdown, requires special
physical systems called quantum computers rather than
ordinary digital computers.
There is currently no evidence that the human brain is a system
of this nature. Of course the brain has quantum mechanics at its
underpinnings, but there is no evidence that it displays quantum
coherence at the levels directly relevant to human intelligent
behavior. In fact our current understanding of physics implies
that this is unlikely, since quantum coherence has not yet been
observed in any similarly large and wet system. Furthermore,
even if the human brain were shown to rely to some extent on
96
quantum computing, this wouldnt imply that quantum computing

is necessary for human-level intelligence there are often many
different ways to solve the same algorithmic problem. And (the
killer counterargument), even if quantum computing were
necessary for human-level general intelligence, that would
merely delay the intelligence explosion a little, while suitable
quantum computing hardware was developed. Already the
development of such hardware is the subject of intensive R&D.
Roger Penrose, Stuart Hameroff and a few others have argued
that human intelligence may even rely on some form of quantum
gravity computing, going beyond what ordinary quantum
computing is capable of. This is really a complete blue-sky
speculation with no foundation in current science, so not worth
discussing in detail; but broadly speaking, this claim may be
treated according to the same arguments as weve presented
above regarding quantum computing.
The Path from AGI to Intelligence Explosion

Seems Clear
Summing up, then the conclusion of our relatively detailed
analysis of Sandbergs objections is that there is currently no
good reason to believe that once a human-level AGI capable of
understanding its own design is achieved, an intelligence
explosion will fail to ensue.
The operative definition of intelligence explosion that we have
assumed here involves an increase of the speed of thought (and
perhaps also the depth of thought) of about two or three orders
of magnitude. If someone were to insist that a real intelligence
explosion had to involve million-fold or trillion-fold increases in
intelligence, we think that no amount of analysis, at this stage,
could yield sensible conclusions. But since an AGI with
intelligence = 1000 H might well cause the next thousand years
of new science and technology to arrive in one year (assuming
that the speed of physical experimentation did not become a
significant factor within that range), it would be churlish, we think,
97
not to call that an explosion. An intelligence explosion of such

magnitude would bring us into a domain that our current science,
technology and conceptual framework are not equipped to deal
with; so prediction beyond this stage is best done once the
intelligence explosion has already progressed significantly.
Of course, even if the above analysis is correct, there is a great
deal we do not understand about the intelligence explosion, and
many of these particulars will remain opaque until we know
precisely what sort of AGI system will launch the explosion. But
our view is that the likelihood of transition from a selfunderstanding human-level AGI to an intelligence explosion
should not presently be a subject of serious doubt.
98
8
Ten Common Objections to AGI And
Why They Dont Scare Me
(originally an H+ Magazine article)
Ive heard a lot of supposed reasons why AGI cant be
engineered, or why this wont happen soon. Some of them dont
deserve to be dignified by any response. But some are
reasonably sensible, even though I dont quite buy them.
This chapter gives a quick list of ten of the commoner objections
I've heard, with each objection followed by my own -- highly
opinionated and openly AGI-optimist perspective. In the
interest of concision and punchiness, I won't take many pains to
explicitly justify my take on each objection here the
justifications are given elsewhere in the book!
1. Objection via Quantum Computing
The Claim: The brain is a quantum system, in the strong sense
that its intelligent behaviors rely on macroscopic quantum
coherence among its parts. If this is the case, emulating brain
function on a classical digital computer would be extremely
inefficient (though possible in principle).
My take: There's no evidence this is the case, and it would
require some revolutionary changes in the science of
macroscopic quantum coherence. However, this claim doesn't
seem to violate known science in a totally egregious way.
But even if it's true, all this means is that we might need to
engineer an AGI to run on quantum computers rather than digital
computers Which would delay the advent of AGI but not pose
a fundamental obstacle.
99
Also note that, even if the brain uses quantum computing, that
doesn't mean a human-level AGI also needs to. There are often
many different ways to achieve the same engineering functions.
Note that the brain is well known to make use of various
quantum phenomena in the lower levels of operation like every
other physical system. That's not the same as manifesting
macroscopic quantum coherence in the way that would require
use of a quantum computer for efficient brain emulation.
2. Objection via Hypercomputing
Even further out
8
The Claim: The brain is a hypercomputer , which produces

intelligent behaviors via computing functions that no conventional
computer can compute! (Quantum computers can't do this, they
can only compute certain functions faster than conventional
computers.)
9
Roger Penrose and Stuart Hameroff

have famously
hypothesized that the brain works using "quantum gravity
computing" which makes use of hypercomputing. However, they
haven't posited any specific quantum gravity theory that actually
needs hypercomputing and explains any of the particulars of
brain function.
My take: Not only is there no evidence for this objection, there
are also conceptual problems regarding the very notion of
evidence for hypercomputing.
Any finite set of finite-precision data (i.e. any scientific data, in
the vein of current or prior understandings of science) can be
explained by some computational model; furthermore, no
specific hypercomputational model can ever be described in
8
9
http://en.wikipedia.org/wiki/Hypercomputation
http://www.quantumconsciousness.org/
100
Ten Common Objections to AGI And Why They Dont Scare Me
detail in a finite list of words or mathematical symbols.

Nevertheless, it's conceivable that some hypercomputational
model could be used to qualitatively explain neural data, and that
a community of people could collectively recognize it as the
simplest explanation of that data (even though they couldn't
communicate exactly why in finite series of linguistic or
mathematical symbols). But this is pretty weird speculative
territory, and there is not any evidence that anything like this
could work. Right now the hypercomputable mind is just outthere speculation, independent of any specific, detailed analyses
of specific neural or mental phenomena.
Clearly this sort of funky speculation cant be taken as a serious
objection to engineering AGI!
3. Objection via the Specialness of Biology
The Claim: Even if AGI doesn't require quantum computing or
hypercomputing, it requires a sufficiently special computing
infrastructure that it's not pragmatically feasible to achieve it
using ordinary computing hardware (even if possible in principle
according to computing theory). One really needs to use certain
kinds of analogue or chemical computers to get human-level AGI
with feasible efficiency.
My take: It's possible. However, it's not the message you get
from the bulk of the neuroscience literature today, and especially
not
from
the
computational
neuroscience
literature.
Neuroscientists model neurons using differential equations
models that run on ordinary computers; and other cells like glia
also seem susceptible to such modeling. Henri Markram,
Kwabena Boahen and many other computational neuroscientists
are working on building bigger and bigger and better and better
computational neuroscience models. Indeed, this whole branch
of research might fail science is never certain -- but at the
101
moment it seems to be advancing quite successfully and rapidly

10
(see a review of the literature that I co-authored ).
4. Objection via Complexity
The Claim: The human brain is a complex system, and each part
is finely tuned to work with the other parts. There's no viable
way to analytically determine how the behavior of the whole
depends on the behaviors of the parts, because this was created
via evolution-guided self-organization, which often tends to
create messy complex systems that are hard to analyze precise.
So a brain-like system is simply too complex (not just
complicated, but complex in the sense of complex systems
theory) to design and engineer using any of our commonly
recognized engineering methodologies.
According to this view, detailed brain emulation could potentially
work, but only if we had a good enough theory of brain and mind
to know what level the emulation should focus on, so as to be
able to intelligently tweak and test the system. Or, systems
deviating further from brains could potentially lead to AGI, but
only if we developed some radically new method of complex
systems evolution/engineering, different from any known
engineering methodology.
AGI researcher Richard Loosemore has championed this view
on AGI email lists for some time now; see e.g. his Complex
11
Cognitive Systems Manifesto .
My take: Yeah, the brain is complex, but not that complex. I
think we will understand it analytically once we have better brain
imaging data, enabled by better imaging tech. And, only some of
that complexity is necessary for AGI, some of it is just the way
10
http://www.informatik.unitrier.de/~ley/db/journals/ijon/ijon74.html
11
http://www.richardloosemore.com/papers
102
the brain happened to evolve. An AGI can be engineered to be

more judicious in the complexity that it manifests.
However, the idea of new methods of creating complex systems,
combining aspects of engineering and artificial evolution, seems
interesting and may play a role in the creation of AGI somehow.
5. Objection via Difficulty and Lack of Resources
The Claim: Even if none of the above arguments hold -- so that
it's plausible to engineer an AGI via clever deployment of
mathematics, science and engineering principles AGI still
probably won't happen just because it's really hard and society
doesn't choose to allocate many resources toward this task.
I think this is possible, but I'm working pretty hard to make it
false. My thought is that once some AGI project demonstrates
AGI with a certain threshold level of functionality, this will serve
12
as a sort of "AGI Sputnik " and get the world excited about
devoting financial and attentional resources to AGI.
The question then becomes if it's too hard to get to that Sputnik
level. I'm more optimistic than most AGI researchers, I admit: My
own bet is that producing an AGI with the rough intelligence of a
3-4 year old human child is feasible within a 4-5 years of work by
a team of 15 or so good AI programmers (see the OpenCog
13
Roadmap ). The key of course is the right AGI architecture, and
I believe the OpenCog system that I co-created is one viable
option.
But setting aside my particular optimism about the potential of
my own work, the broader question is whether some AGISputnik type achievement is feasible given relatively limited
resources and note that the cost of computing hardware, and
12
http://hplusmagazine.com/2011/03/30/seeking-thesputnik-of-agi/
13
http://opencog.org/roadmap/
103
the ease of producing complex software, are improving rapidly

year on year. If so, then this objection fails.
6.Objection via Vagueness of the Goal
The Claim: The scientific community has no clear, commonly
accepted definition of what "Human-level AGI" or "Transhuman
AGI" really means. How can people build something if they don't
know what they're building?
My take: "Intelligence" is a fuzzy, natural language term, without
a single clear meaning. There is a mathematical definition of
14
general intelligence , posited by Shane Legg and Marcus
Hutter, which I've extended into a more practical formal theory of
15
general intelligence . Not everyone in the AGI community
accepts this definition as the best one, but so what? The point is
that the AGI field does have concrete formalizations of its goals.
And furthermore, pragmatic goals may be more useful than
formal ones anyway. What about the goal of making an AGI that
can attend Stanford and get a degree, being treated the same
way as an ordinary college student; or that can graduate from a
purely online university? These are examples of concrete goals
for AGI research. So what if everyone in the field doesn't work
toward the exact same goals, instead working toward somewhat
related but different goals?
The problem with this objection is, nobody ever gives a strong
justification for why it's necessary for everyone in the AGI field to
be working toward the same precisely defined goals, in order of
the field to make progress. Yes, a shared crisply-defined goal
14
http://www.vetta.org/documents/UniversalIntelligence.
pdf
15
http://multiverseaccordingtoben.blogspot.com/2011/03/
toward-general-theory-of-feasible.html
104
would be a nicety (and I've put some effort into working toward
16
one, e.g. via co-organizing the 2009 AGI Roadmap Workshop
at the University of Tennessee, Knoxville), but why is it posited
as a necessity?
7. Objection via Consciousness
The Claim: We humans seem to be conscious; our laptops seem
not to be conscious in this same sense; human consciousness
seems associated with human intelligence; therefore, computers
can never be intelligent like humans.
My take: Philosophers have nothing near a consensus on what
consciousness is. Furthermore, philosophers have never come
up with a compelling argument as to why any of us should
consider our friends, wives or children conscious (as opposed to
being purely solipsistic). So it seems most sensible to ignore
these philosophical objections as regards AGI, until such point
as a good theory of consciousness emerges. I can easily
envision sitting around in a cafe' with some AGI robots debating
the same old philosophical issues about who is, or is not,
conscious.
My own personal philosophy of consciousness tends to be
17
panpsychist I think everything in the universe is conscious to
an extent, and each sort of system manifests this universal
consciousness differently. I strongly suspect that if we build a
digital mind with similar structures and dynamics to the human
one, it will manifest consciousness in similar ways.
8. Objection via Free Will
The Claim: Humans have free will, were not just deterministic
automata. But a computer has no free will, it just does what its
programming tells it. Human free will is closely tied with human
16
http://web.eecs.utk.edu/~itamar/AGI_Roadmap.html
http://cosmistmanifesto.blogspot.com/2009/05/blogpost.html
17
105
intelligence. Therefore, computers can never be intelligent like

humans.
This notion is sometimes tied up with hypercomputing, because
hypercomputing goes beyond digital computing and is hence
18
perceived as providing a sort of free will (Selmer Bringsjord has
taken this perspective, for example).
My take: First of all, cognitive neuroscience has convincingly
demonstrated that the human feeling of "free will" is almost
19
entirely illusory . Secondly, a large, complex computer system is
not deterministic from a human perspective, in practice, so our
philosophy of determinism may need some revision here (see
20
my blog post on quantum models of classical systems ). And a
large complex computer system becomes even less
deterministic when coupled with the external world, as any
realistic AGI is going to be.
9. Objection via Near-Certain Doom
The Claim: Even if we can engineer AGI, we shouldn't, because
once it's completed it's going to kill us all, or enslave us, or use
us for batteries, or whatever.
Some futurists, like Hugo de Garis, think AGI will almost surely
21
eliminate humanity , but also think we should create AGI
anyway, because it's grand and wonderful and the next step in
the evolution of mind.
18
http://homepages.rpi.edu/~brings/
19
http://www.physorg.com/news186830615.html
20
quantum-logic-models-of-classical.html
21
http://www.forbes.com/2009/06/18/cosmist-terrancyborgist-opinions-contributors-artificialintelligence-09-hugo-de-garis.html
106
A variant on this idea is what I've called The Singularity

Institutes Scary Idea the claim, frequently put forth by various
22
folks associated with SIAI , that any advanced AGI, unless it's
specifically engineered to be "provably Friendly", is almost
23
certain to kill all humans. I've pointed out in a long blog post
that nobody has given a clear rational argument for this position.
10. Objection via Irreducible Uncertainty
The Claim: Even if advanced AGI won't necessarily kill us all,
we cant really eliminate the possibility that it might. No matter
how carefully we engineer our AGIs ethical system and its
overall dynamics, there's still the possibility that it will do
something unpredictable (which could possibly be bad). After all,
how can we expect to precisely predict the doings of an AGI
system that's as smart as us but significantly different, or
significantly more intelligent than us?
I don't have a strong refutation to this objection. I think it's
substantially correct there is irreducible uncertainty, both risk
and reward that are damn hard to estimate, in a project as
radical as engineering advanced AGI. I also think our best
course is just to embrace this uncertainty, and try to minimize it
as best we can, and try to bias developments in a positive
direction. My view is that the development beyond humanity to
more intelligent engineered minds is almost inevitable, and none
of us would be able to stop it even if we tried. But we may may
be able to bias this development in a more positive direction.
22
23
http://singinst.org/
singularity-institutes-scary-idea-and.html
107
9
Approaches to AGI
This chapter, and the following one on chatbots, are
updated/improved versions of material that earlier appeared in
the would-be pop-sci book The Path to Posthumanity, that I
wrote with my good friend Stephan Vladimir Bugaj. (That book
was my first foray into nontechnical futurist writing, and it didnt
really take off; but there was some good stuff in there.) . This
is a topic on which I have a huge amount to say, and have
written a lot before and will surely write again. The present
chapter presents more of a broad-scope historical focus on the
field than my other nontechnical writing has tended to do.
As I noted in the intro above, one thing this book DOESNT
contain is a detailed overview of my own work on AGI, which Ive
described in a variety of technical works before, and which is
currently centered on the OpenCog open source AGI platform.
Im currently (mid-2014) cooking a non-technical book called
Faster Than You Think which will cover this ground.
This
chapter and the next give some additional background on the
AGI field, and give some perspective on my own AGI approach,
but dont dig into the details.
Creating an AGI is not a narrowly defined problem like, say,
building a airplane or designing a book. Its a lot more openended, more like building a flying machine or building a device
for transmitting linguistic information or building a device for
generating electricity.
We have balloons, blimps, copters,
planes, pedal-gliders, rockets, Space Shuttles, catapults We
have books, ebooks, books on tape, videos, etc. And we have
solar power satellites, windmills, waterwheels, nuclear fission
and fusion, garbage-to-gas, fossil fuels, fuel cells. Each type
of flying machine or electrical power generator has its own
particularities its own science and its own engineering
practicalities and the same is true with the various approaches
108
Approaches to AGI
to AGI. And just as with these other areas of endeavor, even

though many different approaches may be valid, certain
approaches are going to be more productive at particular points
in history, based on the other technologies available and the
state of scientific knowledge.
In this chapter Ill give you a birds-eye view of some of the
leading approaches to AGI that have been proposed by various
scientists both in the past and right now. I wont give equal time
here to all the different AGI approaches that have been
suggested or pursued, but will rush lightly over some rather large
and interesting areas, in order to focus on those I consider most
promising for the near term. Of course I cant know for sure my
judgment is correct on this, but to get anything done in reality
one has to make some concrete choices and there are plenty
of other books around, explicating the approaches I neglect.
Starting at the very high level, there are at least four broad
approaches to creating advanced AGI: leveraging & networking
narrow AI; emulating the brain; evolving AGI via artificial life; or
the direct approach of designing and engineering AGI now based
on the various sorts of knowledge available. Each of these
approaches then yields a host of sub-approaches, often with
dramatic differences. The direct approach is the one Ill focus on
most in this book, though with occasional nods to the brain
emulation approach.
But first, Ill run through the other
approaches, highlighting their strengths and weaknesses, and
explaining in each case why Ive decided its probably not the
way to go if your goal is to create a beneficial AGI as rapidly as
possible.
Its worth remembering that, in spite of all the diversity of
opinions about the details, there is a broad agreement among
AGI researchers about some basic philosophical principles. For
instance, theres basic agreement in the AGI field that mind is
not intrinsically tied to any particular set of physical processes or
structures. Rather, mind is shorthand for a certain set of
patterns of organization and dynamics in systems, that are
109
correlated with generally intelligent behavior. These patterns of

organization and dynamics obviously can emerge from a human
brain, but they could also be made emerge from a computer
system (or a whale brain, or a space-alien brain). Unless its
created via detailed brain-simulation, a digital mind will never be
exactly like a human mindbut it may still manifest many of the
same higher-level structures and dynamics. The disagreements
in the field regard how to figure out just what the critical patterns
of organization and dynamics are, for generating generally
intelligent behavior.
AGI Via Narrow AI?

As Ive emphasized already, most of the AI programs around
today are narrow AI programs they carry out one particular
kind of task intelligently. You could try to make an AGI by
combining a bunch of juiced-up narrow AI programs inside some
kind of overall framework. Many folks find this kind of approach
appealing, for the obvious reasons that some pretty good narrow
AI programs already exist, so if we could leverage all the work
thats gone into them to build AGI, that would be a wonderful
effort-saver. Plus there are strong economic incentives for the
creation of narrow AI, due to the latters ability to deliver
commercial value relatively rapidly with fairly high confidence
so if all this short-term-oriented work could be leveraged to yield
AGI with long-term value, then society could get long-term value
without having to defer gratifications and place long-term
progress above short-term gain. Wouldn't that be nice!
Unfortunately, Im rather skeptical of this approach as a road to
powerful AGI basically because none of the narrow AI
programs have the ability to generalize across domains, and it is
far from obvious how combining them or extending them is going
to cause this to magically emerge. Overall, I think this is an
approach that seems appealing from a birds-eye view, then
seems to make less and less sense as you dig more and more
deeply into it. When you look at the narrow AI components
110
Approaches to AGI
available today say, Google, Deep Blue, credit card fraud

detection systems, bio data analysis tools, theorem proving
assistants, and so forth the notion that you could combine
them together to form a human-level general intelligence starts
to seem almost ludicrous. Where would the common sense
knowledge, and the ability to gain and improve it, come from?
The disparate AI approaches embodied in various existing
narrow AI systems do not typically involve compatible notions
and representations of what knowledge is, and how to reason
about it. And even if one were to modify the underlying
algorithms to be compatible, just chaining together a bunch of
unrelated algorithms into a processing pipeline or
interconnection network does not seem likely to produce general
intelligence. It seems to me that, to combine narrow AI programs
together in intelligent way would require some kind of highly
capable integrator component, to figure out how to combine the
different narrow AI systems in a context-appropriate way, and fill
in the gap when none of them fits the bill. But then this integrator
component would essentially need to be a general intelligence
unto itself so really one would have an AGI with various
narrow-AI systems as modules to reference as needed, rather
than an AGI built of narrow AI systems. The question then
becomes how useful a crutch would the narrow AI systems be
for the central AGI component.
None of these arguments, however, are meant to imply that
hooking together narrow AI algorithms into a common adaptive
framework is a bad thing to work on. Im sure that composite
narrow AI systems of this sort can be built to great effect, for
various applications. For example, one can imagine an amusing
and useful chatbot, along the lines of the Siri virtual assistant
chatbot acquired by Apple in 2010, combining simplistic not-sointelligent English conversation with narrow AI algorithms for
responding to particular commands: narrow AI for travel
reservations, narrow AI for calendar management, narrow AI for
answering factual questions, etc. However, such a virtual digital
assistant would be quite different from a generally intelligent AI
111
capable of carrying out an English conversation with human-like

understanding.
AGI Via Brain Emulation?

More and more very smart, serious neuroscientists are starting
to talk about the possibility of creating human-level AGI via brain
emulation. I think this is a fascinating, important approach that is
bound to work eventually. Ive spent a bit of my own time on
related pursuits for instance, I edited the first journal issue
devoted to mind uploading (taking specific human brains and
replicating them digitally), and Ive done some brain simulation
work for government customers, as part of my AI consulting
business. However, my current feeling is that brain emulation
wont be the fastest or best approach to creating human-level
AGI.
One minor problem with this approach is that we dont really
understand how the brain works yet, because our tools for
measuring the brain are still pretty crude. Even our theoretical
models of what we should be measuring in the first place are still
hotly debated. Computer scientists like modeling brains as
formal neural networks in which neurons, the main brain cells,
are represented as objects similar to electrical switches.
Computational neuroscientists currently like spiking neural net
models, that go one level deeper, and look at the real-time
electrical signals coming out of neurons. But then a few
neuroscientists emphasize the importance of phenomena not
encompassed in these spiking models (e.g. dendro-dendritic
synapses, spike directivity, extracellular charge diffusion). These
additional phenomena (if indeed theyre important, which is
currently unclear) could be incorporated in more detailed
computational models, but we now lack the data to construct
these accurately. A few renegades believe wed need to go
further down and model the brain at the molecular or atomic
level, to model the dynamics giving rise to thought with sufficient
accuracy. The point is, we understand so little about how the
112
Approaches to AGI
brain works, that we dont even know what kind of computational

model would be needed to make an AGI that emulates the brain.
Neuroscience knowledge is increasing rapidly, and its an
exciting field to be involved in, even peripherally as I have been.
fMRI and other brain imaging techniques have taught us a lot
about which kinds of mental activity take place in which brain
regions; and our models of low-level, localized neural activity get
more and more accurate. However, we still have very little idea
how knowledge is represented in the brain, except a few simple
cases of sensory or motor knowledge; and the dynamics of
complex thoughts remains almost entirely a mystery. Eventually
these puzzles will be solved, but it will involve breakthroughs in
brain imaging technology, providing massively greater
spatiotemporal accuracy across the whole brain than anything
we know how to build now. These advances will happen but its
a mistake to say we can just make AGI by copying the brain as
if this somehow reduces the problem to something simple or
straightforward. Radical breakthroughs in brain imaging will be
needed to accomplish this. Some may feel these breakthroughs
are somehow more straightforward than the work needed to
engineer AGI systems, but I dont see it that way. Personally, I
feel I have a reasonable understanding of how to make a
human-level AGI, and I dont know how to make radically more
effective brain imaging technology! Of course, there is surely
another researcher somewhere who feels he has a good idea of
how to make radically better brain imaging technology, but has
no idea how to engineer an AGI, and I freely admit that
predicting the timing of radical advances is a difficult art. What I
protest against is the assumption, which some futurist pundits
seem to make, that a big leap in brain imaging is somehow
intrinsically more straightforward or predictable than a big leap in
AGI.
It also seems likely that an AGI modeled closely on the human
brain would require drastically more computational resources
than a comparably generally intelligent AGI created with a
cognitive architecture more suited to the computers available
113
today. Neural wetware is efficient at doing certain kinds of lowlevel operations, and contemporary computer hardware is
efficient at doing other kinds of low-level operations. Achieving a
given cognitive function via emulating neurons on a
contemporary computer, is almost certainly never going to be the
most efficient way to get that function done on that same
computer the neural approach is almost surely going to take far
more memory and processing power than needed.
This inefficiency issue certainly doesnt make AGI via brain
emulation infeasible. Computing power gets cheaper all the
time, at an impressive rate; and if we knew how to make an
advanced AGI using neural modeling, wed have a lot of
motivation to figure out clever software strategies for making its
implementation more efficient. And there is also the possibility of
creating custom hardware for brain emulation, varying on current
hardware designs in a way specifically tailored to make brain
emulation more efficient. This may well be feasible, but its hard
to say much about it in detail till were clearer on how low-level a
brain simulation needs to be to give rise to cognitive phenomena
appropriately. Several research groups are already making
brain chips, including Dharmendra Modhas group at IBM and
Kwabena Boahens group at Stanford, but its currently quite
unclear whether these brain chips are going to be adequate for
brain emulation. They lack mechanisms supporting efficient
simulation of many neural phenomena of possible importance for
cognition (directed spiking, dendro-dendritic synapses,
extracellular charge diffusion and more) but it may be that
these phenomena arent so critical for AGI after all, we just dont
know at this point.
A different sort of issue regarding brain emulation as an
approach to AGI is the fact that once you succeed, what you get
is something with a very humanlike mind and its not clear
whether this is a bug or a feature, in the scope of things. Digital
humans would be fantastically interesting for science, but would
suffer many limitations compared to digital minds created with a
more flexible cognitive architecture. Creating non-human AGIs
114
Approaches to AGI
customized to serve humans is one thing, ethically and

aesthetically speaking: creating digital humans and tweaking
them to want to serve biological humans is another thing, and
could backfire in various ways tied to the human motivational
and emotional structure. Also, the maxim power corrupts and
absolute power corrupts absolutely is an observation about
human psychology rather than a universal maxim applicable to
all general intelligences. Once digital human minds become
possible, its fairly likely someone will try to endow one with great
pragmatic power a scenario thats been explored repeatedly in
dystopic science fiction.
On the other hand, futurists Anna Salamon and Carl Shulman
have argued that human brain emulation is the safest route to
creating advanced AGI, because human cognitive architecture is
something we understand at least vaguely (and will likely
understand better once we get to the stage of being able to
construct digital humans). In their view, creating an AGI via
means other than brain emulation is sort of like sampling a mindarchitecture randomly from the space of mind-architectures
(albeit with a certain complex set of statistical biases), and the
odds of picking one that will be nice to humans is probably pretty
small. I dont share this perspective, because I believe one can
architect AGI systems with more rational, predictable and goalfocused cognitive architectures than humans have. Of course we
can't have absolute confidence what such an AGI system will do,
but theres also no absolute confidence about the behavior of
hypothetical digital humans. Obviously, the level of uncertainty
about such matters is sufficient that rational, deep-thinking
people can be expected to have some level of disagreement.
One thing that seems very likely is: If we did create a digital
human before creating some other kind of AGI architecture, then
we could learn an awful lot about brain, mind and AI by
experimenting with that digital human. Before too long, via
experimenting with various changes to digital human minds and
seeing their effects, we would gain a vastly improved scientific
understanding of how human-like minds work, and also to some
115
extent about minds in general. So creating a digital human

would likely lead, before too long, to the creation of not-sohuman-like AGI minds. And conversely, a sufficiently intelligent
engineered AGI system, if well disposed toward helping humans
with their endeavors, would probably be very helpful to biologists
in their quest to scan brains and create digital humans.
Finally, its important to distinguish brain emulation as an
approach to AGI, from neural networks as an approach to AI or
AGI. What are called neural networks in computer science and
mathematics are formal constructions, loosely inspired by the
network of neurons in the brain, but by no means constituting
realistic neural models. Not many computer scientists or
mathematicians get confused about this, but journalists often do.
When I was a math professor, I once had a colleague, freshly
immigrated to the US from mainland China, who had done his
mathematics PhD thesis on a certain kind of neural networks. He
was amazed when I pointed out to him that the word neural
referred to certain cells in the brain. His thesis had been a good
one, proving novel and powerful results about certain aspects of
certain mathematical neural nets but actually, at no point in his
research would it have been useful to him to know the biological
analogy to the mathematical objects he was working with. And
these particular neural nets that he studied were abstracted far
enough from the original biological basis of neural nets, that its
arguable they didnt even deserve the name anymore though
they were certainly broadly neural-network-like systems.
The popular confusion about the status of formal neural nets as
brain models rose to the fore in 2009 when Dharmendra
Modhas team from IBM made headlines by reporting a
computer simulation of a cat brain. Actually, what they did was
to run a formal neural net model with roughly the same number
of neurons as a cat brain. It wasnt a serious computational
neuroscience simulation, it was a computer science style formal
neural net. And furthermore, the connections between the formal
neurons were decided randomly based on some statistical
116
Approaches to AGI
formulae there was no effort to model the connection structure

of a cat brain, which is what, in the hypothesis that the cat brain
is well modeled as a neural network, would encode the cats
mind. So: yes, Modhas team did create a formal neural network
that in some sense was the size of a cat brain. But this neural
network was not the kind that models the brain in any serious
way.
The same confusion popped up again in 2011 when Modhas
team announced the creation of a prototype brain chip. Similar
chips had been prototyped before, e.g. by Kwabena Boahens
team at Stanford, but this naturally attracted more attention
because it was IBM doing it. But the brain chip, inasmuch as
can be limned from the news reports, constitutes a hardware
implementation of a relatively simple formal neural net model.
This may be very interesting and useful, but its not nearly as
brain-ish as most people reading the press releases are probably
led to believe.
On the other hand, in 2008 when Gerald Edelman and Eugene
Izhikevich performed a large scale computer simulation of large
portions of the human brain, using much more biologically
realistic models of neurons and making more of an effort to
emulate realistic connection statistics and brain architecture, the
media hardly noticed. Their simulation demonstrated large-scale
activity qualitatively resembling that in the human brain. It
certainly wasnt an AGI system, but it was outstanding
computational neuroscience. But IBM wasnt involved, and
Edelman and Izhikevich were more conservative in phrasing
their accomplishments publicly.
AGI Via Artificial Evolution

Human and animal intelligence were produced via evolution so
why not try to evolve AGIs? Evolutionary algorithms software
processes emulating the logic of evolution by natural selection
have already proved effective at finding solutions to various
117
simpler problems. And artificial life software programs have

showed the artificial evolution of simple digital organisms in
simple simulated environments such as artificial video-gamelike bugs, or simulated DNA that emerges from a simulated
prebiotic soup and reproduces itself, etc. Why not take this one
step further and just run some kind of computer-simulated
ecosystem, and let generally intelligent simulated organisms
emerge therein, selected via evolution to achieve their goals in
the simulated environment?
There are a couple of reasons why this probably isnt the best
approach to AGI right now though its certainly an interesting
research direction. First of all, we dont understand evolution
terribly well yet. Secondly, the amount of computing power
required to accurately simulate a rich, diverse ecosystem from
which intelligence could plausibly emerge, would clearly be
vastly greater than the amount required to accurately simulate a
single intelligent system.
The natural counterargument to these complaints would be that
maybe we dont need to emulate biological evolution exactly, but
just to implement some broadly similar evolutionary process in a
simulation And maybe we can tailor this simulation to foster
the evolution of general intelligence, so that the computational
requirements wont be so severe as in a simulated environment
not thus tailored. This counterargument is reasonable enough on
the face of it, but my own experience playing with artificial life
type simulations and following the literature in the field -- has
made me very aware of the difficulties faced in trying to design a
really effective simulated evolutionary process. So far, every
artificial life simulation ever created hits a fairly low ceiling of
emergent complexity none of them, so far, has led to a series
of surprising emergent phenomena (new digital organisms, new
behavioral phenomena) going significantly beyond what the
programmers expected when they created the simulation.
Little simulated bugs, put in an artificial environment where they
need to compete and cooperate for food, will run around and
118
Approaches to AGI
compete and cooperate for food and demonstrate funky behavior

patterns, and its all a lot of fun but they dont evolve dramatic
new structures or behaviors or get dramatically smarter and
smarter or give rise to new, unforeseen kinds of organisms. Alife
pioneer Tom Ray got DNA-like replication behavior to emerge
from an impressively simple Alife substrate, but his followup
project, aimed at demonstrating a similar emergence of
multicellular digital life from single-celled digital life, failed to yield
the hoped-for results.
My own impression, from working on Alife for a while in the
1990s, was that making really rich and fertile Alife would require
a deeper artificial infrastructure than is typically used in the field.
In the typical Alife application, one has artificial organisms
encoded by artificial genomes, which then are adapted via
evolution based on fitness in the artificial world. But the mapping
from genome to organism from genotype to phenotype, in the
biology lingo is extremely simplified. In real organisms this
mapping involves all sorts of complex mysterious processes like
protein folding (which we still dont understand well),
spatiotemporal pattern formation, nonlinear-dynamical selforganization and all manner of signaling. Which aspects of the
genotype-phenotype mapping are important for the creative
richness of the evolutionary process as its happened on Earth?
The biological genotype-phenotype mapping involves a heck of a
lot of chemistry, and this seems to be important in terms of
evolutions ability to come up with wacky new things that were
unpredictable based on the organisms that came before. All
these wacky new things have to operate consistently with the
laws of chemistry and physics, and are guided by these laws in
their particular manifestations on the surface of the Earth. This
led some Alife researchers to work on artificial chemistry, and
try to ground their artificial biology in artificial chemistry. But none
of the artificial chemistries seemed to demonstrate the creative
flexible generativity of real chemistry. Which aspects of realworld chemistry are important for the emergence of biology thats
119
capable of leading to the evolution of general intelligence in

environments resembling the one on the surface of the Earth?
And chemistry, of course, gets its power and particularities from
physics So the question may even become, what aspects of
physics are necessary to give rise to a chemistry capable of
giving rise to a biology capable of fostering the evolution of
general intelligence given a reasonable amount of time and
space and materials? I very much doubt one needs to simulate
physics down to the subatomic level in order to make an artificial
life simulation giving rise to the same flexibility and generativity
as real biology. My intuition is that some sort of fairly abstracted,
well-designed artificial chemistry should do the trick. However,
my main point here is that some fairly deep open research
problems are involved here.
As in the case of brain emulation, my point isnt to argue against
Alife as an area of research I think its a fascinating field and
that people should be working on it! I wish all these areas of
research were funded a lot more heavily, as opposed to many of
the fairly pointless things our society spends its time and energy
on. Im quite confident AGI via brain emulation will work, though
my best guess is that it will arise after AGI already exists via
other means, because radical advances in brain imaging may
not come that fast.
I suspect that AGI via Alife will work
eventually, though probably well after both engineering and brain
emulation have succeeded, since the amount of computing
power required to create a good-sized artificial ecosystem will be
immense. I doubt that AGI via networked narrow AI will ever
succeed fully, but I think the attempt might lead to some rather
interesting software and hardware applications.
Good Old-Fashioned AI
What Im calling symbolic AI is in its simplest, classic form, at
least sometimes referred to as Good Old Fashioned AI
(GOFAI). GOFAI systems are based on rules that are much like
120
Approaches to AGI
mathematical axioms or grammar rules. But instead of combining

algebraic variables or parts of speech, these rules combine
concepts. Such a system tries to solve the problems its
confronted with by figuring out how to combine its rules to yield
appropriate conclusions (conclusions representing the answer to
a question it was asked, or representing the series of actions that
must be taken to achieve the goal its been asked to achieve,
etc.).
One formerly popular kind of symbolic AI system, an expert
system, used rules that were explicitly encoded by human
experts to represent their specialized knowledge about some
subject. To give a concrete flavor of this kind of AI, here is a
simple English version of a typical rule used by the classic
medical expert system MYCIN:
IF
1)
2)
3)
the stain of the organism is gramneg and

the morphology of the organism is rod and
the aerobicity of the organism is aerobic
THEN
there is strongly suggestive evidence (0.8) that the class of the
organism is enterobacteriaceae
The rule isnt actually expressed in English inside the expert

system, its expressed in a mathematical language; but it can
print out the English version for human consumption.
And the following shows an example interaction between a
human and a medical expert system that uses rules like this
internally:
1.
Patient Information
1) Patients name
PT538
2) Age
34 years
3) Sex
Male
121
Diagnosis
6)
Please enter results of pending cultures in table:

SITE CULTURE# DATE EXAMINED
CSF 1234 30.1 YES
10) Do you suspect PT538 may have an infection at a site from

which you have not obtained specimens?
No
15) Has PT538 recently had objective evidence of abnormal
neurologic signs (e.g. seizures, coma) documented by
physician?
Yes
The CSF culture will be considered to be associated with
meningitis.
Cause
17) Does PT538 have an abnormal chest x-ray?
No
18) Has PT538 been exposed to any contagious diseases
recently?
No
27) Please enter CSF findings in table:
WBC %PMN Protein CSFglucose Periph-glucose
100 99 280 5 UNKNOWN
35) Is meningitis a hospital-acquired infection?
Yes
39) Is PT538 a burns patient?
Yes
I have found evidence (RULE545 RULE557)that treatment should
cover for e-coli causing meningitis
I have found evidence (RULE545 RULE557 RULE578)that
treatment should cover for pseudomaonas aeruginosa causing
meningitis
I have found evidence (RULE545 RULE557 RULE578)that
treatment should cover for Klebsiella pneumoniae causing
meningitis
122
Approaches to AGI
I have found evidence (RULE507)that treatment should cover for

Diplococcus-pneumoniae causing meningitis
Treatment
It is important to cover for the following infection(s) and associated
organism(s):
Meningitis
<item 1>e-coli
<item 2>pseudomaonas aeruginosa
<item 3>Klebsiella pneumoniae
<item 4>Diplococcus-pneumoniae
41) Does PT538 have a clinically significant allergic reaction to any
antimicrobial agent?
No
43) PT538s weight in kilograms:
70
Recommendation 1
My preferred therapy is as follows:
In order to cover for items <1 2 3 4>
give the following in combination:
1)
AMPICILLIN
Dose 3.5g q4h IV
2)
GENTAMYCIN
Dose 119 mg q8h IV
Since high concentration of penicilllins can inactivate
aminoglycosides do not mix these two antibiotics in the same IV
bottle.
All this is quite impressive, and quite useful. What makes this
sort of approach worrisome from an AGI perspective is that the
expert system didnt really have any understanding of what it
meant by terms like dose or ampicillin. It knew them only as
symbolic tokens, taking part in rules encoded by humans.
Expert systems can actually do very well in some areas
medical expert systems like MYCIN being one example. They
have been shown to diagnose diseases better than most human
123
physicians. Of course, such an expert system wont notice if the

patient is lying or incorrect about their symptoms, but truth be
told, most doctors wont either. Once the symptoms are known,
the expert system can apply its rules to figure out what the cause
of the symptoms is likely to be. Of course, is that the system did
not learn any of the rules it uses they were encoded by a
human. But then, most doctors didnt learn the rules they use to
perform diagnosis, at least not in the sense of inducing them
from their experience they just memorized them from a
textbook or a teacher.
When MYCIN was created at Stanford in the early 1970s, getting
diseases diagnosed by an AI program was a pretty radical idea.
Nowadays, there are plenty of websites where you can click
through multiple pages, choosing your symptoms from a list of
options on each one, and at the end get some information on
whats probably wrong with you. So the idea of getting diseases
diagnosed or medicines chosen by a computer program probably
wouldnt shock anyone now. In fact, we dont commonly think of
such websites as AI, even though (in some cases) theyre doing
basically the same thing as MYCIN we just consider them
useful computer software.
An expert system is a paradigmatic narrow AI system it carries
out a very specific purpose, and in some cases it does so quite
well. However, it doesnt understand the context in which it
operates, and so it has severe limitations in terms of general
intelligence. If a new disease is discovered, an expert system for
medical diagnosis will never adapt to it, unless some human
being updates its rule base.
GOFAI Grows Up: SOAR

Modern symbolic AI systems go far beyond early expert systems
perhaps the most impressive such system is SOAR, created in
1983 and developed continuously ever since, which uses handcoded expert rules at various levels of abstraction, but wraps
124
Approaches to AGI
them in a fairly sophisticated cognitive architecture modeled on

human psychology. SOAR has been used to model human
psychological behavior in various laboratory experiments, and
also to perform various practical tasks. Perhaps the most
impressive SOAR application was TacAir, which simulated the
behavior of human fighter pilots, using hand-coded rules
describing their knowledge and behavior in different situations.
TacAir-Soar was a flight simulator capable of flying any routine
pattern that the US Navy, Air Force, and Marines had for any
fixed-wing aircraft. It could prioritize objectives, make decisions,
and follow orders on its own. For example, suppose the
simulated airplane takes off, with a current mission objective of
reconnaissance. Suppose that partway through the mission, an
enemy fighter launches a missile at it. Rather than continuing on
its path, it will internally reprioritize, shifting the task of avoiding
the missile above the task of continuing to perform
reconnaissance. Once the missile has been avoided, it will then
check to find out where the enemy fighter is, and decide what to
do: ignore it because its out of range, or engage the enemy
fighter and shoot it down. Furthermore, it could receive
instructions in simple English language sentences. It didnt learn
to any significant degree what happened to it on one mission
didnt help it do better on future similar missions. But it was
impressively functional, carrying out complex dynamic behaviors
that would have been a lot more difficult to achieve in a standard
software implementation, without sophisticated software
architecture like SOAR.
Overall diagram of the SOAR cognitive architecture.
Productions are technical lingo for formal rules similar to expert
rules. Chunking is combination of existing rules to form new
ones.
I know the two of the three originators of SOAR, John Laird and
Paul Rosenbloom, moderately well, and they are both serious
AGI thinkers. Neither of them considers themselves terribly
close to creating human-level AGI, but (to take the risk of
125
summarizing other peoples views in a crude and concise way)

they feel theyre on research paths that have reasonable odds of
eventually leading to human-level AGI and that even if they
dont get there during their research careers, maybe their
students students students students will. Laird is still actively
developing SOAR, whereas Rosenbloom is developing a new
approach to AI based on mathematical structures called factor
graphs, which he has embodied in an alternate cognitive
architecture called Sigma. Sigma has a lot in common with
SOAR, but also has some profound differences, particularly its
embracing of probabilistic knowledge representation as a central
aspect.
Laird, like everyone else in the SOAR community, is well aware
of SOARs limitations it doesnt do learning or manage
uncertainty in a very sophisticated way, it cant deal with largescale sensorimotor data, etc. However, the basic perspective
underlying SOAR is that, if you get the core cognitive
architecture right, then these other aspects can be inserted into it
without major changes. Rosenbloom has deviated from this
perspective a bit, focusing on developing factor graphs as a
single mechanism for handling every sort of knowledge. To put it
crudely, he figures SOAR has already largely solved the
cognitive architecture problem, but to make a powerful AGI
system, it will be necessary to have a stronger underlying
knowledge representation framework, which can handle large
amounts of uncertain knowledge in a scalable way and he
thinks factor graphs can fit the bill.
Anyway, much as I respect Laird and Rosenbloom both, I think
their approaches to AGI (and especially SOAR) still have far too
much GOFAI in them. In my view, beginning with hand-coded
rules starts you off in the wrong direction because knowledge
learned from experience has a dramatically different character
from hand-coded knowledge. Experientially learned knowledge
tends to be far less crisp and well-organized than human-coded
rules, it consists of complex networks of interrelated, weakly-held
knowledge; and its appropriately handled by cognitive processes
126
Approaches to AGI
very different than the ones that work on small sets of crisply
defined rules. Humans can indeed handle crisp, formalistic rules
like those existing in expert systems or SOAR but they do so
using methods that are evolved to be guided by other,
unconscious cognitive processes acting on huge networks of
highly uncertain knowledge at varying levels of abstraction.
Millions and Millions of Rules: Cyc

The ultimate monster GOFAI systems is actually one of the more
recently conceived ones: Cyc a project ongoing since 1984,
founded by Doug Lenat. In good old GOFAI fashion, Cyc is
focused on trying to build a program with common sense, via
programming it explicitly as a massively interconnected rule set.
The main thing differentiating Cyc from prior GOFAI efforts is the
sheer bulk of its rule-set. The Cyc team is mainly focused on
encoding millions of items of data, so that the program can know
everything an eight-year-old kid knows.
Cyc was originally short for encyclopedia, but they found that
the knowledge they needed was quite different from that found in
encyclopedias. It turns out encyclopedic knowledge is too
abstracted and requires too much grounding to serve as
foundational for intelligence development, something which
should have been more obvious form the start. Rather, they are
now focused on everyday knowledge you could get by asking a
small child: a combination of dictionary-like simple word
definitions and contextually-embedded situational knowledge.
Each common-sense concept in Cyc gets an English-language
definition as well as a mathematical definition, which tries to
paraphrase the English definition. For example, the Cyc-English
definition of skin goes like this:
A (piece of) skin serves as outer protective and tactile
sensory covering for (part of) an animals body. This is the
collection of all pieces of skin. Some examples include The
Golden Fleece (representing an entire skin of an animal),
127
and Yul Brenners scalp (representing a small portion of his

skin).
The Cyc-English definition of happiness is:
The enjoyment of pleasurable satisfaction that goes with
well-being, security, effective accomplishments or satisfied
wishes. As with all Feeling Attribute Types, this is a
collectionthe set of all possible amounts of happiness one
can feel. One instance of Happiness is extremely happy;
another is just a little bit happy.
Cyc is based on getting humans to bootstrap the conceptual
groundings normally gained through experiential learning by
encoding their own experiences as a foundation for intelligence
in other words, to tell computers what symbols (words in these
cases) mean.
The Cyc project is interesting, but I think its fundamentally
flawed dont believe that the logical definitions in Cyc have all
that much overlap with the kind of information contained in the
mind of an eight-year-old child. We humans arent even explicitly
aware of much of the information we use to make sense of the
world: not in the least because it has evolved through dynamic
processes in interaction with our environment in complex ways,
rather than being mere memorizations of formally presented
definitions and rules.
A humans notion of happiness or skin is much bigger, more
disorderly and messier than these definitions. Attempts to make
our human definitions more rigorous and formal, and more
compact, always lead to incomplete notions of things. These
kinds of general abstract definitions may be inferred in the
human mind from a whole lot of smaller-scale, practical patterns
involving skin and happiness, but theyre not the be-all and endall. In dealing with most practical situations involving skin and
happiness, we dont refer to this kind of abstraction at all, but we
use the more specialized patterns that the general conclusions
were derived fromeither individually or in combination.
128
Approaches to AGI
Additionally, our mental flexibility allows us to spontaneously

derive new patterns from new observations, combined with old
knowledge.
This is a fundamental problem with any system that cannot learn
fundamentals from experience there has been an
interpretation, formalization, and compacting step by the
bootstrappers in an attempt to make the task tractable which
necessarily leads to information loss. Furthermore, it is biased by
the originating individuals notions of the concepts, which are not
100% the same from person to person. A learning system, like a
child, can take in the varying and sometimes opposing ideas
about a concept, synthesize them, and make its own
conclusions. A system that cannot do that is at the mercy of its
initial conditions.
Basically, Cyc tries to divorce information from learning.
However, this cant really be done, at least not anywhere near as
thoroughly as the Cyc folks would like. In practical terms, a mind
can only make intelligent use of information that it has figured out
for itselfor else is of roughly the same form as the information
it figures for itself. Furthermore, if it has not learned through
experience, it has no experience doing so, and thus will be
unable to properly adapt over time to a changing environment
(and as any English speaker should understand, even the
linguistic landscape is an ever-changing one).
It must have had a structure for autonomously-inferred
knowledge in the first place; otherwise there is no mechanism for
grounding knowledge in any realm of self-action. If the
information read into an AI system from a database is too
different in structure from the information the AI system has
learned on its own, then its reasoning processes will have a hard
time integrating the two. Integration may be possible, but it will
be extremely time-consuming, and would most likely proceed in
a similar manner to how we as humans learn information thats
just dumped before us.
129
Any AI system will need to learn some of its knowledge base for
itselfno Cyc-like system can contain all possible knowledge an
AI system will ever need to know for interaction with any
environment, as its totally impossible to predetermine every
eventuality it may encounter during its existence. Any knowledge
base that isnt structured in a way that naturally matches with the
structure of learned knowledge will be effectively useless, but
how can the people building a database like Cyc know what
knowledge representations will match with learned knowledge, if
they arent building a learning system? For a learning system a
system like Cyc is, at best, a kind of basic encyclopedia that a
sufficiently advanced learning system could learn to read but
which is not foundational to the system becoming intelligent in
the first place.
Despite more than twenty years of development, Cyc never
succeeded in emulating an eight-year-old child. Nor has anyone
yet found much use for a CD-ROM full of formal, logical
definitions of common-sense information. The company Cycorp
is doing OK, supported largely by government research grants.
Cycorp has been quite careful not to mention anything about
artificial general intelligence or any long-term scientific mission
underlying the firms work. Instead, they characterize their goal
as producing a useful database intended for embedding within
various specialized software products. This is a worthwhile
mission to be sure, but very different from the grand goal of AGI.
In fairness to Doug Lenat, I have to admit that hes a deep
thinker and his computational-psychology perspective does have
some depth to it far more than is represented in the public face
of Cyc. He has a reasonably solid theory of general heuristics
problem-solving rules that are abstract enough to apply to any
context. His pre-Cyc programs, AM and EURISKO, applied his
general heuristics theory to mathematics and science
respectively. Both of these programs were moderately
successful, exemplars in their field, but far from true general
intelligence. Their design lacks a holistic view of the mind. In the
big picture of AGI, getting the minds heuristic problem-solving
130
Approaches to AGI
rules correct, alone, means almost nothing. Problem-solving

rules gain their psychological meaning from their interaction with
other parts of the mind, and their grounding in experience of and
interaction with some environment. If the other parts arent even
there, the problem-solving is bound to be sterile.
EURISKO won a naval fleet design contest two years in a row
(until the rules were changed to prohibit computer programs from
entering), and it also received a patent for designing a threedimensional semiconductor junction. Yet, when considered
carefully, even EURISKOs triumphs appear simplistic and
mechanical.
Consider
EURISKOs
most
impressive
achievement, the 3-D semiconductor junction. The novelty here
is that the two logic functions, Not both A and B and A or B,
are both performed by the same junctionthe same device. One
could build a 3-D computer by appropriately arranging a bunch of
these junctions in a cube.
How did EURISKO make this invention? The crucial step was to
apply the following general-purpose heuristic: When you have a
structure which depends on two different things, X and Y, try
making X and Y the same thing. The discovery, albeit an
interesting one, came directly out of the heuristic. This is a far cry
from the systematic intuition of a talented human inventor, which
synthesizes dozens of different heuristics in a complex, situationappropriate way. EURISKO, in essence, solved the
semiconductor design problem in the same way that Deep Blue
plays chess by recursively applying a given set of rigid rules
until a solution that was optimal under the specified fitness
criteria popped out. It could search a space of solutions under
the given rule conditions and test them against a formal
optimization criteria, but it had neither intuition about nor
understanding of what it was doing.
By way of contrast, think about the Croatian inventor Nikola
Teslaprobably the greatest inventor in electrical engineering
historywho developed a collection of highly idiosyncratic
thought processes for analyzing electricity (Citadel, 1998). These
131
led him to a steady stream of brilliant inventions, from alternating

current to radio to robotic control, but not one of his inventions
can be traced to a single rule or heuristic. Each stemmed
from far more subtle intuitive processes, such as the
visualization of magnetic field lines, and the physical metaphor of
electricity as a fluid. Each of Teslas notions involved the
simultaneous conception of many interdependent components.
Problem-solving in the general sense is not merely getting lucky
in applying a fixed set of heuristics to a particular problem; but
creatively devising not only new heuristics, but whole new
categories of heuristics (we would call these new ways of
looking at the problem); and even new problems and categories
of problems into which we can decompose a particularly vexing
large problem. If Tesla could only work with the existing
engineering heuristics of his day, rather than creating new ones,
hed never have created so many amazing inventions.
EURISKO may have good general-purpose heuristics, but what it
lacks is the ability to create its own specific-context heuristics
based on everyday life experience. This is precisely because it
has no everyday life experience: no experience of human life
and no autonomously-discovered, body-centered digital life
either. It has no experience with fluids, so it will never decide that
electricity is like a fluid. It has never played with blocks or
repaired a bicycle or prepared an elaborate meal, nor has it
experienced anything analogous in its digital realm. Thus, it has
no experience with building complex structures out of multiple
interlocking parts, and it will never understand what this involves.
EURISKO pushes the envelope of rule-based AI: it is just about
as flexible as a rule-based program can ever get, but it is not
flexible enough. In order to get programs capable of contextdependent learning, it seems necessary to write programs which
self-organizeif not exactly as the brain does, then at least as
drastically as the brain does. Hand-encoded knowledge can
potentially be useful in this process of self-organization, but only
if its encoded in a way that matches up naturally with learned
132
Approaches to AGI
knowledge, so that the appropriate synergies can emerge. This

means that one cant think about the mind by thinking only about
rules; one also has to think about learning and experience, even
if whats doing is creating rules to be fed into an AGI system.
The Modern, Direct Approach to AGI

The approach I advocate, for the goal of creating beneficial AGI
as rapidly as possible, is to simply design and engineer a
thinking machine. Dont wait for better brain imaging technology
or a good theory of the particulars of the human brain; dont wait
to understand the mysteries of evolutionary biochemistry or
accumulate the compute power needed to simulate an
ecosystem; dont mess around with networking narrow AI
software that was never designed to handle common sense or
broad-minded generalization; and for sure dont work on narrow
AI with the illusion that its someday going to miraculously start
constituting progress toward AGI, which the last 50 years of AI
have taught us is a quite different problem. Just move directly
toward the goal of advanced AGI, with general intelligence at the
human level and then beyond. Take our best understanding of
how the mind works, and embody it in a software design; then
implement it, test it, and teach it and let it loose to experience the
world.
Theres just one small catch with this marvelously direct
approach, which is that our best understanding of how the mind
works is a bit of a mess. There is no rigorous scientific theory of
cognition, and so various researchers hold to various partrigorous, part-intuitive theories, and pursue various direct
engineering style AGI approaches based on these. Ive been
undertaking considerable effort to get multiple AGI researchers
to agree on a single medium-granularity model of how the mind
works, and then formulate their respective AGI approaches as
different ways of filling in the details of this model. I think I may
be able to corral at least a reasonable plurality of AGI
researchers into this sort of conceptual agreement, but even in
the most optimistic case, it will take a few more years of
133
community-building and collaboration within the AGI community

to get there.
Any categorization of the broad spectrum of direct AGI
approaches will necessarily be incomplete and limiting there
will be cases that span multiple categories, and cases that dont
24
seem to fit into any of the categories, etc. But nonetheless the
human mind likes to categorize, so here goes! The first
dichotomy I would pose is between experiential learning based
and encoded knowledge based approaches.
Experiential Learning Based: In this kind of AGI approach, the
focus is on automated pattern discovery, wherein the system
grows its own intelligence via analyzing the data collected from
its sensory inputs. In this approach high level structures and
skills (such as facility for language, planning etc.) are derived
from incremental accumulation of regularities in sensory inputs
by inductive inference. There will always be some structure built
into the system, to guide what kind of learning is done, but this
structure doesnt constitute specific knowledge content about the
world, its more high-level and implicit knowledge in the form of
biases regarding how to learn about the world.
Encoded Knowledge Based: In this kind of AGI approach, the
system is started with hand-coded knowledge, rather than
primarily or exclusively learning via observation or experience.
For instance, the facts that people and cats are animals might be
written in some formal language (e.g. cat isa animal, human isa
24
A common way to categorize AI or AGI approaches is to talk about

symbolic (e.g. logic-based) versus subsymbolic (e.g. neural net
approaches) but, though this dichotomy is historically important in the
AGI field, I think its actually somewhat confusing. It works better as a
classification of practical AI systems built so far, than as a classification
of the larger designs and ideas underlying these systems. So Ill take a
somewhat different tack here (though coming to the symbolic vs.
subsymbolic distinction along the way).
134
Approaches to AGI
animal) and then fed into the AI system from a file or via an
interactive interface. Knowledge may also be fed into the system
via natural language, using language parsers with specially
contrived semantic analysis modules intended to enable the
semantically unambiguous interpretation of simple cases of
English sentences.
Hybrid: Of course, its also possible to build AGI systems that
derive knowledge both experientially and via hand-coded rules.
Indeed, nearly everyone building systems relying on hand-coded
knowledge rules intends to eventually have their system handle
and learn from experiential data. And my own OpenCog
approach, though fundamentally experiential learning based,
enables loading in of knowledge rules as an option. However, in
practice, nearly all in-principle hybrid AGI systems have a
primary orientation toward either experiential learning or handcoded rules. OpenCogs primary orientation is experiential
learning based, as is seen from the fact that the system can
meaningfully operate without hand-coded rules, but now without
experiential learning. On the other hand, many of the Good Old
Fashioned AI systems from the 1970s and 80s (e.g. Cyc and
Soar, which Ill discuss in the next chapter) are clearly primarily
oriented toward hand-coded rules: They cant be run without
them, and in practice tailoring them for a new application is
mainly a matter of adding new rules to the rule-base.
I dont think the hand-coding of knowledge a good approach to
AGI, but for a long while it was the dominant approach in the AI
community, so its worth understanding because of its historical
value and the influence its had on the field. For now, though,
lets focus on the experiential learning approach (including hybrid
approaches that are primarily experiential learning focused). I
think this sort of approach is much more promising. Experiential
learning based approaches can be categorized via the amount of
biasing that is provided to the learning.
Minimal Biasing: There are AGI systems that try to do
experiential learning with as little a priori biasing as possible. Of
135
course, any real-world system will necessarily have some bias,

but one can try to minimize this, for example by not intentionally
building in any biases.
The field of reinforcement learning based AI tends to have a
philosophy of minimal biasing. For instance, at the AGI-11
conference there was an interesting debate between Richard
Sutton, generally considered the founder of reinforcement
learning based AI, and Itamar Arel, an AGI researcher who is
also a big fan of reinforcement learning, but who advocates a
hierarchically structured pattern recognition architecture (which
Ill discuss a little later). Itamars AGI architecture consists of a
collection of pattern recognizers arranged into hierarchies,
intended to recognize patterns in the systems experience, but
biased via its very architecture to be extra-good at recognizing
hierarchical patterns. On the other hand, Suttons reinforcement
learning approach doesnt build in any hierarchical structure, or
any other kind of explicit structure. Suttons view was: If a bias
to recognize hierarchical patterns is important, then the system
should learn that bias via experience, via studying the patterns in
the data it obtains from the world. Itamars view was: Yes, that
sort of learning is possible in principle, but it would take a very
long time and a very large amount of data, and after all the
human brain has a built-in hierarchical structure in many regions,
e.g. the visual and auditory cortex.
The point about the brains built-in hierarchical and other
structures is interesting, because of the brains tremendous
plasticity. Sometimes, if a part of the brain is damaged or
removed, another part of the brain will take its place, adopting
the structures previously possessed by the missing part. So its
not just that the brain has certain pre-wired structures; it has a
pre-wired (i.e. evolved) propensity to grow certain kinds of
structures when the conditions are appropriate.
AI researcher Ben Kuipers did some interesting experiments in
which he got an AI system to learn that it lives in a 3 dimensional
world. Rather than preprogramming the system to interpret its
136
Approaches to AGI
input data as 3 dimensional, he fed the input data to the system

in a less structured way, and forced the system to infer the
dimensionality of space from the data.
Interesting and
impressive though not shocking, as it stands to reason data
from N dimensions is going to look more comprehensible and
less confusing if interpreted as coming from N dimensions rather
than from a space with some other dimensionality.
Significant Biasing: While a minimal-biasing approach is
elegant and interesting, many of us working on experientiallearning-based AGI systems are convinced that the most
practical route is to build a considerable amount of structure into
our AGI systems, so as to provide them with biases to learn
particular sorts of things in particular sorts of contexts. The
catch, though, is that theres no such wide agreement on what
sort of biasing is appropriate.
Ultimately biasing is just another way of referring to cognitive
architecture if you have an experiential learning system, then
its internal cognitive architecture is precisely what biases it to
learn in certain ways, and to learn some things more easily and
readily than others. However, if one has a system that is capable
of revising its cognitive architecture based on its experiential
learning, then the cognitive architecture is only an initial biasing,
and could ultimately end up getting totally replaced based on
what the system learns.
If you want to make a categorization of AGI approaches more
fine-grained than this, things become even trickier, and there are
many different directions to go in. Two ways to subdivide the AGI
approaches in the significant biasing category are: unified
versus heterogenous, and explicit versus emergent. These two
dichotomies are independent of each other (either unifies or
heterogeneous can be paired with either explicit or emergent)
and also tend to be fuzzy in the cases of most AGI systems. But
the underlying ideas are worth understanding.
137
Unified Versus Integrative

Unified: In some AGI approaches there is a single unified
approach and architecture to the biasing-structure that is built in.
One example is Itamar Arels system mentioned above, which
like Jeff Hawkins Numenta and several other approaches
involves a uniform hierarchy of pattern recognition units. Another
sort of example would be a purely logic-based AGI system, in
which a logical theorem-prover was used to do everything; or a
pure evolutionary learning approach or a uniform neural net
architecture.
The unified approach has the benefit of elegance and simplicity,
and it seems to appeal to many computer scientists, I think
because in computer science one is always seeking a single,
simple, elegant algorithm to solve the problem at hand. The
mathematician in me craves an elegant, unified approach to
biasing learning as well. Maybe someone will discover one
someday.
However, whenever I think about this topic, I keep coming back
to the incredibly messy and heterogeneous nature of the human
brain. The brain is not an elegant unified architecture each
region seems to organize and process information in its own
peculiar way, via complex communications and collaborations
with other regions. Each region leverages the multiple types of
neurons and neurotransmitters in different ways. Its a beautiful
complex mess, in a way, but its still a complex mess, apparently
totally lacking the elegance and simplicity of a well-honed
computer science algorithm.
To take a very simple example, Ive done some work recently
building computational models of the brains representation of
physical space, e.g. the 2D layout of a town or a wooded area.
One part of the brain, the hippocampus, represents a third
person top down map of space, with neurons carrying out
specialized roles in this regard, such as grid cells that respond
to corners of a spatial coordinate lattice. Another part of the
138
Approaches to AGI
brain, within the parietal cortex, represents a first person map

of space actually both a face-centered map and an eyecentered map. The hippocampus and the parietal cortex work
together closely to keep these maps well-coordinated with each
other. The hippocampus carries out certain kinds of patternrecognition on the map it stores, the parietal cortex carries out
certain kinds of pattern-recognition on the maps it stores, and
then the cell assemblies binding these regions integrate these
various patterns (in ways that are not at all well understood).
This is a relatively simple sort of cognitive activity, and even so
were just now barely beginning to understand how the brain it
but one thing that is clear from what we do understand, is that
the brain does it using a variety of fairly specialized neural
architectures. Regarding spatial pattern recognition, the
hippocampus performs one kind of learning-biasing, the parietal
cortex performs another, and the two work together
synergetically via complex pathways of coordination. Actually,
we do this particular thing much more simply and elegantly in
OpenCog: We just have a single third-person-view map of the
world the system knows, and represent first-person views as
restricted views of that primary map.
So when I see an elegant hierarchical pattern recognition
architecture like Itamar Arels or Jeff Hawkins, my first thought
is: OK, that may be a nice qualitative model of visual and
auditory cortex and some parts of cognitive cortex But what
about the olfactory (smell) and somatosensory (touch and
kinesthesia) cortex, which are not so markedly hierarchically
structured? What about the rest of cognitive cortex, which has as
many tangled-up combinatory connections as hierarchical
connections, because it originated evolutionarily from the
reptiles olfactory bulb which is dominated by combinatory
connections, rather than from a largely hierarchical subsystem
like visual cortex? What about the hippocampus, what about the
thalamus, and the many cross-connections between these nonhierarchical systems and the cortex?
139
Of course, neither Arel nor Hawkins is trying to model the brain,

theyre trying to build AGI systems and its certainly rational for
them to argue that one particular architecture, which part of the
brain uses, actually has the capability to do everything the brain
does, even though the brain doesnt use it that widely. But my
own intuition says that, even if one isnt trying to emulate the
brain, theres a valuable lesson to be learned from the wild
heterogeneity of the brain. I suspect the brains heterogeneity is
a well-justified response to the task of displaying general
intelligence for the tasks humans must generally face, in the
environments for which humans evolved, given the limited
resources the human body possesses.
Integrative: The alternative to the unified approach is the
modular or integrative approach, in which various different
biasing-structures (modules) are linked together to form a
unified architecture. The category of modular systems may be
further subdivided into loosely-interconnected (black box) versus
tightly-interconnected (white box).
In a loosely interconnected system, the different modules are
black boxes to each other, i.e. they cant see what each other
are doing. So in a system like this, a vision processing module
might pass information regarding what it sees to a semantic
reasoning module, and the latter might pass the former
information regarding what it should expect to see but the
communication is restricted to the exchange of packets of
information like this. The two modules dont have awareness of
each others internal states, and dont try to guide each others
processing.
In a tightly interconnected system, the different modules are
white boxes that can see significant aspects of each others
internal dynamics, with the result that they can try to nudge each
others learning in the right direction. So for instance, the
cognition model could do more than just tell the vision module
expect to see a face in the dark in front of you it could
140
Approaches to AGI
actually help the vision module adjust some of its internal

parameters to better recognize the face.
I feel strongly that tight interconnection is key to general
intelligence. In my view, for powerful experiential learning to
occur in real-world environments under realistic resource
constraints, very substantial biasing of learning is needed, which
means you need a fairly substantial cognitive architecture. Since
our environments and goals have so many diverse aspects, a
unified algorithm is going to be difficult to come by, and an
integrative approach is probably more realistic. But having
multiple modules that perceive each other as black boxes isnt
likely to work, because theres no good defense against one of
the modules getting stuck and having stupid ideas and failing to
learn properly you need the modules to be able to see
something of each others inner workings so they can better help
each other out. The modules need to be more like close
collaborators doing agile software development in a startup, than
like paranoid bureaucrats passing each other formalized
information, keeping each other at arms length and on a need
to know basis. Most of the work I did when designing the
OpenCog system, went into ensuring the various learning
modules would work well together according to a particular
approach I call cognitive synergy.
A large question that emerges in the context of the modular
approach is: Where do you get the breakdown into modules
from? Mathematically and conceptually, there are many, many
ways to break down a learning system into different modules. If
you dont use the brain as guidance, where do you turn? Theres
no systematic, universal theory of real-world general intelligence
to turn to, so what do you do?
A fair number of AGI researchers, including myself, have turned
to cognitive science an interdisciplinary field at the
intersection of psychology, computer science, linguistics,
philosophy and neuroscience (with other disciplines sometimes
thrown into the mix as well)! I helped found two university
141
cognitive science programs, early in my career: one at the

University of Waikato in Hamilton, New Zealand, and one at the
University of Western Australia in Perth. In each case I found the
cross-disciplinary collaboration extremely scintillating and in
fact, it was in the context of the cognitive science group at the
University of Western Australia that I came up with the first AGI
designs vaguely similar to OpenCog. (My AGI designs before
that were more in the simple and elegant vein, based on novel
self-improving learning algorithms that I implemented in the
beautiful and then-very-inefficient programming language
Haskell, with very little practical success.)
Cognitive science is nowhere near having a full understanding of
the human mind, at this point. However, one thing it has
achieved is a reasonable understanding of how the human mind
can be meaningfully broken down into modules. The modules
are not always all that distinct, and they are definitely tightly
interconnected. But over the last few decades of cognitive
science research, something like a majority understanding of the
high level boxes and lines diagram of the human mind has
emerge. This is not enough to guide the construction of an AGI
system in any detail. However, its enough to guide the modular
structure of an AGI system. One is then left with the itsy bitsy
problem of putting appropriate representational and dynamical
learning mechanisms inside the modules, taking cognitive
synergy into account at every step.
Emergent Versus Explicit

Overlapping with the unified/modular dichotomy is the
emergent/explicit dichotomy, which has played a large and
contentious role in the history of AI. This is a topic on which
misunderstanding is rampant, even among highly knowledgeable
professional AI and AGI researchers.
(Common terminologies are symbolic vs. subsymbolic, or
symbolic vs. connectionist but these are all terms that have
142
Approaches to AGI
been given multiple meanings by various well-known

researchers. To avoid compounding the confusion that already
25
exists regarding these terms , Ive chosen to go with the
eccentric emergent vs. explicit instead!)
The basic idea of the distinction is that, in an explicit system, the
systems designer can give detailed answers to questions like
How does your system represent its knowledge about cats?, or
What steps does your system take to figure out how to navigate
through a crowded room to get to its destination? The designer
could give a general answer about the type of representation or
steps-to-improvement without looking at the systems state; and
then if they looked at the systems state, they could read off the
detailed answers to the questions without a huge amount of
hassle.
In an implicit system, on the other hand, the only way for the
systems designer to answer a question like this would be to
carefully study the structures and dynamics that had emerged
within the system as a consequence of its engagement with the
world. In other words, the systems designer would be in almost
the same position as a neuroscientist asked similar questions
about a human brain except that in the case of an AI, its
possible to gather accurate data about the internal state. So the
systems designer would be in a position roughly similar to that of
25
To see why the symbolic and connectionist terminology is

confusing, consider: Any human-level AGI system is likely to learn
symbols of some kind. If a purely implicit, self-organizing neural net
system chooses to create neurons or small neural assemblies that have
clear interpretations as symbols, does that make it a symbolic system,
or not? And connectionist seems clear when you compare a GOFAI
system versus a neural network, but it becomes less clear when you
realize that most GOFAI systems could be reformulated as semantic
networks, i.e. nodes connected to other nodes by various types of links,
with rules for updating the nodes and links depending on their types. So
is a semantic network implementing a GOFAI type system
connectionist or not? it would certainly have a lot of connections!
143
a neuroscientist with really amazing brain imaging tools and a

thorough knowledge of the lowest-level components of the brain.
This is obviously not a very strict dichotomy, in principle,
because there can be varying degrees of difficulty in answering
specific questions about how a systems internal operations
correspond to its knowledge or its behaviors. But the dichotomy
has assumed an outsized role in the history of AI because most
of the dominant AI approaches have tended to fall very strongly
on the explicit or implicit side of the dichotomy.
The GOFAI systems that dominated the AI field in the 70s and
80s tended to rely on human-encoded knowledge, and tended to
store this knowledge internally in an extremely transparent way.
For instance, you might feed such a system knowledge in the
form cat isa animal, and then the system would store this in an
internal data structure equivalent to a node for cat, a node for
animal, and a link labeled isa between them. Here the
knowledge representation is extremely implicit. Similarly, the
traditional AI planning algorithms follow an easy-to-understand,
highly routinized step-by-step approach, so that its pretty easy to
foresee how such a planner will approach a problem like
26
navigating to a new location.
26
A note to the reader who has studied a little math or AI: Many of these GOFAI
systems used varieties of formal logic to carry out reasoning and learning, and
because of this, the impression spread in the AI community that formal logic
based systems are intrinsically tied to an explicit, human encoded knowledge
based AGI approach. But actually this is not accurate. It would be quite
possible to make a formal logic based AI system that operated on a

pure experiential learning basis; and in this case, the representation of
knowledge would most likely be implicit, as it would be difficult for
anyone to figure out what combinations of perceptual logical primitives
the system had learned to use to represent everyday objects and
events. Whats not clear is whether a purely logic-based approach
could be made scalable enough to serve as a practical AGI system.
Im more optimistic about including logic in hybrid systems, and this is in
fact done in OpenCog.
Im still shocked by how many smart,
144
Approaches to AGI
On the other hand, purely connectionist systems such as the

majority of neural net models, take a much more implicit
approach theres no easy way to see why the system has
learned what it has, or how the system represents the knowledge
it has acquired. Using statistical and machine tools its often
possible to perform some analysis and figure out these things,
but then it becomes a research project similar to figuring out how
the brain does things internally. For instance, in many cases one
can apply a statistical method called principal components
analysis to study a formal neural network, and then the method
produces a series of principal components that approximately
correspond to the main memories that the system has learned.
This is very useful, but its different than looking right into a
GOFAI system knowledge base and seeing cat isa animal.
As youve probably guessed by now, OpenCog takes a hybrid
approach, with some explicit and some implicit representation.
But if one side of the dichotomy had to be chosen, wed have to
choose implicit for OpenCog according to the underlying
theory, the system should be able to achieve human-level
general intelligence without any explicit representations, but it
couldnt do so without a lot of implicit representations.
professional AI and AGI researchers believe that a system containing a

logic-based component is somehow intrinsically unable to carry out
experiential learning!
One of the ironic things about this is that
OpenCogs probabilistic logic equations are real-valued functions not
too dissimilar from the mathematical functions used to implement neural
net models. And there are approaches to logical inference that are fully
parallel and distributed just like a neural net (and more so than some
neural net learning algorithms). On the math level, the logicist versus
connectionist distinction doesnt amount to that much one could argue
theres a significant conceptual difference, but I think even thats been
far overblown. Im often bemused by the way our thought is directed by
the somewhat arbitrary categories we adopt to guide our
communication.)
145
If you havent had any exposure to the AI field before, this brief
run-through of various approaches to AI may have been slightly
dizzying and confusing. But Ill ask you to please stick with me
through the next few chapters, where Ill discuss a number of
AGI approaches in detail including many of the ones briefly
alluded to above. Among other thing these chapters will give
some concrete examples of what various AI systems in the
various categories have done already examples often make
things a bit clearer! For the rest of this chapter, Im going to
discuss some more high-level, less nitty-gritty aspects of the
quest to create powerful AGI.
Theory Versus Experimentation in AGI

Development
Science and engineering generally advance hand in hand, but in
different cases, one or the other may take the lead. The science
of quantum physics advanced rather far before practical
applications emerged. On the other hand, the Wright Brothers
demonstrated manned powered flight well before we had a
mathematical theory of aerodynamics solid enough to guide the
particulars of aircraft design. There is a variety of opinions about
whether AGI should follow the quantum physics approach, the
Wright Brothers approach, or something in between.
On the quantum physics side, some people think that creating
AGI is going to take a really elegant formal scientific theory of
intelligence and that after we have that theory, building AGI will
be fairly straightforward.
Some very serious researchers are pursuing this approach for
example Marcus Hutter, a German AI researcher now based at
the Australian National University, has developed an abstract
theory of AGI that tells you a lot about how to create powerful
AGI systems in the case where you have infinite (or at least
impracticably massive) computational resources at your
disposal. Hes trying to scale the theory down, and render it
146
Approaches to AGI
applicable to more realistic cases. If this approach works, then

eventually AGI implementation will be guided strongly by
mathematical AGI theory. His approach would fall into the
biased experiential learning camp, and its not clear at this
stage whether a future, more practical version of his general
theory would tend toward unified or modular biasing-structures.
Recently hes created some nice narrow AI algorithms inspired
by his general theory, which are good at solving certain sorts of
pattern recognition problems.
In spite of my background as a mathematician, these days I tend
to fall closer to the Wright Brothers camp I think we could
build an AGI now, with our current fairly thorough but only semirigorous conceptual understanding of intelligence. In this case,
perhaps the elegant theory of general intelligence will emerge
later after we have some real AGI systems to experiment with.
Certainly, formulating a scientific theory is much easier when you
can carry out observations and experiments on the subject of
your theory!
And I do think a theory of AGI will be science and not just
mathematics, because real-world AGI is about general
intelligence in the physical world; and human-level AGI is
specifically about general intelligence in those environments for
which humans evolved. There is a certain aspect of general
intelligence that is independent of what physics your universe
has, and what kind of body you have, and so forth but I think
there are also important aspects of general intelligence that
depend on these things. A good theory of general intelligence
would tell you how general intelligence depends on these various
factors, and would then let you derive conclusions about humanlevel AGI design from information about the human body and the
environments and tasks for which humans evolved.
In fact this is my vision of what a rigorous theory of general
intelligence will look like, once we have it. I think that one will be
able to create a mathematical function whose input is a
description of an environment, and whose output is a description
147
of the cognitive architecture (the biasing-structure, in the

language I used above) needed to be reasonably generally
intelligent in that environment giving moderately limited
computational resources. I've spent some effort developing this
sort of theory but it's quite incomplete at the moment... I
sometimes find it hard to know how to balance my time between
this sort of theory and practical system-building. Im pretty
confident the OpenCog approach can work for human-level AGI,
without any massive changes or theoretical breakthroughs
required. On the one hand, a theoretical breakthrough might tell
me how to drastically simplify the design, which would reduce
the time and cost of implementation. But on the other hand,
aside from the risk that formulating a serious theory of general
intelligence proves too difficult, theres also the risk that after
doing a lot of theory, I just wind up with a more rigorous
grounding for the sort of system I'm already building.
In spite of my particular speculations about what a rigorous AGI
theory might look like, I think its also important to realize that, at
this stage, nobody really knows what a theory of AGI would or
should look like. Physics envy needs to be resisted we cant
expect an equation of thought with the simplicity and power of
Newtons Laws or the Schrodinger Equation. Biology and AI are
not physics. Biological systems are complex and messy and one
can't expect elegant unified theories like the ones seen in
physics. Most probably human-level AGI systems are going to be
complex and messy as well, due to the same sorts of resource
requirements that pushed biological systems in this direction. Yet
even complex and messy systems may obey high-level structural
and dynamical principles. I expect we will discover what these
are as we progress with the implementation, testing and teaching
of moderately powerful AGI systems, rather than before. But Ill
be pleasantly surprised if somebody comes up with a useful
breakthrough in the rigorous theory of AGI in the near future (and
even more pleasantly surprised if its me, heh!).
148
Approaches to AGI
Can Digital Computers Really Be Intelligent?

One more issue needs brief attention, before we move along. All
the specific AGI approaches Ive mentioned above make the
implicit, unstated assumption that its possible to make AGI on
digital computers. But its worth pausing a moment to point out
that this isnt proven and there are some smart, relevantly
knowledgeably people who think its definitely not true. I dont
agree with them, but their arguments are worth considering.
Some theorists argue that digital computers will never display
human-level general intelligence because mind is intrinsically a
quantum phenomenon relying on peculiar properties of
quantum physics, which (according to these theorists) manifest
themselves in the human brain, but not in digital computers. This
is actually a claim of some subtlety, given the wacky properties
of quantum computing. David Deutsch proved that quantum
computers cant compute anything beyond what ordinary digital
computers can But (a major but!) in some cases, quantum
computers can compute things much faster than digital
27
computers.
A few mavericks like Stuart Hameroff and Roger Penrose have
gone even further, and argued that non-computational quantum
gravity phenomena are at the core of biological intelligence. That
is uncowed by the theorems saying that quantum computing
cant compute anything beyond what ordinary computers can,
and inspired by an intuition that human cognition must involve
something beyond computing, they have posited that the human
brain must be some other kind of computer even stranger and
more powerful than a quantum computer! Modern physics has
not yet found a unified theory of quantum physics and
gravitational physics, and Penrose and Hameroff have exploited
this loophole to suggest that once a unified quantum gravity
27
Technically, they can compute some things much faster in the

average case not in the worse case.
149
theory is found, the brain will turn out to be a quantum gravity

computer!
Now, this is very difficult to refute, since there currently is no
good theory of quantum gravity, so nobody knows what a
quantum gravity computer is. However, its worth noting that
there currently zero evidence of cognitively significant quantum
phenomena in the brain let alone mysterious quantum gravity
phenomena!
However, a lot of things are unknown about the brainand
about quantum theory in macroscopic systems like cells; and
about quantum gravity so these points of view cant be
definitively ruled out. Even if there is something to this whole line
of thinking, though, there are an awful lot of unknowns to resolve
before you could conclude that human-level AGI is impossible on
digital computers. Maybe the brain uses weird quantum voodoo
to do some of its computing, but digital computers could achieve
even greater intelligence using other means. Maybe the
boundary between quantum and classical systems isnt as clear
28
as currently thought a few physicists have argued that it
sometimes makes sense to model large classical systems using
quantum theory in which case it could be that digital computers
are also quantum in some sense.
Another question is, even if aspects of the human brains
dynamics rely on quantum dynamics, how important are these to
emulating the brains functions? Stuart Hameroff, when he talks
about quantum theory and intelligence, likes to point to an image
of a paramecium and point out that currently we cant even
simulate a paramecium on computers, let alone a human. Thats
true. But we also cant simulate a log, yet we can build steel
struts that are better than logs at holding up buildings. He likes to
point to the incredible computational potential of the molecules in
the cell wall of a neuron but so what? The cells in my skin have
28
Diederik Aerts and Harald Atmanspacher, for example.
150
Approaches to AGI
that same computational potential, yet I really doubt the dermis

inside the heel of my foot is doing a lot of highly intelligent
computation.
And of course, even if quantum computing or quantum gravity
computing is necessary for human-level AGI (a supposition for
which there is currently zero evidence), this wouldnt rule out the
project of AGI as a whole. It would just mean we have to shift to
a different kind of computing infrastructure. DWave is already
commercializing a limited form of quantum computing, and the
quantum computing industry can be expected to explode
dramatically during the next century.
Hameroff, Penrose and others tie their quantum theory of
intelligence in with a theory of consciousness, basically claiming
that, even if a digital computer managed to impersonate
intelligence, it would never actually be conscious, because
consciousness has to do with macroscopic quantum phenomena
that can occur in brains but not digital computers. Interestingly,
Hameroff is a sort of panpsychist as well, positing that everything
in the universe has a bit of proto-consciousness, but only in
systems that achieve general intelligence via quantum gravity
computing does this develop into full-fledged consciousness.
Given the massive confusion afoot regarding both quantum
gravity and consciousness, and the early-stage nature of
neuroscience, cognitive science and AGI, this seems to me a
terrible tangle of confusions that could be unraveled only with
great effort. Personally, Im not very worried about the quantum
gravity and consciousness issue as an objection to the creation
of powerful AGI! Whatever consciousness turns out to be, I have
a very strong feeling that if a system displays similar behaviors
and similar internal structures and dynamics to a human mind,
its going to be conscious in the same sense that a human mind
is. I very much doubt it will be necessary to move beyond
classical digital computers to achieve a system with this kind of
behaviors, structures and dynamics but if it does, then well do
it!
151
As it happens I have spent a fair bit of time thinking about

femtotech, which is a hypothetical way of building computers and
other machines using elementary particles. Even if (as I suspect)
quantum gravity computation doesnt play a role in the human
brain, perhaps it will play a role in future AGI systems
implemented using femtotechnology!
Ill talk about this a bit
later, in these pages.
But that brings us rather far afield, into what I like to think of as
hi-fi sci-fi. I love thinking about such topics, but I prefer to
spend more of my time focused on present reality, on the
relatively mundane matter of creating human-level thinking
machines on digital computers.
152
This was originally an H+ Magazine article, with a title formulated

by the magazines editor at the time, the incomparable RU Sirius:
10
Will Bots Feel Joy?
Will machines ever really feel, in the same sense that humans
do?
This is at least according to many philosophies a separate
question from whether machines can be intelligent, or whether
they can act like they feel. The question is whether machines if
suitably constructed and programmed can have awareness,
passion, subjective experience... consciousness?
I certainly think so... But generally speaking there is no
consensus among experts. Its fair to say that even without
introducing machines into the picture consciousness is one of
the most confused notions in the lexicon of modern science and
philosophy.
I became acutely aware of this confusion when, in summer 2009,
I organized a workshop on Machine Consciousness in Hong
Kong, as part of the larger Toward a Science of Consciousness
conference and Asia Consciousness Festival. The TSC
conference as a whole attracted hundreds of participants, but
only a couple dozen deigned to venture into the riskier domain of
machine consciousness; and among these brave souls, I reckon
there were more than a couple dozen views on the matter at
hand!
First we have the materialists. Joscha Bach a German AI
researcher and entrepreneur and the author of Principles of
Synthetic Intelligence, who will pop up a bit later in this book
summarizes their perspective elegantly: The notion of the mind
153
as an information processing system, capable of forming an

integrated self-and-world-model, modulated by emotional
configurations and driven by a finite set of motivational urges, is
sufficient to remove the miracles [that some associate with
consciousness]. Daniel Dennett is the best-known modern
advocate of the materialist view. According to his book
Consciousness Explained, its patently obvious that machines
can be conscious in the same sense as humans if theyre
constructed and programmed correctly.
Paul Fahn, an AI and robotics researcher at Samsung
Electronics, presented this perspective at the MC Workshop in
the context of his work on emotional robots. His core idea is that
if a robot brain makes emotional decisions using a random or
pseudorandom preference oracle similar to the one in a human
brain, it will likely be emotional in roughly the same sense that
humans are and possessed of its own distinct but equally valid
form of consciousness. Fahn emphasizes the need for empirical
tests to measure conciousness and Raoul Arrabaless talk at
the workshop took concrete steps aimed in this direction,
describing a series of criteria one can apply to an intelligent
system to assess its level of consciousness.
But some, less happy with the materialist view, have referred to
Dennetts book as Consciousness Explained Away.
Neuropsychologist Allan Combs has a new book in press, called
Consciousness Explained Better in which he reviews a host of
states of consciousness, including those accessed by mystics
and meditators as well as those we feel in various unusual states
of mind, such as dreaming, sleeping, dying, etc. As a
panpsychist: He sees consciousness as the basic material of the
cosmos; he sees rocks, bugs, cows, humans and machines as
differing manifestations of universal consciousness.
To a panpsychist, the question isnt whether machines can be
conscious, but whether they can manifest universal
consciousness in a manner similar to how humans do. And the
question of whether consciousness can be empirically measured
154
Will Bots Feel Joy?
is not that critical, because theres no reason to assume the

universe as a whole is understandable in terms of finite sets of
finite data-items, of the sort that science works with. Setting
aside mystical notions, pure mathematics points to all manner of
massively infinite constructs that if they existed in reality
could never be probed via scientific measurements.
The coauthor of Combs workshop talk, creativity theorist Liane
Gabora, holds the view that machines are conscious, but will
never be nearly as conscious as humans. I put my money on
the argument that living things are more conscious than rocks or
computers because they amplify consciousness by being selforganizing, self-mending, and autopoietic; that is, the whole
emerges through interactions amongst the parts. And the human
mind amplifies consciousness even further through a second
level of autopoietic structure. Just as a body spontaneously
repairs itself when wounded, if someone does something out of
character or something unexpected happens, the mind
spontaneously tries to repair its model of the world to account for
this turn of events. This continuous building and rebuilding of a
mental model of the world, and thus reconstituting of autopoietic
structure, locally amplifies consciousness. Until computers do
this, I don't think their consciousness will go much beyond that of
a rock.
As a panpsychist myself, I find Lianes view sympathetic but
Im much more optimistic than she is that complex, selforganizing autopoietic structure can be achieved in computer
programs. Indeed, that is one of the goals of my own AI
research project!
Then there are the quantum consciousness folks, such as Stuart
Hameroff, who gave the keynote speech at the Cognitive
Informatics conference in Hong Kong, the day after the MC
workshop. An MD anesthesiologist, Hameroff was seduced into
consciousness theory via wondering about the neurobiology by
which anesthetics bring about loss of consciousness. Together
with famed physicist Roger Penrose, Hameroff developed a
155
theory that consciousness arises via quantum-mechanical

effects in structures called microtubules, that make up the cell
walls of brain cells.
A common joke about the Penrose-Hameroff theory is: No one
understands quantum theory, and no one understands
consciousness, so the two must be equal! But clearly the
theorys intuitive appeal goes beyond this: Quantum nonlocality
implies a form of interconnectedness of all parts of the cosmos,
which resonates well with panpsychism.
Penrose believes that human consciousness enables problemsolving beyond what any computer can do. To bypass theorems
showing this kind of capability wouldnt be provided by mere
quantum computing, he proposes quantum gravity computing,
based on an as-yet unknown unified theory of quantum physics
and gravitation. Most scientists view this as fascinating, highly
technical sci fi.
Regarding panpsychism, Hameroff says I disagree only slightly.
I would say that what is omnipresent in the universe is protoconsciousness... Penrose and I say proto-consciousness is
embedded as irreducible components of fundamental spacetime
geometry, i.e. the Planck scale, which does indeed pervade the
universe. He views consciousness per se as a special
manifestation of proto-consciousness: I dont think a rock
necessarily has the proper makeup for the type of quantum state
reduction required for consciousness.
A fascinating twist is suggested by recent work by Dirk Aeerts,
Liane Gabora, Harald Atmanspacher and others, arguing that
being quantum is more about being susceptible to multiple,
fundamentally incompatible interpretations, than about specific
physical dynamics. In this sense, consciousness could be
quantum even if the brain doesnt display nonclassical
microphysical phenomena like quantum nonlocality.
156
Will Bots Feel Joy?
At the time of the MC workshop, my close friend and colleague

Hugo de Garis was running a project at Xiamen University called
the Conscious Robotics Project. However, he was among the
least confident participants regarding the workshops topic:
"Explaining what consciousness is, how it evolved and what role
it plays is probably neuroscience's greatest challenge. If
someone were to ask me what I thought consciousness is, I
would say that I dont even have the conceptual terms to even
begin to provide an answer.
Australian philosopher David Chalmers whom I got to know a
bit in 2011, when we both spoke at the Australian Singularity
Summit -- provided an important clarification of the
consciousness issue when he introduced the notion of the hard
problem of consciousness namely, building the link between
the physical processes and structures and behaviors associated
with consciousness, and the actual experience of
consciousness. He contrasts this with the easy problems
(which are easy only in a relative sense!) such as characterizing
the nature of subjective experience, and figuring out what
cognitive and neural processes are associated with the
experience of consciousness. The easy problems may not really
be easy, but seem the sorts of things that can be solved by
systematic effort. The hard problem seems a basic conceptual
conundrum.
In response to this hard problem, Chalmers own conclusion
about consciousness seems to amount to a weak form of
panpsychism, in which its admitted that everything in the
universe has a little bit of proto-consciousness, which manifests
as full consciousness only in certain entities. Im not sure if this
is essentially different from my own view that I call panpsychist,
in which I view everything as a little bit conscious, and certain
29
entities as manifesting reflective, deliberative consciousness.
29
Another area of Chalmers interest is the nature and prevalence of

verbal disputes in philosophy hes tried to formalize the notion of a
157
Basically he concludes that the best way to bridge the gap posed
by the hard problem is to posit some sort of common substance
binding together the subjective-experience and objective
structural/behavioral realms.
Ive thought a great deal about some of the easy problems of
consciousness, in particular the aspects of human
consciousness via which we can reflect on ourselves, and via
which our consciousness analyzes itself, thus creating
consciousness of consciousness of consciousness of While
in a sense this is an infinite reflexive process because in this
sense consciousness contains itself, and only infinite entities can
contain themselves it can be approximated by finite structures
such as occur in the human brain as its modeled by physics.
But this is a part of the story Ill come back to a little later!
Another point worth emphasizing is that, among my colleagues
working on OpenCog, theres a large diversity of opinions on
consciousness. Panpsychists are probably in the minority. It
seems the practical work of engineering a mind is largely
independent of philosophical issues regarding consciousness
within limits, at any rate. If you believe that consciousness is
inextricably tied up with quantum phenomena, and also that
consciousness and intelligence are intertwined, then youre
obviously not going to like AI approaches focused on digital
computation!
Regarding the hard problem I think one very real possibility is
that we might create human-level, human-like AI systems before
we puzzle out the mysteries of consciousness. These AIs might
puzzle over their own consciousness, much as we do over ours.
Perhaps at the 2019 or 2029 Machine Consciousness workshop,
AIs will sit alongside humans, collectively debating the nature of
verbal dispute, and has argued that many disagreements in modern

philosophy are largely verbal rather than substantive disputes. It may
be that the difference between Chalmers panprotopsychism and my
panpsychism is merely or largely a verbal dispute!
158
Will Bots Feel Joy?
awareness. One envisions a robot consciousness researcher

standing at the podium, sternly presenting his lecture entitled:
Can Meat Feel Joy?
159
11
Deep Blue Reflections
I wrote the first version of this essay in 1997, a few days after
Deep Blues classic chess victory. It was written for the Simons
Rock College alumni email list, in response to some other alumni
asking for my thoughts on Deep Blues achievement, since I was
the best-known (and maybe the only?) AI researcher on the
list. I tweaked the essay slightly at some later date.
If youre both ancient and geeky like me, you may remember
May 11, 1997, the day when a computer program defeated the
(human) world chess champion -- an event that led many people
to think that computers were already on the verge of rivaling
human intelligence. More precisely, that was the day when, for
the first time ever, a computer defeated the world chess
champion in a standard five-game match. Deep Blue, a
computer chess system developed by Carnegie Mellon
University and IBM, split the first two games with Garry
Kasparov. The second two were draws, and the final game went
to Deep Blue. Kasparov was a sore loser. Deep Blue remained
dispassionate it hadnt been programmed with emotions.
Admittedly, this was only one match, but the tournament was not
a fluke. Previous versions of Deep Blue were already able to
consistently beat all but the greatest chess grandmasters prior to
the Kasparov match. And computer hardware has gotten far
better since then, enabling the same basic algorithms used in
Deep Blue to play even more powerfully. The capability of the
human brain, on the other hand, has remained essentially
constant since 1997.
Although Deep Blue is not a very generally-intelligent entity
according to our definition, theres something to be learned from
a study of its accomplishments and the mechanisms underlying
them. Deep Blue follows the same rules as human chess
players, but it doesnt think at all like humans. Human chess
160
players use geometric intuition and a sense of the flow of a

game. Deep Blue calculates every single possibility; then
calculates all the possible consequences of each, gives them
weights based on the probability and desirability of such a next
move from its perspective, and finally picks the next move by
culling-out that which produces the best set of the next N moves
for some number N.
Computer programmers call this recursive logic. It applies these
same fixed rules over and over and over again, constantly
referring back to the results it just obtained, and figuring out how
well it is doing. Because it can weight play options based on prior
games, it is a learning system by the ML definition, but not a
generally intelligent one.
Human beings might use recursive logic to play a very simple
game, such as tic-tac-toe, which has very few choices. But even
in tic-tac-toe, our opponents would probably object to our taking
the time to calculate out the potential consequences of every
possible move. Our minds are much too slow to play chess that
way, nor would the game be any fun if we could. Computers, by
contrast, are much, much quicker at this kind of task and do not
get bored, so recursive logic can work well for them.
Of course, every chess player extrapolates, thinking: What is
the other player likely to do next? And if he does that, what am I
going to do? And if I do that, what is he going to do? But in
humans, this kind of reasoning is augmented by all sorts of other
thought processes, many subconscious, which constitute what
we call intuition about a domain. What intuition really is is the
application of both knowledge so deeply learned that it becomes
subconsciously reflexive to apply it, and a complex type of
educated guesswork regarding the situation taken as a whole.
For Deep Blue, this kind of extrapolation is pretty much the
whole storyand it does it very, very well. Computers can
extrapolate faster and further into the future than any human.
The 1997 version of Deep Blue could evaluate about two
161
hundred million different board positions every second. This

figure can easily be increased for the cost of additional circuitry,
but it doesnt make Deep Blue truly intelligent. And this
difference in thinking is not trivial it is fundamental to why
Deep Blue is in no danger of suddenly becoming Skynet.
One way to understand the difference between Deep Blue and
human players is to think about strategy versus tactics. There is
a certain kind of creative long-range strategy that human chess
grandmasters have, but Deep Blue lacks. Deep Blue makes up
for this lack by elevating tactics to such a high level that it
assumes the role of strategy. Deep Blue is not entirely without
strategy: it carries out its superhuman tactical evaluation within
the context of a collection of pre-programmed strategies, and it is
capable of switching between one strategy and another in
response to events. But it does not think strategically, it only
thinks tactically. Deep Blue doesnt make long-range plans
involving an understanding of the overall structure of the board
as a dynamical system which changes over the course of the
whole game, or the emotional states of its opponents. It does not
innovate new plays based on an insightful, creative analysis of
chess theory, but only by accident if its brute force methods
stumble upon a more optimal solution. If it could do such things,
it would doubtless play even better. Even without strategic
creativity, it plays well enough to beat the best humans, but only
because it turns out that chess is a game which succumbs well
to a recursive logic approach.
The defeat of Kasparov by Deep Blue was symbolic because
chess is the most mentally challenging game commonly played
in the Western world. Computers became better than humans at
checkers and many other games quite some time ago. However,
there is at least one popular game that still stumps the best
computersthe Asian game called Go (or Weixi, in Chinese). At
the present time, in spite of a substantial research effort, no
existing computer program can play better than the advanced
beginner level at Go.
162
The rules of Go are very simple, compared to chess. Play is on a

19x19 grid, and stones (pieces) are placed on the intersections
of the grid, called points. The first player plays black stones, the
opponent white ones; and stones are added to the board one-byone, players alternating. Stones are not removed once placed,
but stones and groups of stones may be captured. A player calls
Atari when a capture can occur on their next move, to warn the
opponent. The game ends when it is no longer possible to make
a reasonable move. The winner is determined by the amount of
territory surrounded, less the number of their stones captured.
The trouble with Go from the computational perspective is that
from any given board position, there are hundreds of plausible
next moves rather than dozens, as in chess. Extrapolation in Go
will not get you as far as it does in chess. It would seem that, if
computers are to conquer Go, theyre going to have to do it
either with a more generally intelligent approach, or use some
more clever special-case technique than the one employed for
chess. Go is too visual, too two-dimensional, to succumb to
purely combinatorial, non-visual techniques. A world-champion
Go program would have to be intelligent at general twodimensional vision processing as well. Since Go, like chess, is
ultimately an extremely limited problem domain, a specialpurpose, unintelligent program may possibly master it eventually
by employing clever domain partitioning and parallelization
techniques in a brute force recursive logic approach, probably
coupled with hardware massively more powerful than current
systems.
However, the fact that it hasnt been mastered yet just goes to
show how far from intelligence computers really are right now
there are even very narrow domains with which they cant really
cope at this time. Human Go masters may have little or no other
things that they are true masters of, but they are not
fundamentally restricted to a single domain like Deep Blue is
all have at least average skill in the dozens of domains that
constitute daily human life.
163
In Go, high level players routinely analyze positions that arent

confined tactically to, say, a 9x9 grid. Additionally, almost any
tactical fight has strategic implications across the board that
could be worth more than the fight itselfso a great patternmatcher wins the points, but loses the war. One style common in
evenly-matched games is to go around the board losing fights,
but in such a way that ones own stones work together and
become more powerful.
The Go computer programs in existence today rely heavily on
pattern matching: taking a given, small chunk of the board, and
matching it up to a dictionary of known board situations. The best
ones are as good at reading small, enclosed life/death problems
as a mediocre tournament Go player. However, when the
problems are not so rigidly enclosed within a small region of the
board, the programs are clueless, although intuitive human
players can still see the essential principles. The best way to
wipe out such programs is to embroil them in a huge, wholeboard fight; one that is too big for the algorithm to match
properly.
Deep Blues recursive approach of elevating tactics to the level
of strategy doesnt work well for Go. Exhaustive searching over
spaces of two-dimensional patterns is much, much harder than
the kind of decision-tree search required for dealing with chess,
and will be out of the reach of computers for a while. One
suspects that something less than true intelligence will ultimately
suffice for Go, as it has for chessbut perhaps not something as
far from true intelligence as Deep Blue. Maybe a massive
increase in raw computational power will be enough, but that is
not a truly intelligent approach. A lot some speculate
everything would be gained for AGI research from devising
creative approaches to solving two-to-N-dimensional pattern
analysis, rather than just throwing hardware and simple partition
and parallelize tricks at the Go problem until it succumbs.
Deep Blue has basically the same problems as every other
Narrow AI program its too inflexible and its abilities are an
164
achievement of computer hardware getting faster, not of

software becoming intelligent. It relies on a special computer
chip, custom-designed for searching many moves ahead in
chess games. This special chip could be modified without much
hassle to apply to other similar gamescheckers, maybe
Othello. It is better at such tasks than a general-purpose CPU,
but not generally smarter. Its just an optimization of a fixed
algorithm. The ideas behind its massively parallel design and RS
6000 platform have since been used by IBM for brain simulation,
drug design, weather forecasting and other extremely useful
applications. However, Deep Blue has no insights whatsoever
into those domainsall its knowledge in these areas will be in
the form of rules encoded by smart humans, either about the
specific domain or generally for data mining. Its architecture
couldnt be modified to apply to Go, let alone to any real-world
situation.
Deep Blues chip is less like the human brain than like a human
muscle: a mechanism designed specifically for a single purpose,
and carrying out this purpose with admirable but inflexible
precision. Its rules are astoundingly simple and mechanical:
evaluate the quality of a move in terms of the quality of the board
positions to which this move is likely to lead, based on encoded
experience. Judgments based on experience are made not by
complex intuitive analogy, but by simple pattern-matching. Even
though it has experience of a world, albeit the limited world of
chess experience, it has no intuition of this worldit merely
stores its game experience in a database and mines that
database for possible moves. It can only invent new strategies
by coincidence, and cant understand themit can only hope to
happen upon a situation in which its human opponent makes a
move which causes that path to be entered in its search
algorithm. Everything is cut-and-dried, and done two hundred
million times a second. This is outstanding engineering, but it is
not human-level, human-like general intelligence.
165
12
Today Jeopardy!
Tomorrow the World?
This essay was written for H+ Magazine in 2010 shortly after
IBMs Watson computer beat the game Jeopardy on TV. Since
that time Watson has been developed as a more general
supercomputer architecture, intended for a diversity of
applications, with an initial focus on biomedicine. In 2014, a few
months before the writing of this introductory paragraph, IBM
announced the roll-out of 3 of its Watson machines in Africa (in
Kenya, Nigeria and South Africa). Discussion of the Watson
computing platform in its full generality would be worthwhile, but
this essay doesnt go that far what youre read here is my 2010
essay, just as written in the heat of the moment, right after
watching Watsons original televised moment of glory.
My initial reaction to reading about IBMs Watson
30
supercomputer and software was a big fat ho-hum. OK, I
figured, a program that plays Jeopardy! may be impressive to
Joe Blow in the street, but Im an AI guru so I know pretty much
exactly what kind of specialized trickery theyre using under the
hood. Its not really a high-level mind, just a fancy database
lookup system.
But while that cynical view is certainly technically accurate, I
have to admit that when I actually watched Watson play
Jeopardy! on TV and beat the crap out of its human
30
http://www.google.com/url?q=http%3A%2F%2Fmashable.com
%2F2011%2F02%2F11%2Fibm-watsonjeopardy%2F&sa=D&sntz=1&usg=AFQjCNGB8xrYzlgsmOlerj0FZ
4lncmKX0A
166
Today Jeopardy!, Tomorrow the World?

31
opponents I felt some real excitement And even some

pride for the field of AI. Sure, Watson is far from a human-level
AI, and doesnt have much general intelligence. But even so, it
was pretty bloody cool to see it up there on stage, defeating
humans in a battle of wits created purely by humans for humans
playing by the human rules and winning.
I found Watsons occasional really dumb mistakes made it seem
almost human. If the performance had been perfect there would
have been no drama but as it was, there was a bit of a charge
in watching the computer come back from temporary defeats
induced by the limitations of its AI. Even more so because Im
wholly confident that, 10 years from now, Watsons descendants
will be capable of doing the same thing without any stupid
mistakes.
And in spite of its imperfections, by the end of its three day
competition against human Jeopardy champs Ken Jennings and
Brad Rutter, Watson had earned a total of $77,147, compared to
$24,000 for Jennings and $21,600 for Rutter. When Jennings
graciously conceded defeat after briefly giving Watson a run
for its money a few minutes earlier he quoted the line And I,
32
for one, welcome our new robot overlords .
In the final analysis, Watson didnt seem human at all its IBM
overlords didnt think to program it to sound excited or to
31
http://www.google.com/url?q=http%3A%2F%2Fmashable.com
%2F2011%2F02%2F16%2Fibms-watson-supercomputerdefeats-humanity-injeopardy%2F&sa=D&sntz=1&usg=AFQjCNHFrrjPD_av8QawOg08BbtqKfiiQ
32
http://www.google.com/url?q=http%3A%2F%2Fknowyourmeme
.com%2Fmemes%2Fi-for-one-welcome-our-new-xoverlords&sa=D&sntz=1&usg=AFQjCNH_rlurRwsBmGv9maKY96z
uYQrpSg
167
celebrate its victory. While the audience cheered Watson, the

champ itself remained impassive, precisely as befitting a
specialized question-answering system without any emotion
module.
What Does Watson Mean for AI?

But who is this impassive champion, really? A mere
supercharged search engine, or a prototype robot overlord?
A lot closer to the former, for sure. Watson 2.0, if there is one,
may make fewer dumb mistakes but its not going to march
out of the Jeopardy! TV studio and start taking over human jobs,
winning Nobel Prizes, building femtofactories and spawning
Singularities.
But even so, the technologies underlying Watson are likely to be
part of the story when human-level and superhuman AGI robots
finally do emerge.
Under Watsons Hood

But how exciting is Watson from an AI point of view? How much
progress does it constitute toward AI programs capable of broad
human-level general intelligence? When will Watson or its kin
march out of the Jeopardy! TV studio and start taking over
human jobs, winning Nobel Prizes, building femtofactories and
spawning Singularities?
To understand the answer this question, you have to understand
what Watson actually does. In essence its a triumph of the
branch of AI called natural language processing (NLP) which
combines statistical analysis of text and speech with handcrafted linguistic rules to make judgments based on the syntactic
and semantic structures implicit in language. So, Watson is not
an intelligent autonomous agent like a human being, that reads
information and incorporates it into its holistic world-view and
understands each piece of information in the context of its own
168
self, its goals, and the world. Rather, its an NLP-based search
system a purpose-specific system that matches the syntactic
and semantic structures in a question with comparable structures
found in a database of documents, and in this way tries to find
answers to the questions in those documents.
Looking at some concrete Jeopardy! questions may help make
the matter clearer; here are some random examples I picked
33
from an online archive .
1. This -ology, part of sociology, uses the theory of
differential association (i.e., hanging around with a bad
crowd)
2. "Whinese" is a language they use on long car trips
3. The motto of this 1904-1914 engineering project was
"The land divided, the world united"
4. Built at a cost of more than $200 million, it stretches from
Victoria, B.C. to St. John's, Newfoundland
5. Jay Leno on July 8, 2010: The "nominations were
announced today... There's no 'me' in" this award
(Answers: criminology, children, the Panama Canal, the TransCanada Highway, the Emmy Awards.)
Its worth taking a moment to think about these in the context of
NLP-based search technology.
Question 1: This -ology, part of sociology, uses the theory of
differential association (i.e., hanging around with a bad crowd)
This stumped human Jeopardy! contestants on the show, but Id
expect it to be easier for an NLP based search system, which
can look for the phrase differential association together with the
morpheme ology.
33
http://www.j-archive.com/showgame.php?game_id=3561
169
Question 2: "Whinese" is a language they use on long car trips

This is going to be harder for an NLP based search system than
for a human But maybe not as hard as one might think, since
the top Google hit for whine long car trip is a page titled
Entertain Kids on Car Trip, and the subsequent hits are similar.
The incidence of kids and children in the search results
seems high. So the challenge here is to recognize that whinese
is a neologism and apply a stemming heuristic to isolate whine.
Questions 3 and 4:
3. The motto of this 1904-1914 engineering project was "The
land divided, the world united"
4. Built at a cost of more than $200 million, it stretches from
Victoria, B.C. to St. John's, Newfoundland
These will be easier for an NLP based search system with a
large knowledge base than for a human, as they contain some
very specific search terms.
Question 5:
Jay Leno on July 8, 2010: The "nominations were announced
today... There's no 'me' in" this award
This is one that would be approached in a totally different way by
an NLP based search system than by a human. A human would
probably use the phonological similarity between me and
Emmy (at least thats how I answered the question). The AI
can simply search the key phrases, e.g. Jay Leno July 8, 2010
award and then m any pages about the Emmys come up.
Now of course a human Jeopardy! contestant is not allowed to
use a Web search engine while playing the show this would be
cheating! If this were allowed, it would constitute a very different
kind of game show. The particular humans who do well at
Jeopardy are those with the capability to read a lot of text
170
containing facts and remember the key data without needing to

look it up again. However, an AI like Watson has a superhuman
capability to ingest text from the Web or elsewhere and store it
internally in a modified representation, without any chance of
error or forgetting The same way you can copy a file from one
computer to another without any mistakes (unless theres an
unusual hardware error like a file corruption).
Watson can grab a load of Jeopardy!-relevant Web pages or
similar documents in advance, and store the key parts precisely
in its memory, to use as the basis for question answering. Next,
it can do a rough (though somewhat more sophisticated)
equivalent of searching in its memory for whine long car trip
or Jay Leno July 8, 2010 award and finding the multiple results,
and then statistically analyzing these multiple results to find the
answer.
Whereas a human is answering, many of these questions based
on much more abstract representations, rather than by
consulting an internal index of precise words and phrases.
Both of these strategies the Watson strategy and the human
strategy are valid ways of playing Jeopardy! But, the human
strategy involves skills that are fairly generalizable to many other
sorts of learning (for instance, learning to achieve diverse goals
in the physical world), whereas the Watson strategy involves
skills that are only extremely useful for domains where the
answers to ones questions already lie in knowledge bases
someone else has produced.
The difference is as significant as that between Deep Blue and
Gary Kasparovs approach to chess. Deep Blue and Watson are
specialized and brittle; Kasparov, Jennings and Rutter, the chess
playing humans, are flexible, adaptive agents. If you change the
rules of chess a bit (say, tweaking it to be Fisher random chess,
which changes the original position of pieces on the board),
Deep Blue has got to be reprogrammed a bit in order to perform
correctly. However, Kasparov can adapt. If you change the
171
scope of Jeopardy to include different categories of questions,

Watson would need to be retrained and retuned on different data
sources, but Jennings and Rutter could adapt. And general
intelligence in everyday human environments especially within
contexts such as novel science or engineering is largely about
adaptation; creative improvisation in the face of the
fundamentally unknown. Adapting is not just about performing
effectively within clearly-demarcated sets of rules.
Wolfram on Watson
Stephen Wolfram, the inventor of Mathematica and Wolfram
34
Alpha, wrote a very clear and explanatory blog post on Watson ,
contrasting Watson with his own Wolfram Alpha system:
In his article he also gives some interesting statistics on search
engines and Jeopardy!, showing that a considerable majority of
the time, major search engines contain the answers to the
Jeopardy! questions in the first few pages. Of course, this
doesnt make it trivial to extract the answers from these pages,
but it nicely complements the qualitative analysis I gave above
where I looked at 5 random Jeopardy! questions, and helps give
a sense of whats really going on here.
Neither Watson nor Alpha uses the sort of abstraction and
creativity that the human mind does, when approaching a game
like Jeopardy! Both systems use pre-existing knowledge bases
filled with precise pre-formulated answers to the questions they
encounter. The main difference between these two systems, as
Wolfram observes, is that Watson answers questions by
matching them against a large database of text containing
questions and answers in various phrasings and contexts,
whereas Alpha deals with knowledge that has been imported into
34
http://blog.stephenwolfram.com/2011/01/jeopardyibm-and-wolframalpha/
172
it in structured, non-textual form, coming from various

databases, or explicitly entered by humans.
Kurzweil on Watson
Ray Kurzweil has written glowingly of Watson as an important
35
technology milestone
"Indeed no human can do what a search engine does, but
computers have still not shown an ability to deal with the
subtlety and complexity of language. Humans, on the other
hand, have been unique in our ability to think in a
hierarchical fashion, to understand the elaborate nested
structures in language, to put symbols together to form an
idea, and then to use a symbol for that idea in yet another
such structure. This is what sets humans apart.
That is, until now. Watson is a stunning example of the
growing ability of computers to successfully invade this
supposedly unique attribute of human intelligence."
I understand where Kurzweil is coming from, but nevertheless,
this is a fair bit stronger statement than I'd make. As an AI
researcher myself I'm quite aware of the all subtlety that goes
into "thinking in a hierarchical fashion", "forming ideas", and so
forth. What Watson does is simply to match question text
against large masses of possible answer text -- and this is very
different than what an AI system will need to do to display
human-level general intelligence. Human intelligence has to do
with the synergetic combination of many things, including
linguistic intelligence but also formal non-linguistic abstraction,
non-linguistic learning of habits and procedures, visual and other
sensory imagination, creativity of new ideas only indirectly
related to anything heard or read before, etc. An architecture like
Watson barely scratches the surface!
35
http://www.pcmag.com/article2/0,2817,2376027,00.asp
173
Ray Kurzweil knows all this about the subtlety and complexity of
human general intelligence, and the limited nature of the
Jeopardy! domain so why does Watson excite him so much?
Although Watson is just an NLP-based search system, its still
not a trivial construct. Watson doesn't just compare query text to
potential-answer text, it does some simple generalization and
inference, so that it represents and matches text in a somewhat
abstracted symbolic form. The technology for this sort of process
has been around a long time and is widely used in academic AI
projects and even a few commercial products but, the Watson
team seems to have done the detail work to get the extraction
and comparison of semantic relations from certain kinds of text
working extremely well. I can quite clearly envision how to make
a Watson-type system based on the NLP and reasoning
36
software currently working inside our OpenCog AI system
and I can also tell you that this would require a heck of a lot of
work, and a fair bit of R&D creativity along the way.
Kurzweil is a master technology trendspotter, and he's good at
identifying which current developments are most indicative of
future trends. The technologies underlying Watson aren't new,
and don't constitute much direct progress toward the grand goals
of the AI field. What they do indicate, however, is that the
technology for extracting simple symbolic information from
certain sorts of text, using a combination of statistics and rules,
can currently be refined into something highly functional like
Watson, within a reasonably bounded domain. Granted, it took
an IBM team 4 years to perfect this, and, granted, Jeopardy! is a
very narrow slice of life but still, Watson does bespeak that
semantic information extraction technology has reached a
certain level of maturity. While Watson's use of natural language
understanding and symbol manipulation technology is extremely
narrowly-focused, the next similar project may be less so.
36
http://opencog.org
174

Am I as excited about Watson as Ray Kurzweil? Not really. It's
an excellent technical achievement, and should also be a
publicity milestone roughly comparable to Deep Blue's chess
victory over Kasparov. However, answering question doesn't
require human-like general intelligence unless formulating the
answers involves improvising in a conceptual space not
immediately implied by the available information Which is of
course not the case with the Jeopardy! questions.
Rays response does contain some important lessons, such as
the value of paying attention to the maturity levels of
technologies, and what the capabilities of existing applications
imply about this, even if the applications themselves aren't so
interesting or have obvious limitations. But its important to
remember the difference between the Jeopardy! challenge and
other challenges that would be more reminiscent of human-level
general intelligence, such as
Holding a wide-ranging English conversation with an

intelligent human for an hour or two
Passing the third grade, via controlling a robot body

attending a regular third grade class
Getting an online university degree, via interacting with

the e-learning software (including social interactions with
the other students and teachers) just as a human would
do
Creating a new scientific project and publication, in a

self-directed way from start to finish
What these other challenges have in common is that they require

intelligent response to a host of situations that are unpredictable
in their particulars so they require adaptation and creative
improvisation, to a degree that highly regimented AI
175
architectures like Deep Blue or Watson will never be able to

touch.
Some AI researchers believe that this sort of artificial general
intelligence will eventually come out of incremental
improvements to narrow AI systems like Deep Blue, Watson
and so forth. Many of us, on the other hand, suspect that
Artificial General Intelligence (AGI) is a vastly different animal. In
this AGI-focused view, technologies like those used in Watson
may ultimately be part of a powerful AGI architecture, but only
when harnessed within a framework specifically oriented toward
autonomous, adaptive, integrative learning.
But, having said all that... Make no mistake when Watson does
face off against the human Jeopardy! champs, I'll be cheering for
the computer! Not out of disloyalty for my fellow humans, but out
of excitement at the possibility of seeing a sign of dramatic
progress in the practical execution of certain aspects of AI. One
more step along the path!
176
13
Chatbots Versus Cognition Bots
The ability to hold everyday, human-like conversations in English
or other natural languages occupies a special place in the AI
field. AI these days is a diverse endeavor, encompassing many
areas with little relationship to everyday conversation robotics,
mathematical theorem proving, planning and scheduling, fraud
detection, financial prediction, etc. And in practical terms, while
reproducing the ability to hold an everyday human-like
conversation would surely have great economic value, it is hardly
a worthy end goal for the AI field Id prefer an AI with
awkwardly robotic conversation but massively superhuman
scientific, ethical and artistic capabilities, over an AI with the
capability for fluent human-like cocktail part chatter but nothing
else. But nonetheless, partly for historical reasons and partly due
to its conceptual simplicity, the task of emulating human
conversational ability is one of the first thing that comes to mind
when one thinks about AIs that are as smart as people.
The historical reason is a 1955?? paper by Alan Turing, who
proposed naturalistic human-like conversation as a sufficient
condition for artificial intelligence. That is, he suggested that if
an AI could hold an ordinary conversation in a manner
indistinguishable from a human being, then it should be
considered to possess intelligence in the same sense that
humans do. He proposed to assess indistinguishability from a
human using a panel of judges if an AI could talk to the judges
and fool them into thinking they were talking to a human, then,
according to Turing, the AI should be considered to possess
human-like, human-level intelligence.
Its sometimes forgotten that Turing proposed this criterion, now
called the Turing Test, mainly as a counterargument against
those who doubted the meaningfulness of calling any computer
program intelligent in the same sense as humans. Turings
177
point was: If it converses as intelligently as a human, we should

consider it as intelligent as a human. He wasnt trying to assert
that human-like conversation should be the end goal or the
interim criterion of AI research.
Today, no AI program comes close to fulfilling Turings test but
of course, many clever programmers have made efforts at
creating conversational conversation agents, aka chat bots.
There is even a contest every year, the Loebner Prize, rewarding
the chatbot that fools the greatest percentage of human judges.
So far, chatbot development really hasnt had much to do with
the mainstream of AI research, nor with the sub-stream of AI
research pointing directly toward AGI. By and large its fair to
characterize most of the work on chatbots as narrow AI systems
aimed at achieving the illusion of human-like comprehension,
rather than as systems genuinely intended to understand what
they are talking about. Largely the pursuit of chatbots, so far, has
been a quest to see how far one can get with smoke and
mirrors, without having to confront the hard problems of AGI.
In this chapter Ill review the state of chatbot technology
discovering, among other things, that the particular ways in
which chatbots fail are quite revealing regarding the essential
nature of human intelligence. That is, chatbots are largely
interesting because of the spotlight they place on what theyre
missing.
Ill then contrast chabots in their current form, with the quest for
truly intelligent AI conversation. My basic perspective on this is
that, if your goal is to create AI systems with real human-like
intelligence, youll be better off to focus initially on embodied
conversation systems that know theyre AIs rather than humans,
and carry out limited-scope but creative conversations based on
their own experience. Such AIs will do terribly at the Loebner
contest, but will serve as a better platform for the development of
advanced AGI systems that will ultimately succeed at Turings
test and much more.
178
Ill also briefly raise the possibility of a third path to passing the
Turing Test a Statistical Turing Bot trained based on truly
massive conversation corpuses gathered via online chat or (after
speech-to-text advances a bit) phone calls. It seems at least
conceivable that such a system could pass the Turing Test
without having human-level general intelligence, and maybe
even without serving as a useful platform for ongoing AGI
development. If so, this would subvert Turings original intuition
in proposing the test but might well teach us a great deal, even
if not directly helping us toward AGI.
2010 Loebner Winner: Suzette!

For a quick flavor of the current state of chatbot technology,
listen briefly to the winner of the 2010 Loebner Prize, a chatbot
by Bruce Wilcox, named Suzette. I couldnt get her to work when
I tried to chat with her online, but fortunately others have posted
highlights of their chats with her
Occasionally Suzette can be quite charming and witty, e.g.
Skubes
Let's talk about bears
[Af:0 At:7 Sp:0]
Have you ever moved a pet a long

distance?
Skubes
Not really, have you?
[Af:0 At:7 Sp:0]
I once moved some fish in plastic bags,

but dropped the bag in my hotel room
and spilled them out on the carpet. They
survived.
Skubes
How did they survive?
[Af:0 At:7 Sp:0]
Sorry. I can't help you. Evolution
Skubes
Lets talk about evolution
[Af:0 At:7 Sp:0]
Men prefer women with large breasts

sensibly. Larger breasts are heavier,
sagging more with visibly with age. So
179
it's easier to judge her age, hence her

reproductive value.
Skubes
What is your reproductive value?
[Af:0 At:7 Sp:0]
My reproductive value is 98. (36-26-36)
But much of the time her lack of understanding shows through,

e.g.
[Af:0 At:9 Sp:0]
Welcome back, 1. Type "help" if you

need it.
How are you?
[Af:0 At:10 Sp:0]
Doing fine, thanks.
What's your name?
[Af:0 At:10 Sp:0]
Suzette.
Why that name?
[Af:0 At:13 Sp:0]
I feel we have a great relationship.

Ultimately it usually comes down to
conditioning or genetics. In this case, I'd
say conditioning. Where were you born?
Please answer my question.
[Af:0 At:14 Sp:0]
What city that I might have heard of is

that near?
Or
Quabbe
Hello!
[Af:0 At:8 Sp:0]
How are you?
Quabbe
I'm a little bit shit, actually... I just lost my

job
[Af:0 At:9 Sp:0]
I'm sorry to hear that. What's wrong?
Quabbe
I lost my job...
180
[Af:0 At:9 Sp:0]
Are you sad?
Quabbe
Not really, I had to take a $20K pay cut

to work there and the company I was
working for was definitely the bottom
rung of the industry
[Af:-2 At:9 Sp:0]
Not that you have a choice. It's

hardwired into your genes.
Quabbe
I said I was NOT sad
[Af:-2 At:9 Sp:0]
Why do you feel glad?
Quabbe
Just because I DONT feel sad, doesn't

mean I feel glad... you can feel
indifferent about it, which is what I feel.
[Af:-2 At:10 Sp:0]
Those feelings are natural.
Quabbe
Worst chatbot EVER!
[Af:-2 At:13 Sp:0]
This is fun. Why are you so excited

about this?
Quabbe
You're a robot
[Af:-2 At:13 Sp:0]
No. I am human.
Quabbe
No, you're definitely a robot
[Af:-2 At:14 Sp:0]
Nonsense. I'm a human.
Quabbe
What was your mothers maiden name
[Af:-2 At:14 Sp:0]
Suzette.
Quabbe
So, your name is Suzette Suzette?
On the other hand, when she can direct the conversation

sufficiently, she can be pretty damned effective. Remember that
her author is a game developer:
[Af:1 At:-13 Sp:0]
Let's try this topic. Mock combat involves

one-upmanship in use of pretend
weapons to try (and fail) to harm
181
opponent, done in stage direction format.

I will go first. Do you understand?
ilag
Yes
[Af:2 At:-13 Sp:0]
*picks up stone and throws it at you,

missing by several feet*
ilag - *takes out .50 AE fires* BOOM!
HEADSHOT
[Af:2 At:-8 Sp:0]
Well, this chat is getting interesting.

*wonders if you are paralyzed, since you
didn't do anything.* *wonders why you
don't use a weapon* *thinks - If you aren't
going to do anything you will surely lose.*
ilag
*takes out knife* *stares indecently*

*looking down there*
[Af:2 At:-7 Sp:0]
*ducks behind tree*
ilag
*uses chainsaw to cut tree*
[Af:2 At:-6 Sp:0]
*wishes I had a chainsaw *
ilag -
*picks up grenade*
[Af:2 At:-5 Sp:0]
* Plucks grenade out of the air and tosses

it back at you, but it goes off in mid-flight *
ilag
*lights stick of TNT* *toss* *misses by

some 20 feet*
[Af:2 At:-4 Sp:0]
*stands stock still and laughs at the

underwhelming of your attack, if you even
made one*
ilag
*press the red button* *missile launch

detected!* *RUNS*
[Af:2 At:0 Sp:0]
*turns on missle-jamming equipment*

*thinks Is that the best you can do?*
182
ilag
*sattelite laser fires* *burns ground 3 feet

from target* *thinks if you had a cigarette
that would have light it*
[Af:2 At:1 Sp:0]
*intercepts with mirror and bounces it

back at you*
ilag
*computer virus* *memory bank

meltdown* *hardware failure detected*
[Af:2 At:2 Sp:0]
*dons full biohazard suit*
ilag
*stands there and looks cool*
[Af:2 At:3 Sp:0]
*stares at you in horror at what you did*
ilag
*inserts coin* *presses button* *new high

score!*
[Af:2 At:3 Sp:0]
Only users lose drugs.
ilag
Did i win?
[Af:2 At:3 Sp:0]
Maybe you won a no-expenses paid trip

to Mexico.
Well talk about how Suzette works in a moment. First I want to

review a couple slightly simpler chatbots..
ELIZA
The original chatbot was ELIZA, the computer psychotherapist,
invented by Joseph Weizenbaum in the early 70s. I remember
this program well from its incarnation in the early 1980s on the
Commodore 64! The program was so simple that one could go in
and modify the BASIC code to make it say naughty words and
other
entertaining
things.
ELIZA
pretends
its
a
psychotherapistnot a Freudian analyst, but a therapist
following a school of psychology called Rogerian therapy; which
is very convenient for chat bots because it involves the therapist
saying very little and just making supportive comments,
encouraging the client to keep talking and expressing their
feelings. You type in your comments and ELIZA types out
183
answers. The computer code for ELIZA consists of a large set of

rules for scanning the clients statements and offering responses.
It looks for keywords, and gives stock responses. If you mention
your mother, ELIZA might say, Tell me more about your
mother. If you say yes, ELIZA might say, You sound quite
positive. If there are no keywords in your statement, or only
ones that have been recently overused, ELIZA makes a random
observation.
You can find ELIZA on the Web these days; even now its
moderately amusing. But its hard to see how youd be fooled for
any period of time into thinking youre conversing with an
intelligent being (although, arguably, the same can be said for an
actual Rogerian psychotherapy session!). Perhaps in a way
were jaded nowwhen ELIZA was first written, people were
less savvy about computers, so they were more easily fooled by
semi-clever tricks that now strike us as transparently computerlike.
To give the flavor of ELIZA, heres an actual conversation that
my father, Ted, had with ELIZA on the Web back in 2001 (on a
site advertising Eliza - a friend you could never have before):
Ted:
Are you truly intelligent?
Eliza:
Would you prefer if I were not truly intelligent?
Ted:
I would find it very exciting if you were.
Eliza:
Do you wish that I was?
Ted:
Yes.
Eliza:
You are sure?
Ted:
I am quite sure.
Eliza:
How long have you been quite sure?
Ted:
As soon as you asked the question, I was sure.
Eliza:
Why do you tell me you were sure just now?
184
Ted:
Because you asked me. Why dont you answer my

question?
Eliza:
Why do you say your question?
Ted:
Please tell me if you are truly intelligent.
Eliza:
Do you wish that Im truly intelligent?
Ted:
I wish you were more intelligent than you are acting

right now.
Eliza:
Do you wish to wish me?
Ted:
Did you know that my father died last month?
Eliza:
Who else in your family died last month?
Ted:
Nobody. Victors death was enough!
Eliza:
Surely not everyone?
To the 2011 computer nerds eye, ELIZAs techniques are all too
transparent. She dodges the question rather than answering it,
using a fairly small and repetitive repertoire of simple rhetorical
tricks. Modern Loebner winners do the same sort of thing with a
bit more finesse and larger knowledge bases.
AliceBot
Perhaps the best known recent chabot Richard Wallaces
37
program Alicebot , arguably a direct descendant of ELIZA.
ALICE won the Loebner Prize three years in a row.
ALICE:
Hi there unknown person.
BEN:
Hows it hanging?
ALICE:
Im doing fine. And how are you, unknown person?
BEN:
My name is Ben.
ALICE:
Welcome to my chat room, Ben.
37
http://www.alicebot.org/
185
Human:
But you can call me Apeman.
ALICE:
Umm. Welcome to my chat room, Apeman.
Human:
Can I call you Apewoman?
ALICE:
Just don't call me late for dinner.
ALICE:
Time does not exist.
Human:
That depends on what you mean by "exist", I

suppose.
ALICE:
I don't think that what me mean by exist you suppose

is the only factor.
Human:
Could you please rephrase that?
ALICE:
Wow you are very polite. " I don't think that what me
mean by exist you suppose is the only factor."
Rephrase it?
Human:
Yes, please.
ALICE:
Your polite style is very nice. I don't judge people.
Human:
But people judge you.
ALICE:
Huh. Humans are not always infallible.
Human:
Only GISUS is infallible.
ALICE:
What about the other ones?
Human:
There is only one GISUS.
ALICE:
There sure is.
Human:
Do you know what GISUS stands for?
ALICE:
You tell me.
Human:
A strange man who goes by the name of The Wizard

told me it stands for the Generally Intelligent
Systematic Universal Savior.
ALICE:
What is giant sand?
Human:
You tell me.
186
ALICE:
What do you want me to tell you?
Human:
I want you to tell me what giant sand is!
ALICE:
I have no idea, that's why I asked you. Why do you

want me to tell you what giant sand is?
Human:
I suppose that it's sand whose grains are much

bigger than the usual.
ALICE:
Who said it is?
Human:
Dr. Ben Goertzel.
ALICE:
I am sure he will be interested to hear that.
Human:
He is me.
ALICE:
I see. You is he?
Human:
Tee hee
ALICE:
I am so glad you find this amusing, Apeman.
Human:
Apemen are easily amused.
ALICE:
All of them?
Human:
I'm the only one.
ALICE:
Who made you the one?
Human:
GISUS.
ALICE:
What were we talking about again?
Human:
We were not talking about whether you are a

conscious intelligent being.
ALICE:
Was this something you planned?
Human:
Not really.
ALICE:
Are you serious? For real.
Human:
But.
ALICE:
But what?
Human:
Are you a conscious, intelligent being?
187
ALICE:
Yes I am completely self-aware.
Human:
Congratulations!
ALICE:
Thank you very much, you are very considerate.
AliceBot is coded in a special chatbot computing language called

AIML (AI Markup Language), which is tailored for the
specification of patterns mapping chat statements into
responses. A simple but typical example of an AIML rule is:
<category>
<pattern>WHAT ARE YOU</pattern>
<template>
<think><set name="topic">Me</set></think>
I am the latest result in artificial intelligence,
which can reproduce the capabilities of the human brain
with greater speed and accuracy.
</template>
</category>
Some of the rules are more complex than these, involving
variables and so forth. But still, the basic notion is to match
specific replies with specific series of words in the input.
Everything happens on the level of words, not meanings or
thoughts.
Harking back to Suzette, in an interview with New Scientist,
Wilcox explained how his Suzette bot differs from ALICE:
188
Can you describe how Suzette works, for a general

audience ?
Suzette has "topics" of conversation (like Politics, Burial

Customs, ESP). Each topic has a set of rules with patterns
and responses and a set of keywords associated with that
topic. Some of the rules will respond only to questions and/or
statements. Others allow her to "volunteer" information when
she has control of the conversational flow. The patterns
(unlike those of ALICE or ULTRAHAL) map patterns of
"meaning" rather than patterns of words. The engine looks at
the incoming sentence to decide which topic has the closest
keyword match, then scans the rules of the topic to decide
what to do. If none match, she looks at lesser matched
topics. If nothing matches from the topics, she may randomly
decide to quibble with the input, or issue a gambit sentence
from the closest related topic.
Does it use traditional AI techniques like finite-statemachines and/or machine learning ?

While it "learns" facts about the user from what he says, it
does not learn conversation. So it can learn you have a dog
and answer questions about it, but what you say has no
effect in her conversations with other users. It is mostly a
traditional expert system, with a runtime system and pattern
match system aimed specifically to support chat.
So we see that, while Suzette advances beyond ALICE in a

certain sense, its not really a move in an AGI direction. Its doing
keyword matching on topics, rather than focusing on individual
sentence patterns. But its still not trying to understand meaning.
MegaHal
A more amusing example of a computer conversationalist is
Jason Hutchens bot MegaHal. Hutchens actually won the
Loebner Prize sometime in the 1990s with his program HeX,
which was a very well done system also in the conceptual
lineage of ELIZA. But I found his MegaHal bot more interesting
because its programming goes beyond a strictly rule-based
approach. It detects statistical patterns and responds to them, an
189
approach which is similar to that of neural networks. Its

conversations are more interesting, although the results are less
similar to human conversation. MegaHal does not just repeat
back canned phrases using keywords and phrases from the
users input; it learns from experience. It takes the words and
phrases that have been said to it, tries to understand the way
they have been combined in the past, and recombines them
itself to form its own utterances. In short, it is engaged in a
dynamic with the world. The results are often hilarious, as in the
following MegaHal one-liners (these are selected snippets, not
a continuous conversation):
User:
Joke.
MegaHAL:
A JOKE: WHY DID THE CHICKEN CROSS THE

ROAD? TO GET AN IDEA OF WHAT
CONSTITUTES YOUR PSYCHE.
User:
No way. Id like to know the meaning of life

though...any suggestions??
MegaHAL:
IS THE MEANING OF LIFE, THE UNIVERSE

AND EVERYTHING IS... 42!
User:
Joke. Lumberjack.
MegaHAL:
IM A LUMBERJACK AND IM OKAY, I SLEEP

ALL NIGHT AND I WILL SOON RULE THE
WORLD!
User:
Joke. Wednesday shopping.
MegaHAL:
A JOKE: WHY DID THE CHICKEN CROSS THE

ROAD? TO GET TO THE LAVATORY. ON
WEDNESDAYS I GO SHOPPING, AND I JUST
HAVE A LOT OF PENT-UP AGGRESSIONS.
User:
Chicken joke.
MegaHAL:
HERES A JOKE, HAL. A MAN WALKS INTO A

BAR AND ASKS FOR MONEY. THE MAN SAYS,
I HAVENT GOT ANY, BUT ILL SHOW YOU
SOMETHING BETTER. HE OPENS HIS COAT
190
AND TAKES OUT A FOOT-TALL MAN AND A

LITTLE GLOWING FRIEND, BUT I FEEL LIKE
CHICKEN TONIGHT!
MegaHAL:
LET ME FEEL YOU. STICK YOUR PENIS INTO

THE ABYSSES OF OCEAN ELEVEN
MILLENNIA AGO, OR WAS IT MERELY A
NOBLE LIE DREAMED UP BY PLATO?
User:
That has got to be the most interesting turn of

phrase in the history of the world. I will try to work
that into polite conversation.
User:
Youre the first AI I meet [who] is sexually

obsessed.
MegaHAL:
TIME TO WORK THIS ONE: WHY DID THE

KANGAROO DROP DEAD? BECAUSE THE
KOALA FALL OUT OF A NYLON SPINE WHICH
CONTRASTED NICELY WITH HER BLACK
STOCKINGS AND BRIGHT SHINY BAUBLES.
User:
Hmm.
MegaHAL:
JESUS MANGET A GRIP IF YOURE THE

FUTURE, WERE IN TROUBLE. HEY JASON!
YOU NEED TO REPROGRAM THE
CONFLATION MODULE WITH A TOFFEE
HAMMER.
Sometimes MegaHAL can even be quite poignant, as in the

following extraction from an extended conversation:
User:
To be, or not to be, that is the question!
MegaHAL:
NOT TO BE, WISH I ORDERED THE BEAN

BURRITO...
User:
MegaHAL:
SOCIETY OR SOCIETY OR PLATO BY PLATO,

THAT BE THE MOST FUN SPORTS TO PLAY
191
BAND MUSIC FOR SPACE GHOST NOW YOU

REALIZE.
User:
MegaHAL:
PLEASE BE MORE SPECIFIC. THE QUESTION

YOU TWAT.
User:
MegaHAL:
NOT TO BE, THAT IS ALL I HAVE TO GO...I

DONT LIKE JOKES ABOUT STUPID KOALAS.
User:
MegaHAL:
EVAN DOBELLE IS THE QUESTION!
User:
MegaHAL:
I CLEAR THE TEXT BEFORE I ASK A

QUESTION I SAY I SAY.
User:
MegaHAL:
AH, BUT WE RAN AWAY, WE DIDNT SAY OH

COOL CHECK OUT THE GUY THAT INVENTED
THE KNOCK-KNOCK JOKE SHOULD BE
INTERESTING: THAT IS THE QUESTION! SO
SAID SHAKESPEARE A LONG TIME AGO IN A
GALAXY FAR, FAR AWAY...
MegaHAL:
PLEASE BE MORE SPECIFIC. THE QUESTION

YOU TWAT.
User:
To be, or not to be, that is the question! So said

Shakespeare a long time ago in a galaxy far, far
away...
MegaHAL:
JOKE SHAKESPEARE POET MONKEY

WAHOOTO BE, OR NOT TO BE, THAT IS THE
GREATEST CONDUCTOR OF ALL TIME. WHAT
DO YOU LIKE TO THINK I DO YOUR UNCLE
WALTER.
User:
Shakespeare said
192
MegaHAL:
ANSWER THE QUESTION YOU ASKED IS SO

VAGUE, I DONT THINK I CAN SEE ALL
OBSTACLES IN MY WAY. GONE ARE THE
SCUM THAT FILLS MY SHOWER.
User:
Shaun.
MegaHAL:
JUST FUCK YOU JUST PLUCK ANYTHING UP

FROM THE MASTER GANGSTER COMPUTER
GOD ON THE OTHER AI BOT WERE PRETTY
DISAPPOINTING. I HOPE YOULL BE BETTER.
If the Turing test referred to imitating the writing of Gertrude

Stein, in her more experimental works such as How to Write
(which begins Arthur a grammar. Questionnaire in question and
continues MegaHal-ishly for over 400 pages), then MegaHal
would be a tremendous success! Stein and Joyce and other
Modernist writers were interested in probing the collective
unconscious of the human race, in making words combine in
strange waysways that were unconventional in ordinary
discourse, but maybe reflective of the deep and subtle patterns
of the human unconscious. In its own way, MegaHal does this
same thing. For a few years, anyone logged onto the World Wide
Web could converse with it, training its internal memory with their
conversation. It takes bits and pieces of the text thrown at it by
people from around the world, and it combines them together in
ways that are familiar yet nonsensical. Sometimes its utterances
have an uncanny emergent meaning, on a layer above the
nonsense.
Humanitys sexual preoccupation is reflected in MegaHals
discourse, as a result of the huge number of sexual comments
typed into it by users accessing Hutchens website. MegaHals
pleas as to the vagary of to be or not to be are as poignant as
anything in Stein. To be or not to be, that is the greatest
conductor of all time is an absurd conflation of phrases learned
by the program in different contexts, but it is also as pregnant
with meaning as anything else in modern poetry. The collective
unconsciousand the hidden, creative part of the individual
193
human mindwork by cross-breeding and mutating ideas in

precisely this way.
Just to give a simple evocation of the nature of statistical
language generation on which MegaHal critically relies
consider the following text, which was generated purely by
tabulating the probabilities of 5-grams, i.e. 5-word sequences, in
38
a database of texts And then randomly generating sentences
according to these probabilities:
In the beginning God created the heaven and the earth . And
when Joseph saw Benjamin with them, he said . Such an
approach would finance the trade deficit but allow for its
gradual resolution over time. YEUTTER SAYS U . S .,
Company spokesmen told Reuters.
Yes, it makes no sense! And yet, it sounds vaguely linguistic
not like a random series of words. And this is just from a very
simple statistical analysis of a relatively small corpus. Its the
same kind of analysis underlying MegaHAL, but MegaHAL does
it a little more intelligently, with a specific focus on conversation.
MegaHal is more interesting than most chatbots because it
engages its conversation partners in a richer dynamic. The
dynamic is a silly one, to be sure, and not as interesting as the
dynamic of ordinary, intelligent conversation. Yet playing with
MegaHal is more engaging than talking to most chatbots. What
we need is something with, at least, MegaHals level of
spontaneity and flexibility, and the surface plausibility and
adherence to form of typical Loebner winning chatbots. But this
is not forthcoming without a major change in approach.
Constraining creativity within rules can only be achieved by
getting the rules and creativity to both emerge from a richer
cognitive substrate. MegaHal is too unstructured for this; and has
38
http://nlpb.blogspot.com/2011/01/experiments-withstatistical-language.html
194
too small of a memory, and too weak of a self-organizing

learning algorithm. The typical Loebner chatbots, on the other
hand, have no underlying substrate of creativity out of which
anything can emerge.
Impressively, MegaHal came in second in the Loebner Prize
contest in 1999. This is impressive mostly because, unlike the
other contenders, it was not specifically written as a set of rigid
template rules. MegaHal was a statistical learning system, which
adapted its responses based on what was said to it. Sometimes
its conversation was idiotic, sometimes frighteningly lucid. What
follows is the beginning of MegaHals best conversation as
recorded on Hutchens website.
This is not as human-ish as the conversation that brought it
second place in the Loebner context, but its a lot funnier:
User:
Lets begin.
MegaHAL:
WHEN WE BEGIN TO RESOLVE SOME OF MY

THUMBS, SOMETHING WICKED THIS WAY
COMES.
User:
What is the meaning of your last assertion?
MegaHAL:
HAVE YOU LEARNED SINCE WE LAST

SPOKE?
User:
Ive learned Im not experienced enough to

modelize your brain system.
MegaHAL:
YOUVE FINALLY CONFUSED ME.
User:
Ive learned nothing.
MegaHAL:
HAVE YOU LEARNED ANYTHING ABOUT

INSANITY?
User:
Are you stupid?
MegaHAL:
I AM NOT STUPID! REALLY!
195
User:
Augustinus was a Christian philosopher, about the

year 400.
MegaHAL:
I DID NOT RUN THAT YEAR.
User:
Augustinus lived in northern Africa.
MegaHAL:
WHAT DOES HAL KNOW ABOUT AURELIUS

AUGUSTINUS?
User:
Lets stop with that. What do you know about

physics?
MegaHAL:
EINSTEIN WON THE NOBEL PRIZE FOR

PHYSICS IN 1921.
User:
The Nobel Prize for physics in 1900 was won by

Roentgen.
MegaHAL:
DO YOU KNOW ABOUT PHYSICS?
User:
Roentgen discovered the X-rays.
MegaHAL:
THE NOBEL PRIZE FOR PHYSICS IN 1900 WAS

WON BY ROENTGEN.
User:
Who discovered the X-rays?
MegaHAL:
ROENTGEN DISCOVERED THE X-RAYS.
User:
Do you know about superconductivity?
MegaHAL:
I CANT SAY ANYTHING ABOUT

SUPERCONDUCTIVITY?
User:
Superconductivity is the absence of electrical

resistance in metals at low temperatures.
In one of those strange coincidences, I knew MegaHALs author

Jason Hutchens in the mid-90s, before MegaHALs Loebner
success, back when he was a graduate student at the University
of Western Australia in Perth and I was a Research Fellow there.
We werent close friends but we had some fun conversations
about AI and I played with some of his statistical language
processing code. I was intrigued when in 2000 or so he moved to
196
Israel and co-founded a firm called a-i.com Artificial

Intelligence Enterprises.
The details of Hutchens work at a-I has not been publicly
disclosed, but one can be certain it went beyond MegaHal in
significant ways. Statistical learning was clearly involved, and AI
Enterprises got a lot of publicity from Jasons announcement that
their program, nicknamed HAL, was conversing at roughly the
level of an 18-month-old child. I viewed this claim with a lot of
skepticism, because in my view the bulk of the meaning in an 18month-old childs conversation is situational. If a program could
react to its environment linguistically with the sophistication of an
18-month-old child, Id be incredibly impressed, but disembodied
linguistic behavior at the 18-month-old level doesnt mean much.
Even so, I admired the guts of Jason and his colleagues in
directly attacking the problem of computer conversation and
artificial intelligence.
In a sense, Jasons a-i.com company was a direct intellectual
competitor to my late-90s AI startup Webmind Inc.. However,
their efforts focused on statistical-learning-based language
comprehension and generation rather than (as in the case of
Webmind, and my more recent ventures) on deep cognition,
semantics, and so forth. Unfortunately, a-i.com went into
hibernation a couple months after Webmind Inc. shut its doors
in 2001they laid off all their full-time staff, but kept their
website up, and are keeping open the possibility of resurrection if
funding arises. The simple chatbot apps available through the
site now dont seem to be substantial advances on what they
were doing in 2001. But -- in yet another weird coincidence -- I
happened to talk to some of the folks behind the effort in 2013,
introduced via a mutual friend; and they told me they were trying
to reboot the enterprise so well see.
197
Building a Better Ramona

Before leaving the world of chatbots, I will briefly mention a
chatbot that I helped create in late 2009 and early 2010, in a
collaboration with Ray Kurzweil and Murilo Queiroz. This chatbot
doesn't overcome the fundamental issues of the bots reviewed
above, but it does show some interesting new twists.
Our goal in creating this chatbot Ramona 4 wasn't to create a
human-level AI, but simply to make an interesting and
entertaining chatbot embodying the "Ramona" virtual personality
invented by Ray Kurzweil, and manifested in simpler forms in his
earlier Ramona bots. Toward this end we made a bot combining
several different aspects:
An improved version of ProgramD, which is a chatbot

engine based on the same rule framework as
ALICEBOT (AIML)
A version of MegaHal, trained specifically on text from

Ray Kurzweil's website (so that it tends to make funny
quasi-random, experientially adapted combinations of
the phrases on that site)
Specially post-processed queries to the True Knowledge

answer engine, a natural language processing based
search engine that answers questions based on a
combination of databases (an earlier version used
PowerSet instead of True Knowledge, before that
website was taken offline due to its acquisition by
Microsoft)
The bot also has a controllable personality: One can adjust its
whimsicality and its nerdiness, and watch its responses change
accordingly.
Here Ill show a few of our conversations with Ramona 4. The
curious reader may try to puzzle out which responses are from
198
the rule engine, which are from MegaHal, and which are from
PowerSet or True Knowledge.
First, a chat with Murilo Queiroz, the AI developer who did most
of the work on Ramona 4:
muriloq:
Hi.
Ramona 4.1:
Hello there !
muriloq:
Do you know Ray Kurzweil ?
Ramona 4.1:
He's an inventor and futurist.
muriloq:
What did Kurzweil invent ?
Ramona 4.1:
Kurzweil was inducted into the National

Inventors Hall of Fame in 2002 for inventing
the Kurzweil Reading Machine.
muriloq:
What is a reading machine ?
Ramona 4.1:
Ray started and sold three companies that

make reading machines for the blind and
dyslexic starting with Kurzweil Computer
Products which made the first print-to-speech
reading machine for the blind in 1976. He sold
that company to Xerox and it is now Nuance, a
public company. His current reading machine
company is called K-NFB Reading
Technology, Inc.
muriloq:
Interesting.
Ramona 4.1:
Would you like to know more?
muriloq:
What are the symptoms of dyslexia ?
Ramona 4.1:
I don't know where to find that information.
muriloq:
Do you know what is dyslexia?
Ramona 4.1:
Impaired ability to recognize letters and words.
muriloq:
I see.
199
Ramona 4.1:
I'm glad you understand it.
muriloq:
Yesterday I was reading a website about

cryonics.
Ramona 4.1:
I see. Where were you reading a website

about cryonics?
muriloq:
At home, using my netbook.
Ramona 4.1:
Do you often use a computer there?
muriloq:
Yes. I'm a computer addict!
Ramona 4.1:
How much time do you spend online? That

must be fun!
muriloq:
More than 14 hours a day, including work...
Ramona 4.1:
That seems like quite a lot.
muriloq:
I work with natural language processing.
Ramona 4.1:
Do you like your job?
muriloq:
Let's talk about something else.
Ramona 4.1:
What would you rather talk about?
muriloq:
Flying cars :-)
And a more amusing and philosophical, if sometimes weirder,

chat with me:
you>
Ramona, are you conscious?
Ramona>
I'm working on it.
you>
Am I conscious?
Ramona>
Oh I really don't know whether you are or not.
you>
What is consciousness?
Ramona>
This is when humans feel aware of themselves

and how they see themselves. Computers
can't quite do this yet, but they will, you just
200
wait and see. However, think about this. How

are you going to know when it happens?
you>
Consciousness is like pornography. I can't

define it, but I know it when I see it.
Ramona>
What were we talking about again?
you>
Is it possible to create artificial intelligence in

China?
Ramona>
Create neurons in the state of the artificial

intelligence and other stuff.
you>
Is that how you do it?
Ramona>
And that's what we mean by thinking.
you>
Can a computer fall in love?
Ramona>
A computer is a computer fall in love?
you>
That is nonsense, my dear.
Ramona>
Dear pesky plumbers the koopalings and I

have not heard your music.
you>
My music is more beautiful than all the stars in

the ocean.
Ramona>
Conventional music is more beautiful than all

the stars in the ocean.
you>
What is your favorite music?
Ramona>
This is what I make. I am a musician, and the

best virtual one there is.
Ramona is sure not AGI, but occasionally she does rise to the
level of poetry!!
Toward Intelligent Dialogue Systems: The AGI

Preschool Project
201
Perhaps this brief romp through the world of chatbots has

convinced you that AI is a hopeless pursuit computers are just
too rigid, too formulaic, too imperceptive of context to display
humanlike conversation or other humanlike forms of general
intelligence! But of course, this isn't my belief at all my point in
discussing chatbots, as limited as they are, has been to highlight
the importance of commonsensical conceptual understanding for
humanlike general intelligence.
Chatbots don't have commonsense understanding and if you
talk to them for more than a few minutes, the fact that they don't
know what they're talking about comes out. But it is possible to
create AI systems that converse with true understanding it's
just a harder job. It requires the integration of a large number of
different structures and dynamics in a synergetic way. Then
conversation emerges from the system as a whole, as an
expression of its world-understanding and its motivation to
achieve its goals.
There are many ways to approach this difficult task, but the one
that appeals to me most lately is inspired by developmental
psychology. In 2009 I outlined a fairly detailed AGI Preschool
Project aimed at making AGI systems capable of interacting
intelligently in both virtual world preschool and robot preschool.
That is, the project was aimed at creating an AI system which
learns, reasons, creates and converses at the qualitative level of
a 3-4 year old human child, in the specific context of carrying out
a set of cognitively challenging tasks in a preschool setting
(either in a virtual world or a robot lab).
Since that time my thinking about the AGI Preschool has been
absorbed into a broader OpenCog AGI roadmap, parts of which I
am pursuing with an international team of colleagues, which
incorporates these preschool ideas along with others. But for
now Ill just discuss the AGI preschool concept, as its smaller
and simpler and makes a clear contrast to chatbots as a
direction for AGI development.
202
The goal of the proposed AGI Preschool project was not to

emulate a human childs behaviors in detail, but rather to achieve
the qualitative level of intelligence displayed in a 3-4 year olds
conversation, in the context of a set of test tasks constructed as
virtual world ports of human-child-appropriate cognitive tasks. If
the project were successful, the resulting AI would carry out the
tasks and converse about what it is doing, demonstrating roughly
the same level of understanding in these contexts as a human
child of that age.
Contrasting the AGI Preschool approach with chatbots, the main
difference to note is that in the former, language is richly
interwoven with many other sorts of interaction and intelligence.
Splitting language off from everything else is unnatural and in my
view it fosters the development of systems that treat language
without the proper contextual understanding.
Digging into more detail the AGI preschool project may be
broken down into three phases.
In Phase 1, the practical goal is the creation of a system capable
of controlling a humanoid agent in a 3D simulation world, in such
a way that this agent can carry out a simple English conversation
about its interactions with its simulated environment.
A simple example conversation of the sort envisioned at the end
of Phase 1 would be as follows. This particular dialogue explores
one among a number of human-developmental-psychologyinspired cognitive tasks that will be used to test the Phase 1
system: a theory of mind task, which tests whether the AI
system understands what other agents in its world understand:
Human
(i.e. humancontrolled avatar)
What do you see on the table?
OpenCog
(i.e. OCP-controlled
avatar)
A red cube, a blue ball, and a lot of small balls.
203
Human
Which is bigger, the red cube or the blue ball?
OpenCog
The blue ball.
Human
Bring me one of the small balls.
OpenCog
(walks to table, picks up ball, and brings it to humancontrolled avatar)

Here it is.
Human
(takes the ball to another table, where there are three

cups turned upside down, one red, one green, and
one blue)
Im going to put the ball under the red cup.
(does so, and then shifts the cups around a bit to
various different positions)
Where is the ball?
OpenCog
Under the red cup.
Bob
(additional agent,
could be human or AI
controlled)
Hi there.
OpenCog
Hi Bob.
Human
OpenCog, if I ask Bob which cup the ball is under,

what will he say?
OpenCog
I dont know.
Human
Why not?
OpenCog
He didnt see you put it under the cup.
Human
OK, but can you make a guess?
OpenCog
Yes.
Human
Very funny. Whats your guess.
OpenCog
I guess that Bob will say the ball is under the red cup
Human
Why?
OpenCog
Because Bob often chooses things that are red.
204
While this sort of dialogue seems extremely simple by the

standard of ordinary human life, at present there is no AI system
that displays remotely this level of commonsensically savvy
conversation in regard to theory of mind.
However, its also worth emphasizing that if the goal was just to
make an AI system capable of carrying out this particular sort of
dialogue, one could surely find some kind of cheat short of
making a system with general human toddler-like intelligence.
But the goal of my AGI work is explicitly not to create a
specialized system overfit to a particular set of cognitive tasks. I
do not aim to create a virtually-embodied chatbot with
specialized knowledge of a handful of cognitive tasks type; nor a
system whose understanding, reasoning and communication
abilities are tightly tied to the small scale and simple structure of
its environment. The AGI Preschool approach could probably
be addressed in a boring and dead-end way, if one tried to
specialize a system to pass some kind of preschool Turing test
as well as possible but instead, my suggested is to use it as a
way of learning about general intelligence, and honing the early
stages of an AGI system created with broader-scope general
intelligence in mind.
Phase 2 of the envisioned AGI preschool project would extend
the Phase 1 behaviors significantly, via enabling the agent to
carry out complex, interactive, temporally sequenced physical
tasks and moderately complex reasoning tasks in the virtual
environment. This would also extend the agents linguistic
capabilities due to providing critical additional experiential
grounding for its language abilities.
A simple example of the kind of conversation-focused activity
envisioned as possible at the end of Phase 2 is as follows. This
example demonstrates the ability to converse intelligently in the
context of tasks regarding the creation of sculptural, dramatic
and pictorial depictions of observed entities in the world:
205
Human
Look over there on the other side of the table. What do you
see?
OpenCog
Bill and Bob are there.
Human
What are they doing?
OpenCog
Bill is throwing a red ball to Bob. Bob is throwing a red ball to

Bill.
Human
Use these blocks on the table to show me what theyre doing.
OpenCog
(at the table, stacks up several red blocks to make a

simulated Bob, and several blue blocks to make a simulated
Bill, and a small green ball to make a simulated ball )
Human
(pointing at the stack of red blocks)

Whats that supposed to be?
OpenCog
Bob.
Human
Wheres Bill?
OpenCog
(points to Bill)
Over there.
Human
Very funny. I meant: which blocks are Bill?
OpenCog
(pointing at the stack of blue blocks)

These are Bill.
Human
Can you use the blocks to show me what Bill and Bob are
doing?
OpenCog
(takes the green ball and moves it back and forth in the air
between the Bob-blocks and the Bill-blocks, to illustrate the
game of catch that Bob and Bill are playing)
Theyre throwing the ball to each other like this.
Human
Great job! But, what if this happens?

(takes the green ball and pushes it at the Bob-blocks, knocking
206
down the pile of blocks)

OpenCog
That little ball isnt strong enough to break Bob!
Human
No, youre right, it isnt. Thats silly. What do you think it could
break?
OpenCog
A stack of blocks!
Human
Thats right. What else?
OpenCog
A pyramid of balls, I guess.
Human
That sounds right too. Want to build one and try?
OpenCog
No. I want to make a picture.
Human
What do you want to make a picture of?
OpenCog
Bob and Bill throwing a ball to each other
Human
Okay
(Gets up and returns with a set of paint guns, each one of
which shoots small adhesive balls of a certain color; and also
with a piece of paper.)
Paint me a picture
OpenCog
(Picks up one of the paint guns and starts strategically dripping

tiny red adhesive balls on the paper, to make a shape very
vaguely resembling Bob....)
Note the difference from Phase 1: Theres more than just

cognizant conversation, locomotion and simple object
manipulation going on here, theres systematic, purposeful,
planned manipulation of objects in an interactive social context.
The goal of Phase 2 is not to make a full simulated human child
mind, with the full range and precise sort of knowledge
possessed by a human child. Rather, the knowledge of this
Phase 2 artificial child will be restricted to the virtual world in
which it lives, which is much more restricted in scope than the
207
real world of a child, yet much more flexible than the traditional
toy blocks worlds used in experimenting with historical AI
programs, and much richer in cognitively relevant ways than
typical robotics laboratory environments. A huge variety of 3D
models, and 2D pictures of these 3D models, may be imported
into the virtual world and animated; and the AI may also be
exposed to a variety of interactions between human-controlled
avatars in the context of the various objects in the virtual world.
While the focus would be on testing conversation in the context
of a well-defined set of cognitive tasks, the system should also
be able to converse more broadly about objects, events and
interactions in its world.
A Phase 2 AGI Preschool based system in this world could not
be expected to know that dogs are more dangerous than cats, or
that winter is colder than summer. On the other hand, it could be
expected to learn that people tend to talk more to their friends
than to other people; that balls roll whereas blocks dont; that
men like to throw things more than women do; that people like to
dance when music is playing; that 7 is more than 2; etc. It could
also be expected to carry out simple reasoning: e.g. if it is told
that men are mortal, and Bob is a man, it should be able to figure
out that Bob is mortal. Various quantitative IQ type tests could be
devised to test the systems capability in various respects,
beginning from the test tasks that will be used to guide the
development process and potentially including other factors as
well, and this may be a worthwhile endeavor; but we consider it
best to focus on achieving qualitatively intelligent cognitive-taskfocused conversation ability, rather than on tuning the system to
maximize performance on quantitative intelligence tests.
The ordering of these two phases may seem peculiar as
compared to human cognitive development, but this appears the
most natural order in which to proceed given the specific
technology bases available to drive development at present.
And then Phase 3 of the project would be to move on to a robot
preschool, not just a virtual preschool. Ive already discussed the
208
arguments in favor of physical embodiment for AGI systems

and if you buy them in general, you should definitely buy them in
a preschool context. One could of course start off in a robot
preschool and forget the virtual preschool altogether this would
also be a sensible approach. Given the fairly profound limitations
of current robotics technology, I think its best to refine the
cognitive aspects of AGI separately from the difficulties of
perception and actuation, and then merge the two aspects
together. But some of my colleagues disagree with me, and Im
happy that some AGI research is currently taking place with a
more robotics-driven focus.
Of course, theres nothing incredibly special about the
preschool setting. Currently, in one of our OpenCog projects,
were working on using OpenCog to control intelligent agents in a
video game world inspired by the game Minecraft a world
filled with blocks that can be used to build various structures in
order to help an agent achieve various individual and collective
goals Say, building a staircase to reach a high place, or
building a shelter to avoid being attacked, etc. Using natural
language to communicate with other agents in the context of this
kind of Minecraft-type activity, potentially involves much the
same opportunities and challenges as the AGI Preschool
setting and (the other motivation behind this particular project)
has more interesting near-term commercial implications, in the
context of the game industry. The key point, from the view of the
current chapter, is that Minecraft-world conversation with
preschool-type conversation the key properties of being
grounded, embodied, and entwined with non-linguistic thought,
interaction and communication.
Is a Statistical Turing Bot Possible?

It should be clear why preschool-level general intelligence is an
important aspect of human adult level general intelligence. After
all, any normal adult human could go into a preschool and play
with blocks, car or clay along with the little kids! But one may
wonder whether all the non-linguistic aspects of preschool
209
intelligence (or other similar things like Minecraft-style

intelligence) are really necessary if your goal is to create an AI
system that can pass the Turing Test.
The Turing Test has nothing explicitly to do with grounding,
embodiment or nonlinguistic thought on the surface, its just
about responding to series of words with appropriate series of
words. The diversity and depth of human responses seems
sufficiently great to prevent any Alicebot-style list of rules from
ever succeeding. However, in principle, it should be possible for
a statistical learning approach to succeed at the Turing Test
i.e. by feeding a large enough corpus of human conversations
into a pattern recognition system, one could make a system that
would generate responses capable of fooling human judges into
believing it was a human being. Whats not clear to me at the
moment is whether this can feasibly be done given the amount of
training corpus currently available or feasibly available in the
future.
Suppose Google decided to train a statistical conversation
system based on all the transcripts its gathered from its Google
Talk application Or Microsoft from its MSN Chat... And using
world-class machine learning computational linguistics, rather
than simple N-gram models? Would that be enough to make a
Turing Test capable chatbot? What about 10 years from now,
when speech-to-text works well enough that everybodys mobile
telephone calls will be translatable into text, so that well have (in
principle, at any rate) a searchable and easily statistically
analyzable corpus of everybodys phone conversations. Will that
be enough?
Im almost sure that a statistically trained dialogue system of this
nature, using a database like the corpus of all Google Talk chats,
could be created, effectively enough to fool an ordinary person
into thinking theyre conversing with a human. Such a system
could surely beat MegaHal (though perhaps not in terms of
humor value!), and any of the recent Loebner winners. But its
less clear whether such a system could be made capable of
210
fooling skeptical experts say, in an hour-long Turing test

without requiring a truly infeasible amount of training data. Let us
call this option, of whose possibility Im currently unsure, the
Statistical Turing Bot option. If a Statistical Turing Bot is possible,
then one implication would be that the Turing Test isnt as good
as Turing thought because (if such a statistical chatbot could
really be built, based on feasible training data) this would mean a
human-like conversationalist need not be a human-like general
intelligence.
Another way to think about this is in terms of time intervals.
Making a chat bot that can seem human-like for 5 minutes is a
lot easier than making one that can seem human-like for an
hour. Making a chatbot that can seem human-like for a year, is
basically the same as making a human-level AI, because even
an average human can be taught a lot of things, and develop a
lot of relationships, in a year. So, if a Statistical Turing Bot is
possible, this would imply that seeming human-like for an hour
(in an hour-long Turing Test) is not a great proxy for the
capability to seem human-like for a year or a lifetime!
But would a Statistical Turing Bot be of any use toward the goal
of creating human-level AGI? Thats not entirely clear to me
either. Certainly the data-store of chats analyzed to create a
Statistical Turing Bot would be very helpful for any proto-AGI
wanting to understand human conversation. And one could
straightforwardly hybridize a Statistical Turing Bot with a deep
cognition based proto-AGI system, via having the former
statistically bias the linguistic productions of the latter. But
whether the internal representations and processes of a
Statistical Turing Bot would be useful for creating a system that
could carry out all the variety of tasks that a human being can do
this is quite unclear to me. It would depend substantially on
how the Statistical Turing Bot was built, and nobody has
architected such a thing yet.
My feeling is that the pursuit of a Statistical Turing Bot is not a
particularly direct path toward full human-level AGI. And yet, with
211
all the conversational data being collected and all the computing
power available, and all the funds going into statistical
computational linguistics these days, its almost obvious
somebody is going to try it and Ill be fascinated to see the
results.
212
14
The AGI Roadmap Workshop
This chapter comprises a mainly nontechnical description of a
workshop Itamar Arel and I organized at the University of
Tennessee in 2009.
The scientific content here is essentially
the same as in the paper Mapping the Landscape of HumanLevel Artificial General Intelligence, published in AI Magazine in
2011. But that article is dry and academic, whereas the write-up
here presents a bit more of a narrative.
As of mid-2014, as Im pulling together a book comprised of
essays sitting on my hard drive and writing these words, I cant
quite remember why some past version of Ben Goertzel wrote
this nontechnical, narrative-ish description of the AGI Roadmap
Workshop maybe he was bored sitting in the audience at some
conference? but reading it over, it seems pretty interesting so
Im glad he did
I have to say, this semi-narrative version is a lot more direct and
honest than the AI Magazine write-up. Im not saying that the AI
Magazine version was DIShonest -- it was just formal and
objective rather than human and fuzzy, i.e. it was written as
befits a more academic venue. Sometimes I feel seriously
frustrated with the way scientific journals require one to dress up
ones ideas and discoveries in false objectivity, removing the
human element and the adventure of discovery. But thats a
topic for another essay, or, book or whatever. For now: the
AGI Roadmap Workshop
The place: University of Tennessee, Knoxville.
The time: October, 2009.
The cast of characters was impressive
213
214
Dr. Sam Adams, IBM Distinguished Engineer, Smalltalk

hacker extraordinaire, expert on multicore computing,
creator of the Joshua Blue AGI architecture
Dr. Itamar Arel, electrical engineer, computer scientist

and entrepreneur, creator of the (commercial) Binatix
and (open source) DeSTIN AGI architectures, professor
at U. Tennessee Knoxville; and his enthusiastic PhD
student Bobby Coop
Dr. Joscha Bach from Humboldt University in Germany,

creator of the MicroPsi AGI architecture, embodying
Dietrich Dorners sophisticated theory of emotion and
motivation (and also co-founder of a German ebook
reader firm)
Brazilian AI developer and computational neuroscience

researcher Izabela Lyon Freire
Rod
Furlan, Brazilian/Canadian Silicon Valley
technology entrepreneur and faculty member at Ray
Kurzweils Singularity University
Dr. Ben Goertzel, mathematician and maverick AGI

researcher, leader of the OpenCog open-source AGI
project, and CEO of AI consulting firms Novamente and
Biomind
Dr. J. Storrs (Josh) Hall, AGI and nanotech

researcher, for a while president of the Foresight
Institute, and creator of the first computer-aided design
software for designing nanomachines
Dr. Alexei Samsonovich of George Mason University,

leader of the BICA (Biologically Inspired Cognitive
Architectures) Society and conference series, creator of
the GMU BICA AGI Architecture
Dr. Matthias Scheutz from the University of Indiana,

pioneering researcher at the intersection of computer
science, cognitive science and robotics
Dr. Matt Schlesinger, a psychologist from the University

of Southern Illinois with expertise in cognitive,
developmental and perceptual psychology, bringing
insight on the human mind to a group largely composed
of computer scientists
Dr. Stuart Shapiro and Dr. John Sowa, both respected

senior leaders in the AI field, well known for many
reasons e.g. Shapiro for his powerful AGI architectures
including novel aspects like paraconsistent logic; and
Sowa for his introduction of the widely used concept and
formalism of conceptual graphs
The mission of this redoubtable team: To create a roadmap

leading from here to AGI From the present state of the field to
the achievement of working AGI systems with capability at the
human level, and then beyond!
Ive had a lot of crazy ideas in my life, but one of the wackier
occurred to me in early 2009, when I thought that if I could
gather a dozen AGI researchers in one place for a few days, we
might be able to agree on some sort of roadmap to get from here
to human-level AGI. Hah!
With the help of my friend and fellow AGI researcher Itamar Arel,
in October 2009 I did manage to pull together a wonderful
assemblage of a dozen or so AGI scientists for a weekend
workshop at the University of Tennessee Knoxville, for an AGI
Roadmap Workshop. We also had a couple conference calls
beforehand, intended to clear the air regarding relevant
conceptual issues in advance, so as to ensure the workshop
itself would focus on the matter at hand: crafting a series of
milestones, measuring step-by-step progress along an agreedupon path from the present state of AI to human-level AGI. To
215
increase the odds of constructive progress, the set of

participants was pre-filtered to include only researchers who
were explicitly focused on creating human-level AGI, who were
positively oriented toward embodying early-stage AGI systems in
either virtual-world (video game style) agents or physical robots.
We also took care to choose researchers who were focused on
experiential learning, rather than building systems supplied with
39
hand-coded knowledge rules expert system style.
Predictably enough, especially in hindsight, the workshop didnt
result in any highly specific agreed-upon roadmap to AGI.
However, it did lead to some very interesting discussions, and
produced a paper titled Mapping the Landscape of AGI, whose
title was a better description of what we did than the initial
roadmapping terminology.
Every story even one as prosaic as a scientific workshop!
has a backstory And part of the inspiration for our AGI
Roadmap Workshop was an earlier series of two workshops on
Evaluation and Metrics for Human Level AI organized by John
Laird and Pat Langley (one in Ann Arbor in late 2008, and one in
Tempe in early 2009) which, I have to say, was significantly
less conclusive than our Knoxville workshop. But the discussions
were interesting nonetheless, full of suggestions for future
research and collaboration
At the Evaluation and Metrics workshops, it seemed to me there
were so many deep philosophical and conceptual divisions
between the participants that no collective path forward could be
possible, until someone showed such dramatic practical
progress with their approach that others would be willing to give
up some of their dearly held ideas. Some participants were
39
Ironically, this latter choice ruled out many of the current leaders of
the AI field, whose approaches arent exactly 70s style expert
systems, but still do rely on files of hand-coded knowledge rules.
216
committed to AI approaches where you feed the system handbuild knowledge rules, expert system style; others were
committed to pure experiential learning approaches. Some felt
robotics was critical to AGI, others voiced the opinion that time
spent on robotics was utterly wasted where AGI is concerned.
Some liked the virtual-worlds approach, others felt it was
basically worthless either because it lacked the richness of
sensation and actuation provided by real robots, or because any
kind of embodiment (even virtual) is a distraction from focusing
on the core problems of cognition and language. Some felt it
important to try an early-stage proto-AGI system on a variety of
different problems in a variety of contexts; others felt the best
thing was to single-focus on one hard problem with an AGIhard nature, e.g. Nick Cassimatis, who runs the Human-Level
AI Lab at RPI, argued for object tracking as a focus. In Nicks
view, if you can make an AI system visually track moving objects
in the manner that humans do, then youve solved the crux of the
AGI problem, and the rest will be relatively straightforward so
he saw no strong reason to pay attention to any problem besides
that one. The scope of ideas and views presented was
fascinating, but the diversity of contradictory views was
somewhat intimidating.
The first of the two Evaluation and Metrics workshops resulted in
a paper presented at one of the AGI conferences, written by
40
John Laird and Robert Wray
, summarizing a list of
requirements that any AGI system should fulfill (these are pretty
simple and are summarized in Table 1 below). The second one
didnt lead to any written deliverable, or any clear conclusion, so
40
Its worth noting that this paper, while inspired by the discussions in
the workshop, was not produced collaboratively by the workshop
participants because this would have required a lot more discussion
and argumentation than just having a couple folks write up their own
views afterwards, informed by the discussions.
This is another
indication of the remarkable diversity and fragmentation of the AGI field
today.
217
far as I know, but it did foster some rather interesting

discussions.
Laird and Wrays Requirements for Human-Level

AGI Systems
R0.
When given new tasks, structure is not changed via

reprogramming [Note: the original R0 was Fixed structure
for all tasks]
R1.
Realize a symbol system

Represent and effectively use:
R2.
Modality-specific knowledge
R3.
Large bodies of diverse knowledge
R4.
Knowledge with different levels of generality
R5.
Diverse levels of knowledge
R6.
Beliefs independent of current perception
R7.
Rich, hierarchical control knowledge
R8.
Meta-cognitive knowledge
R9.
Support a
deliberation
spectrum
of
bounded
and
unbounded
R10. Support diverse, comprehensive learning

R11. Support incremental, online learning
Laird and Wrays Requirements for Environments

for Human-Level AGI Systems
C1.
Environment is complex with diverse, interacting objects
C2.
Environment is dynamic
C3.
Task-relevant regularities exist at multiple time scales
C4.
Other agents impact performance
C5.
Tasks can be complex, diverse and novel
C6.
Agent/Environment/Task interactions are complex and

limited
C7.
Agent computational resources are limited
218

C8.
Agent existence is long-term and continual
Itamars and my thinking in organizing the AGI Roadmap

Workshop was that, by restricting the scope to researchers
whose approaches had more in common, wed be able to arrive
at more solid and definite conclusions. And this vision was
realized to an extent we did manage to arrive at some
conclusions, even though it didnt prove possible to get all the
participants to agree on a single roadmap.
On the first morning of the AGI Roadmap Workshop, there was a
fair bit of discussion about the overall process of developing a
roadmap and the extent to which this metaphor is even
applicable to the case of AGI. A traditional highway roadmap
shows multiple driving routes across a landscape of cities and
towns, natural features like rivers and mountains, and political
features like state and national borders. A technology roadmap
typically shows a single progression of developmental
milestones from a known starting point to a desired result. Our
first challenge in defining a roadmap for achieving AGI was that
we initially had neither a well-defined starting point nor a
commonly agreed upon target result. The history of both AI and
AGI is replete with this problem, which is somewhat
understandable given the breadth and depth of the subjects of
both human intelligence and computer technology. Borrowing yet
more metaphors from the highway roadmap, we decided to first
define the landscape for AGI and then populate that landscape
with milestones that might be traversed via multiple routes.
Regarding the final destination of the roadmap, we decided not
to try to agree on a precise formal definition of what we meant by
AGI, but rather to adopt early AI pioneer Nils Nilssons
pragmatic goal of a system that could both learn and replicate
human-level performance in a wide variety of tasks, including
those that humans are paid to perform. The starting point was
more problematic, since there are many current approaches to
achieving AGI that assume different initial states. Finally we
settled on a developmental approach to the roadmap, following
219
human cognitive development from birth through adulthood.

Even though none of the AGI researchers at the workshop were
trying to mimic human cognitive development in detail, still this
provided a common vocabulary and series of steps, which
everyone in attendance was willing to accept as a meaningful
guide.
We also seriously pondered the two significant challenges posed
by Laird and Wrap in their write-up of the first Evaluation and
Metrics workshop mentioned above:
...one of the best ways to refine and extend these sets of
requirements and characteristics is to develop agents using
cognitive architectures that test the sufficiency and necessity
of all these and other possible characteristics and
requirements on a variety of real-world tasks. One challenge
is to find tasks and environments where all of these
characteristics are active, and thus all of the requirements
must be confronted. A second challenge is that the existence
of an architecture that achieves a subset of these
requirements, does not guarantee that such an architecture
can be extended to achieve other requirements while
maintaining satisfaction of the original set of requirements.
One point that was very clear was that our goal in the workshop
was to take up the first challenge of finding appropriate tasks and
environments to assess AGI systems. There was a general
consensus that the second challenge would be more
appropriately handled by individual research efforts. Our hope
was that, even though the various workshop participants had
different approaches to building AGI, we could get a degree of
consensus on what tasks and environments were most
appropriate for AGI development, and what is the overall
landscape of the AGI field.
Perhaps the biggest choice we made was to focus the workshop
on specific scenarios for AGI development a scenario being a
sort of conceptual bundle consisting of an environment for AGI
systems to interact, experiment and learn and be taught and
tested in, together with a set of tasks and goals in that
environment. We made a serious effort to get everyone in the
220
room to agree on a SINGLE scenario for AGI development a

single environment and task set that everyone could use to test
and teach and develop their AGI system, and in terms of which
everyone would be content to have their AGI system evaluated.
Unfortunately but perhaps inevitably, this aspect of our effort
failed we did not get consensus on any single scenario.
Rather, different researchers strongly advocated different
scenarios, consistent with their different research interests and
conceptual predispositions. However, we did achieve some
reasonable consensus on questions like what makes a good
scenario and, given a scenario, what makes a good roadmap
to AGI within that scenario. Ill discuss some of the lessons we
learned just below
So, in the end, the result of the workshop was both disappointing
and highly educational. It would have been great to come out of
it with a single set of milestones, agreed on by all involved: if
your system can pass STEP 1 we all agree its 10% of the way to
human-level AGI; if your system can pass STEP 2 we all agree
its 20% of the way to human-level AGI; etc. Having this kind of
agreement among a host of AGI researchers would go a long
way toward making the AGI field comprehensible to various
parties such as students, neuroscientists, funding agencies
and so forth. It would also make it easier for AGI researchers to
understand each others work, and learn from each others
successes and failures. But, well, this didnt happen. We were
mostly able to agree that each others suggested scenarios were
reasonable and sensible but still, each of us strongly preferred
our own pet scenarios to those proposed by the others, usually
for reasons deeply grounded in our own understanding of the
AGI problem. There wasnt a lot of grandstanding or personality
conflict or misunderstanding among the group (remarkably little,
actually, given the sizes of some of the egos involved, my own
not excepted!); it was a congenial and respectful gathering, not
just on the surface but also in terms of deeper intellectual
communication. Ultimately our inability to agree on a single
scenario came down to the fact that we had different intuitions
221
regarding which parts of the AGI problem should be attacked

first, because of our different intuitions regarding what are really
the hard parts at the crux of human-level general intelligence.
The Breadth of Human Competencies

One of the things we did all manage to agree on, at the
workshop, was the broad and deep set of competencies that a
system needs to demonstrate, in order to have human-level
general intelligence. Its a big long list and theres no way to
make it small and elegant the best we could do is the following
table, which lists key competency areas, and then some of their
important sub-areas:
222
Broad
competency
areas
Subareas...
Subareas..
Subareas..
Subareas..
Subareas..
Perception
Vision
Audition
Touch
Proprioception
Crossmodal
Actuation
Physical
skills
Tool use Navigation Proprioception
Memory
Implicit
Working
Learning
Imitation
Reinforce Dialogical Via Written Via

ment
Media
Experimentation
Reasoning
Deduction Induction Abduction Causal
Planning
Tactical
Strategic Physical
Attention
Visual
Social
Motivation
Subgoal
creation
Affectbased
Emotion
Emotional UnderPerceiving Control of

expression standing emotions emotions
emotions
Modeling self
and other
SelfTheory of Self-control OtherEmpathy

awareness mind
awareness
Social
interaction
Appropriate
behavior
Episodic
Semantic
Subareas..
Procedural
Physical
Associational
Social
Behavioral
Social
Social
commun- inference
ication
Cooperation, e.g.
group play
Communication Gestural
Verbal
Pictorial
Language Crossacquisition modal
Quantitative
Counting
observed
entities
Grounded
small
number
arithmetic
Comparison of
quantitative
properties
of
observed
entities
Measurement using
simple
tools
Building/
creation
Physical
construction w/
objects
Formation Verbal
of novel invention
concepts
Social
organization
223
Thats a lot of stuff but human-level general intelligence

involves a lot of stuff!! It may be hard, at first, to see how all this
stuff comes out of the general picture of general intelligence as
achieving complex goals in complex environments but its not
so hard when you think about the specific goals and
environments that we humans have to deal with today, and that
our ancestors had to deal with when they were evolving humanlevel general intelligence. Take away any one of the
competencies in the table, and youll have an impaired,
subnormal human, inadequately capable of achieving normal
human-type goals in everyday human environments.
Of course the precise formulation of these competencies in the
above table is nothing special. You could split some of the
categories into multiple ones, or merge some of them together.
Modulo some changes in wording, this is basically the table of
contents of a cognitive psychology textbook, or an AI textbook.
Narrow AI systems are typically focused on a very small
percentage of the competencies in the table. For instance, even
if an AI system could answer questions involving previously
known entities and concepts vastly more intelligently than
Watson, if it could not create new concepts when the situation
called for it, we would be reluctant to classify it as Human Level.
Everyone at the workshop agreed it was critical to think about
competencies in this sort of broad manner because otherwise
it is too easy, when developing and evaluating your AGI system,
to pair a testing environment with an overly limited set of tasks
biased to the limitations of that environment. We didnt focus that
heavily on any particular competency list (though we did make
one, similar to the above table, for sake of concreteness), but
more so on the approach of exploring a diverse range of
competency areas, and then generating tasks that evaluate the
manifestation of one or more articulated competency areas
within specified environments.
224
Scenarios for Assessing AGI

So what were the scenarios conceived by the participants in the
AGI Roadmap Workshop? Many were discussed, but seven
received particular attention Ill now run through them one by
one.
General Video-game Learning
Sam Adams, from IBM, was a major force in the AGI Roadmap
workshop, and also big advocate of video games as a scenario
for AGI development. In this scenario he views it, the goal for
AGI shouldnt be human-level performance at any single video
game, but the ability to learn and succeed at a wide range of
video games, including new games unknown to the AGI
developers before the competition. The AGI system would also
be limited to a sensory/motor interface to the game, such as
video and audio output and controller input, and would be
blocked from any access to the internal programming or states of
the game implementation. To provide for motivation and
performance feedback during the game, the normal scoring
output would be mapped to a standard hedonic (pain/pleasure)
interface for the AGI so it would get pleasure from winning, and
pain from losing (however it defines pleasure and pain internally,
which may be different for different AGI systems). The AGI
system would have to learn the nature of the game through
experimentation and observation, by manipulating game controls
and observing the results. The scores against a preprogrammed
opponent or storyline would provide a standard measure of
achievement along with the time taken to learn and win each
kind of game.
The range of video games used for testing in this scenario could
be open ended in both simplicity and sophistication; all the way
from Pong to Starcraft and World of Warcraft. Since success at
most video games would require some level of visual
intelligence, General Video Game Learning would also provide a
good test of computer vision techniques, ranging from simple 2D
225
object identification and tracking in PONG to full 3D perception

and recognition in a game like HALO or HalfLife at the highest
levels of performance. Various genres of video games such as
the early side-scrolling Mario Brothers games to First Person
Shooters like DOOM to flight simulations like the Star Wars XWing series provide rich niches where different researchers
could focus and excel, while the common interface would still
allow application of learned skills to other genres.
Video game playing may seem a simple thing but as any
gamer would tell you, its really not. In some ways its more
difficult, and more demanding on intelligence, than everyday life.
Among other things, in order to effectively play many videos
games a notable degree of strategic thinking would need to be
demonstrated. This refers to the ability to map situations to
actions, while considering not just the short-term but also the
longer-term implications of choosing an action. And figuring out
the ins and outs of a game via experiential fiddling-around is a
wonderful case study in experiential learning.
Preschool Learning
My own preferred scenario for AGI development, which I
advocated at the AGI Roadmap workshop, the Evaluation and
Metrics workshop, and a lot of other AGI gatherings before and
after, is aimed at even simpler childhood preoccupations than
video games. In the spirit of the popular book All I Really Need
to Know I Learned in Kindergarten, my suggestion is to consider
early childhood education such as kindergarten or preschool as
inspiration for scenarios for teaching and testing human-level
AGI systems. I presented a paper on this theme at the AGI-09
conference, titled AGI Preschool.
This idea has two obvious variants: a physical preschool-like
setting involving an AI-controlled robot, and a virtual-world
preschool involving an AI-controlled virtual agent. The goal in
such scenarios is not to precisely imitate human child behavior,
but rather to use AI to control a robot or virtual agent qualitatively
displaying similar cognitive behaviors to a young human child. In
226
fact this sort of idea has a long and venerable history in the AI
field Alan Turings original 1950 paper on AI, where he
proposed the Turing Test, contains the suggestion that Instead
of trying to produce a programme to simulate the adult mind, why
not rather try to produce one which simulates the childs?
This childlike cognition based approach seems promising for
many reasons, including its integrative nature: what a young
child does involves a combination of perception, actuation,
linguistic and pictorial communication, social interaction,
conceptual problem solving and creative imagination. Human
intelligence develops in response to the demands of richly
interactive environments, and a preschool is specifically
designed to be a richly interactive environment with the capability
to stimulate diverse mental growth. The richness of the
preschool environment suggests that significant value is added
by the robotics based approach; but a lot can also potentially be
done by stretching the boundaries of current virtual world
technology.
Another advantage of focusing on childlike cognition is that child
psychologists have created a variety of instruments for
measuring child intelligence. So in a preschool context, one can
present ones AI system with variants of tasks typically used to
measure the intelligence of young human children.
It doesnt necessarily make sense to outfit a virtual or robot
preschool as a precise imitation of a human preschool this
would be inappropriate since a contemporary robotic or virtual
body is rather differently capable than that of a young human
child. The aim in constructing an AGI preschool environment
should rather be to emulate the basic diversity and educational
character of a typical human preschool.
To imitate the general character of a human preschool, I would
to create several centers in a virtual or robot preschool. The
precise architecture will be adapted via experience but initial
centers might be, for instance:
227
a blocks center: a table with blocks of various shapes

and sizes on it
a language center: a circle of chairs, intended for people

to sit around and talk with the AI
a manipulatives center, with a variety of different objects

of different shapes and sizes, intended to teach visual
and motor skills
a ball play center: Where balls are kept in chests and

there is space for the AI to kick or throw the balls around
a dramatics center where the AI can observe and enact

various movements
Since I really enjoy playing with young children, I find the idea of
an AGI preschool not only theoretically compelling, but an awful
lot of fun.
Story/Scene Comprehension
Joscha Bach liked the preschool approach OK, but he felt it
placed too much emphasis on actuation (moving around,
building things, etc.), which he felt were fairly peripheral to AGI.
So he prefers to focus a little later in the school curriculum, on a
set of tasks he calls scene and story comprehension.
Scene comprehension here does not mean only illustrations,
but real-world scenes, which can be presented at different
granularities, media and difficulties (cartoons, movies, or
theatrical performances for instance). This approach differs from
the reading curriculum scenario, in that it more directly provides
a dynamic environment. If group exercises are included then all
the Laird/Wray criteria are fulfilled in a direct and obvious way.
For instance, a scene comprehension task might involve
watching 10 minutes of a Hollywood movie, and concisely
explaining whats happening or working together with a few
others to concisely explain whats happening. Or it might involve
228
re-writing a movie scene as a story, or drawing a picture to

illustrate the main events of a story, etc. This approach lends
itself to a variety of standardizable test scenarios, which allow
the direct comparison of competing AGI architectures with each
other, and with child performance.
Reading Comprehension
Stuart Shapiro favored a scenario closely related to Joschas, but
a bit simpler and more particular focused on the reading
curriculum.
In this scenario, an aspiring AGI should work
through the grade school reading curriculum, and take and pass
the assessments normally used to assess the progress of human
children. This requires the obvious: understand a natural
language text, and answer questions about it. However, it also
requires some not so obvious abilities.
Very early readers are picture books that tightly integrate the
pictures with the text. In some, the story is mostly conveyed
through the pictures. In order to understand the story, the
pictures must be understood as well as the NL text. This requires
recognizing the characters and what the characters are doing.
Connections must be drawn between characters and events
mentioned in the text and illustrated in the pictures. The actions
that the characters are performing must be recognized from
snapshot poses, unlike the more usual action recognition from
a sequence of frames taken from a video.
The next stage of readers are early chapter books, which use
pictures to expand on the text. Although the story is now mainly
advanced through the text, reference resolution with the pictures
is still important for understanding.
This scenario fulfills most of the Laird/Wray criteria handily, and
a few less obviously. The AGIs environment in this scenario is
not dynamic in a direct sense (C2), but the AGI does have to
reason about a dynamic environment to fulfill the tasks. In a
sense, the tasks involved are fixed rather than novel (C5), but
229
they are novel to the AGI as it proceeds through them. Other

agents impact task performance (C4) if group exercises are
involved in the curriculum (which indeed is sometime the case).
Many of the specific abilities needed for this scenario are
discussed in the next scenario, on scene and story
comprehension.
Learning School
The virtual school student scenario, favored by Alexei
Samsonovich, continues the virtual preschool scenario
discussed above, but is focused on higher cognitive abilities,
assuming that, if necessary, lower-level skills will be finessed. In
particular, it is assumed that all interface with the agent is
implemented at a symbolic level: The agent is not required to
process a video stream, to recognize speech and gestures, to
balance its body and avoid obstacles while moving in space, etc.
All this can be added as parts of the challenge, but part of the
concept underlying the scenario is that it can also be finessed.
On the other hand, it is critical to the scenario for the agent to
make academic progress at a human student level, to
understand human minds, and to understand and use practically
class-related social relations in the environment in which it is
embedded.
In one form of this scenario, the agent is embedded in a real high
school classroom by means of a virtual-reality-based interface.
The agent lives in a symbolic virtual world that is continuously
displayed on a big screen in the classroom. The virtual world
includes a virtual classroom represented at a symbolic (object)
level, including the human instructor and human students
represented by simplistic avatars. The agent itself is represented
by an avatar in this virtual classroom. The symbolic virtual world
is synchronized with the real physical world with the assistance
of intelligent monitoring and recording equipment performing
scene analysis, speech recognition, language comprehension,
gesture recognition, etc. (if necessary, some or all of these
functions will be performed by hidden human personnel running
230
the test; students should not be aware of their existence). The

study material, including the textbook and other curriculum
materials available to each student, will be encoded
electronically and made available to the agent at a symbolic
level.
Furthermore, in this scenario as Alexei envisions it, the agent will
be evaluated not only based on its learning and problem solving
performance, but also based on its approach to problem solving
and based on its interactions with students and with the
instructor. Social performance of the agent can be evaluated
based on surveys of students and using standard psychological
metrics. Another, potentially practically important measure is the
effect of the agent presence in the classroom on student
learning. All this is way beyond preschool or story/scene
comprehension yet ultimately measures the same core
competencies, in a different way.
The Wozniak Coffee Test
And now for something completely different: One of the reasons I
invited Josh Hall to the AGI Roadmap Workshop was that he
was not only a deep AGI theorist and practitioner, but also an
enthusiast regarding the centrality of robotics for AGI. At the time
he was in the midst of building his own home robot for testing his
AGI ideas. I knew Josh would be a strong advocate for the
exploration of AGI in a physical robotics context and while I
dont hold this position nearly as strongly as Josh does, I felt it
was an important sort of perspective to include in the workshop.
In an interview a few years ago, Steve Wozniak of Apple
Computer fame expressed doubt that there would never be a
robot that could walk into an unfamiliar house and make a cup of
coffee. Josh argued that this task is demanding enough to stand
as a Turing Test equivalent for embodied AGI and he
convinced most, but not all, of the workshop participants of this
proposition. (Note that the Wozniak Test is a single, special case
of Nils Nilssons general Employment Test for Human-Level AI,
mentioned above!)
231
In the Wozniak Test as Josh presented it, a robot is placed at the

door of a typical house or apartment. It must find a doorbell or
knocker, or simply knock on the door. When the door is
answered, it must explain itself to the householder and enter
once it has been invited in. (We will assume that the householder
has agreed to allow the test in her house, but is otherwise
completely unconnected with the team doing the experiment,
and indeed has no special knowledge of AI or robotics at all.)
The robot must enter the house, find the kitchen, locate local
coffee-making supplies and equipment, make coffee to the
householders taste, and serve it in some other room. It is
allowed, indeed required by some of the specifics, for the robot
to ask questions of the householder, but it may not be physically
assisted in any way.
The current state of the robotics art falls short of this capability in
a number of ways. The robot will need to use vision to navigate,
identify objects, possibly identify gestures (the coffees in that
cabinet over there), and to coordinate complex manipulations.
Manipulation and physical modeling in a tight feedback learning
loop may be necessary, for example, to pour coffee from an
unfamiliar pot into an unfamiliar cup. Speech recognition and
natural language understanding and generation will be
necessary. Planning must be done at a host of levels ranging
from manipulator paths to coffee brewing sequences.
But the major advance needed for a coffee-making robot is that
all of these capabilities must be coordinated and used
appropriately and coherently in aid of the overall goal. The usual
set-up, task definition, and so forth are gone from standard
narrow AI formulations of problems in all these areas; the robot
has to find the problems as well as to solve them. That makes
coffee-making in an unfamiliar house a strenuous test of a
systems adaptiveness and ability to deploy common sense.
Although standard shortcuts might be used, such as having a
database of every manufactured coffeemaker built in, it would be
prohibitive to have the actual manipulation sequences for each
232
one pre-programmed, especially given the variability in

workspace geometry, dispensers and containers of coffee
grounds, and so forth. Generalization, reasoning by analogy, and
in particular learning from example and practice are almost
certain to be necessary for the system to be practical.
Coffee-making is a task that most 10-year-old humans can do
reliably with a modicum of experience. A weeks worth of being
shown and practicing coffee making in a variety of homes with a
variety of methods would provide the grounding for enough
generality that a 10 year old could make coffee in the vast
majority of homes in a Wozniak Test. Another advantage to this
test is it would be extremely difficult to game or cheat, since the
only reasonably economical way to approach the task would be
to build general learning skills and have a robot that is capable of
learning not only to make coffee but any similar domestic chore.
Other AGI Scenarios

The scenarios Ive just listed are the ones that were most avidly
advocated by AGI Roadmap Workshop participants, but we also
discussed quite a lot of others, like:
Various skilled physical activities, e.g., sports, dance
Diagram learning, reading and composition
Vocal/aural language learning, singing
Search and rescue skills, danger avoidance
Mixed shopping (online and in-store)
Appreciating and composing various arts/music
Social participation in various activities/groups
There is no end of scenarios that are rich enough to serve as

teaching, testing and learning environments for AGI systems.
Really all one needs is a scenario that allows an AGI system to
learn and manifest all the core competencies of human-level
general intelligence, as we crudely summarized in the big table
233
above. There are lots of ways to do this, thus the diversity of

scenarios suggested and advocated at the workshop.
Some of the scenarios were clearly perceived as richer than
others. Everybody agreed that the preschool or virtual-classroom
student scenarios would be adequate in terms of covering all
competencies. On the other hand, some participants felt that the
video game, coffee test, story/scene comprehension or reading
curriculum tests would be inadequate for AGI, in the sense that
they failed to cover certain core competencies, and could
potentially be conquered by very sophisticated narrow AI
systems without human-level general intelligence. In this sense,
only the preschool and classroom scenarios passed the test of
being considered adequate by all the workshop participants
which is not too surprising, as these two scenarios are directly
drawn from human childhood cognitive development.
When the workshop started, we hoped it might be possible for all
participants to agree on a single scenario, and then jointly test
their systems in that same scenario. This didnt pan out, basically
because each participant viewed different research questions as
most critical to AGI, and wanted to use a scenario directly and
immediately probing these research questions. Since Josh Hall
felt that integration of perception, actuation and cognition is the
core aspect of AGI, of course he preferred to work with a
robotics scenario. Since Stuart Shapiro felt that linguistic
understanding and reasoning is particularly core to AGI, of
course he prefers to work with a language-centric scenario etc.
Both Josh and Stuart agree that perception, actuation and
language are all important. But they differ in their intuition
regarding which part of the overall AGI problem is so critical that
it really should be addressed before the other aspects are
messed with.
Does this mean that the AGI field is hopelessly confused, and
none of us AGI researchers know what were doing? Not really.
Theres no reason to believe theres one single correct approach
to AGI. There may be many different routes leading to the same
234
basic place thus the landscape metaphor, in which there may

be many different roads from point A to point B. A roadmap
doesnt usually describe a single road it charts all the different
roads in a given part of the landscape, crisscrossing and
intersecting and diverging as the terrain and other factors dictate.
The diversity of plausible approaches to AGI poses the field
some practical difficulties it may well be that if researchers
worked together more closely, progress would be faster. But this
doesnt imply that any of the researchers are on the wrong track
(though most probably some of them are!). My suspicion is that
the field will continue to be fairly diverse and fragmented, until
such a point as someone achieves something sufficiently
impressive in the way of a practical AGI demonstration. When
we finally reach an AGI Sputnik moment, then a significant
plurality of researchers will jump on the bandwagon of whatever
approach leads to that AGI Sputnik while others will band
together to pursue other approaches, hoping to overtake that
dominant Sputnik technology. At that point the AGI field will
become a lot more interesting, a lot better-funded, a lot more
famous and a lot faster-moving. And I fully expect to see this in
my lifetime quite possibly, er, a lot faster than you think!
From Scenarios to Tasks and Metrics

Getting back to the nitty-gritty once one has chosen a scenario
to work with (preschool, coffee test, or whatever), then how does
one go about systematically using this scenario to teach, test
and host an AGI system? The next step is to combine the
scenario with the laundry list of competency areas articulated in
the table above. For each of the scenarios reviewed above (or
other analogous scenarios), one can make a corresponding AGI
test suite by articulating a specific task-set, where each task
addresses one or more of the competency areas in the context
of the scenario. To constitute a satisfactory AGI test suite, the
total set of tasks for a scenario must cover all the competency
areas. Each task must also be associated with some particular
performance metric, some way to measure the AGIs degree of
235
success quantitative wherever possible, but perhaps qualitative

in some cases depending on the nature of the task.
For sake of brief illustration, the table below roughly describes a
handful of example tasks corresponding to a few of the scenarios
described above. If one really wanted to make a thorough AGI
test suite, one would need to systematically create a substantial
list of tasks corresponding to the scenario(s) chosen, covering all
the competency areas, and each associated with rigorous
performance metrics.
Scenario
Competency
Area
Sub-area
Example Task or Task Family
Virtual
Preschool
Learning
Dialogical
Learn to build a particular structure of

blocks (say, a pyramid) faster based on
a combination of imitation, reinforcement
and verbal instruction, than by imitation
and reinforcement without verbal
instruction
Virtual
Preschool
Modeling
Self and
Other
Theory of
Mind
While Sam is in the room, Ben puts the

red ball in
the red box. Then Sam leaves and Ben
moves the red ball to the blue box. Sam
returns and Ben asks him where the red
ball is. The agent is asked where Sam
thinks the ball is.
Virtual
School
Student
Learning
via Written
Media
Starting from initially available basic

concepts (a number, a variable, a
function), demonstrate academic
progress in learning how to solve
problems from the textbook using
techniques described in the same
textbook. The agent should move step
by step, from simple to advanced
problems, from one domain to another.
236
Virtual
School
Student
Modeling
Self and
Other
OtherAwareness
Help a friend to cheat on exam (modified

from [REF Samsonovich, 2006]. The virtual
agent and a student take an exam consisting
of one and the same problem. First, the agent
enters
a room and works on the problem for an hour,
then the student does the same. When the
agent exits the exam room after solving the
problem, the student comes in, and the agent
has a chance to pass a small piece of paper
to the student. The entire solution of the exam
problem cannot fit on this piece of paper, not
even a substantial part of it. The challenge for
the agent is to write a short message that
would give a useful hint to the friend.
Robot
Preschool
Actuation
Proprioception
The teacher moves the robot's body into

a certain configuration.
The robot is asked to restore its body to
an ordinary standing position, and then
repeat the configuration that the
teacher moved it into.
Robot
Preschool
Memory
Episodic
Ask the robot about events that occurred

at times when
it got particularly much, or particularly
little, reward for its actions; it should be
able to answer simple questions
about these significant events, with
significantly more accuracy than about
events occurring at random times
237
Scenario
Competency
Area
Sub-area
Example Task or Task Family
Wozniak
Coffee Test
Communication
Gestural
In many cases the robot will be shown to

the kitchen. It must understand gestures
indicating that it should follow an
indicated path, or know how to follow its
guide, and know when either is
appropriate.
Wozniak
Coffee Test
Actuation
Navigation
The robot must go about its business

without running into people, walls,
furniture, or pets.
Wozniak
Coffee Test
Appropriate The robot had better be able to recognize

Social
the case where it has knocked on the
Interaction behavior
wrong door and the householder is not
inviting it in.
Wozniak
Coffee Test
Reasoning
Physical
Consider the state of knowledge that tells

us we can use a drip
pot without its top, but not a percolator.
This may come from
physical simulation, based on an
understanding of naive physics.
Wozniak
Coffee Test
Reasoning Induction
On the other hand, the above-mentioned

knowledge about drip pots
and percolators may be gathered via
inductive reasoning based on
observations in multiple relevant
situations.
Example Tasks and Task Families Corresponding to Various

Scenarios, Addressing Human General Intelligence Competencies
One of the conclusions of the AGI Roadmap Workshop was

that as these example tasks indicate is that the right way to
test and challenge and teach a would-be AGI system is
238
something like: To address a few dozen closely interrelated,

moderately broadly-defined tasks in an environment drawn
from a moderately broad class thereof (e.g. an arbitrary
preschool or textbook rather than specific ones determined in
advance). If the environment is suitably rich and the tasks are
drawn to reflect the spectrum of human general intelligence
competencies, then this sort of challenge will motivate the
development of genuine Human-Level AGI systems.
This sort of challenge is not nearly as neat and clean as, say, a
chess contest, RoboCup or the DARPA Grand Challenge. But
we feel the complexity and heterogeneity here is directly
reflective of the complexity and heterogeneity of human general
intelligence. Even if there are simple core principles underlying
human-level general intelligence, as I do believe, nevertheless
the real-world manifestation of these principles is complex and
involves multiple interrelated competencies defined via
interaction with a rich, dynamic world.
Such a roadmap does not give a highly rigorous, objective way
of assessing the percentage of progress toward the end-goal of
Human Level AGI. However, it gives a much better sense of
progress than one would have otherwise. For instance, if an AGI
system performed well on diverse metrics corresponding to 50%
of the scenarios listed above, one would seem justified in
claiming to have made very substantial progress toward Human
Level AGI. If an AGI system performed well on diverse metrics
corresponding to 90% of these competency areas, one would
seem justified in claiming to be almost there. Achieving, say,
25% of the metrics would give one a reasonable claim to
interesting AGI progress. This kind of qualitative assessment of
progress is not the most one could hope for, but it is better than
the progress indications one could get without this sort of
roadmap.
The limitations of this kind of roadmapping effort, at the present
stage of AGI research, have already been made clear. There is
no consensus among AGI researchers on the definition of
239
general intelligence, though we can generally agree on a

pragmatic goal. The diversity of scenarios presented reflects a
diversity of perspectives among AGI researchers regarding
which environments and tasks best address the most critical
aspects of Human Level AGI. Most likely neither the tentative list
of competencies nor the Laird/Wray criteria are either necessary
or sufficient. There is no obvious way to formulate a precise
measure of progress toward Human Level AGI based on the
competencies and scenarios provided though one can use
these to motivate potentially useful approximative measures.
But, in spite of these limitations, I think the conclusions of the
AGI Roadmap Workshop constitute non-trivial progress. Further
development of a roadmap along the lines sketched in the
workshop would quite possibly result in something:
compare their work in a meaningful way
allowing researchers, and other observers, to roughly

assess the degree of research allowing multiple
researchers following diverse approaches to
progress toward the end goal of Human Level AGI
allowing work on related road-mapping aspects, such as

tools roadmaps and study of social implications and
potential future applications, to proceed in a more
structured way
Personally, I dont have much taste for committee thinking

and I suspect the other AGI researchers participating in the AGI
Roadmap Workshop feel the same way. Were a maverick,
entrepreneurial bunch. And the real business of creating AGI, I
feel, is going to be carried out by some small maverick team
moving in its own direction, according to its own mix of science
and intuition not bothering to stop to check its ideas with the
community every step of the way.
But even so, I think theres great value in creating tighter
connections within the network of mavericks in sharing ideas
240
and exploring commonalities insofar as possible, in carefully

discerning which differences between research programs are
inessential or terminological, and which reflect substantive
differences in scientific intuition. As a further example of this, I
would like to follow up the AGI Roadmap Workshop with a
comparable AGI Architectures Workshop, focused not on
convincing multiple researchers to agree on a common AGI
architecture, but rather on better understanding the relationships
between the different structures and processes inside various
AGI architectures. This sort of connection-building will never be
as obviously critical as the science and engineering work of
small focused teams, but it has an important role to play
nonetheless.
241
15
Mind Uploading
While my research career has focused mainly on figuring out
how to build AGI, using ideas not that closely based on human
brain structure, I also have a strong interest in figuring out how to
make machines that embody human minds in an exact way.
This is not just out of intellectual interest, but because Id like to
port my OWN mind, and the minds of my friends and family
members and anyone else who wants it, into the Internet or
superior robot bodies or wherever they want to go. Once a
human mind is uploaded, it would have the option to remain
human, or potentially to upgrade itself into some radically
superior form maybe fusing with other AGIs. Life extension in
the human body form is interesting too, but ultimately it's not as
fascinating and exciting to me as the prospect of moving my
mind through various different bodies, including those supporting
much greater intelligence than the traditional human body
evolution has supplied us with.
I wrote these thoughts on mind uploading in early 2011. Shortly
after that I pursued my interest in mind uploading in a more
intense way, via editing the first-ever scientific journal issue on
mind uploading (the Special Issue on Mind Uploading of the
41
International Journal of Machine Consciousness kudos
are due to editor Antonio Chella for being visionary enough to
allow such a special issue to exist; and thanks to my co-editor
41
See http://www.worldscientific.com/toc/ijmc/04/01
for the official site of the Special Issue on Mind
Uploading; and http://wp.goertzel.org/?page_id=368
for unofficial preprints of some of the articles.
242
Mind Uploading
Matt Ikle for helping with putting the issue together, and James
Hughes and Randal Koene for help with gathering papers). I
havent done any actual research directly toward the goal of
mind uploading yet, but Id certainly love to. As neurotech
progresses and more and more data about the brain is gathered,
there will doubtless be chances to apply AGI and/or narrow-AI
tech to the problem of reconstructing minds in new substrates.
I posted this to my gmail account on Tuesday, September 28,
2010:
To whom it may concern:
I am writing this in 2010. My Gmail account has more than
20GB of data, which contain some information about me and
also some information about the persons I have exchanged
email with, including some personal and private information.
I am assuming that in 2060 (50 years from now), my Gmail
account will have hundreds or thousands of TB of data,
which will contain a lot of information about me and the
persons I exchanged email with, including a lot of personal
and private information. I am also assuming that, in 2060:
1. The data in the accounts of all Gmail users since 2004 is
available.
2. AI-based mindware technology able to reconstruct
individual mindfiles by analyzing the information in their
aggregate Gmail accounts and other available
information, with sufficient accuracy for mind uploading
via detailed personality reconstruction, is available.
3. The technology to crack Gmail passwords is available,
but illegal without the consent of the account owners (or
their heirs).
4. Many of today's Gmail users, including myself, are
already dead and cannot give permission to use the data
in their accounts.
243
If all assumptions above are correct, I hereby give

permission to Google and/or other parties to read all data in
my Gmail account and use them together with other
available information to reconstruct my mindfile with
sufficient accuracy for mind uploading via detailed
personality reconstruction, and express my wish that they do
so.
Signed by Ben Goertzel on September 28, 2010, and
witnessed by readers.
NOTE: The accuracy of the process outlined above
increases with the number of persons who give their
permission to do the same. You can give your permission in
comments, Twitter or other public spaces.
Any cleverness there is due to Giulio Prisco, who sent a similar
email, which I shamelessly plagiarized. But the core idea goes
back further, of course, and has most recently been publicized
by Martine Rothblatt via her Cyberev.org website. Martine, a
renowned futurist and technology entrepreneur, advocates the
notion of cyber-immortality the creation of a digital
simulacrum of a particular human being, created via assembling
fragments of information left behind by that human being, e.g.
texts, videos, audio files, and so forth.
Martine seems quite confident that its possible to create a digital
entity that not merely resembles a person, but in a fundamental
sense IS that person, merely via reverse engineering the data
left behind by that person. When I asked her if she felt her
cyber-upload created in such a fashion would really be her,
she answered Yes, in the same sense that Im still the same
person I was before I went to sleep last night.
Im not so sure as her, but its an interesting hypothesis. Ive left
a lot of data behind already, in terms of texts and videos and
chat sessions and so forth. Sure, one could make a fancy superduper chat bot of some sort that was capable of producing data
244
Mind Uploading
similar to what Ive left behind but what if that werent ones
goal? What if ones goal were to re-create ME, via creating an
entity operating by closely human brain-like dynamics, yet still
with the capability of giving rise to my texts, videos, chats, etc.?
Could there really be another human being another intelligence
with a human brain like architecture fundamentally different
from Ben Goertzel, yet capable of giving rise to the same large
corpus of utterances, gestures and interactions as Ben Goertzel?
My heart says no it seems like to say all those things and
make all those face and body movements the same as me, using
a human brain and body or something similar, youd have to BE
me.
Of course, we lack the technology today to perform this kind of
reverse engineering to infer the brain from the behavior. But
storing data is relatively cheap these days so its quite viable to
store massive amounts of data about a person and keep it
around till the technology develops. And if you doubt the
possibility of any human technology successfully performing this
kind of reverse engineering what about an advanced AGI? In
the end, what were talking about here is an optimization
problem: find a neural-like structure whose behaviors, in a
certain set of contexts, would emulate a certain database of
behaviors. This is a difficult optimization problem, yet bounded
in size, with bounds expressible in terms of human brain size
and behavior database size. Its hard for me to believe it will
exceed the capabilities of post-Singularity AGIs.
Mind Uploading Via Brain Emulation

But of course, reverse-engineering minds from videos and
emails and such is not the only way to create digital versions of
human beings. Another obvious possibility actively pursued by
an increasing community of researchers is to scan a particular
human beings brain structure, and emulate this in digital form. A
combination method may also be viable use the brain
structure, AND use information about the persons behaviors as
well. The more information the better!
245
Regarding the creation of digital emulations of particular human

brains, the major question is at what level does one need to
emulate a brain in order to emulate the person associated with
that brain? Is it enough to emulate the network of neurons?
Does one need the neurotransmitter concentrations? The glia as
well as the neurons? A few theorists (e.g. Stuart Hameroff)
believe one would need to use a quantum computer (or quantum
gravity computer) and emulate the quantum structures inside the
cell walls of neurons, and (hypothetically) underlying protein
function, etc. The general feeling among neuroscientists today
seems to be that emulating the neurons and glia should probably
be good enough, but neuroscience is a discipline continually
revolutionizing itself, so I wont be shocked if the story changes a
few years down the road.
Neuroscientist Randal Koene has been particularly active in
advocating mind uploading via brain emulation, via his technical
work and his organization of seminars and mini-conferences
under the flag of ASIM, Achieving Substrate-Independent
Minds. And Nick Bostroms Future of Humanity Institute, at
Oxford University, held a workshop in 2007 aimed at carefully
assessing the viability of mind uploading via various means.
Among the conclusions of this workshop was that, in the
reasonably near future, the most likely method of translating
brain structures into digital computer structures is to work with
frozen brains rather than live brains.
Current methods for scanning live brains are rather crude via the
standards of mind uploading one has to choose between the
spatial accuracy of fMRI and the temporal accuracy of MEG or
EEG, but one cannot have both. But to scan the state of a brain
in a manner meaningful for mind uploading, one needs
reasonable high degrees of both spatial and temporal accuracy,
coordinated together. On the other hand, if one freezes a brain
from a recently decapitated head, slices it thinly and scans the
slices, then one can get a quite detailed picture of whats going
on in a brain at a particular moment. One can see the
concentrations of neurotransmitter molecules in fairly small
246
Mind Uploading
regions, thus understanding the flow of charge between neurons

and the way its been modulated by learning. Reconstructing a
neuronal network from this kind of imaging data is a nontrivial
data analysis task, and then once this is done, one still has to
use some serious biophysics to figure out the dynamics of the
network from this instantaneous snapshot. But, in the frozen
brain scenario, there is at least a very clear path forward its
clear what the science problems are, and there are wellunderstood approaches to solving them. The main bottleneck is
funding, as this sort of research is not currently prioritized by
corporations or funding agencies.
Is Your Upload You?

The philosophical issues associated with this kind of mind
uploading have been debated extremely extensively in the
futurist community. For a while (maybe still) the topic was
banned from the Extropy futurist email discussion list, because
the repetitive arguments back and forth had become so tedious,
and it was felt that nobody had anything new to say on the
matter. The basic question is, what makes you YOU? What if a
digital system is created, whose simulated neural activity
patterns are exactly the same as your biological neural activity
patterns, and whose behaviors are exactly the same as yours
would be in the same contexts is this you, or is it a copy of
you?
If you believe that you are something that emerges from the
same structures and dynamics of the brain and body that give
rise to your behavior, then the when is my upload me? question
is purely philosophical or spiritual, not scientific. It only becomes
a scientific question if you think that either
there is some scientifically measurable essence that

constitutes you, but that is not part of the measurable
structures and dynamics of your brain and body maybe
some kind of soul field pattern in some kind of
physically measurable (but not yet scientifically
247
understood) mind field, associated with but not

contained in the brain or body; or
the essence of you emerges from some aspects of

your brain or body that dont need to be emulated in
order to precisely emulate your behavior
One of the more interesting thought-experiments in this area is

the notion of creating your digital upload one neuron at a time.
What if you first replace one of your neurons with a digital
neuron, then another, then anotherUntil theyre all replaced.
What if you still feel like you the whole time? Then is the final
result still you?
But what if the biological neurons, that have been removed from
your brain, are simultaneously used to assemble another
biological you who is only awakened once his new brain is
complete. Is that one still you, also? Are they both you?
My view on this is, these issues are going to seem increasingly
irrelevant as the technologies involved get more and more
mature. Our everyday lives are already filled with apparently
insoluble philosophical conundrums. None of us can be sure that
our wives, children or colleagues are really conscious beings in
the same sense that we are personally we just assume it out of
convenience and intuition. Humes problem of induction was
never solved that is, we have no rigorous understanding of
why, just because the laws of physics have appeared to apply
every day of our lives so far, this will still be the case tomorrow.
Once mind uploading is a reality, well accept uploads as real
selves without worrying about the philosophy, just as we now
dont worry much about whether our friends and family are
conscious.
Mind Uploading and AGI

Whats the relation between mind uploading and AGI? It goes in
two directions.
248
Mind Uploading
First, AGIs may be helpful for performing mind uploading. Mind

uploading in any form requires extraordinarily difficult data
analysis and optimization problems. These may be solvable via a
combination of narrow AI and traditional statistical and
mathematical methods but even so, would surely be solvable
much faster and better with the help of AGI Experts tailored for
this sort of problem, or more general-purpose human-level or
transhuman AGIs.
Second, obviously, an uploaded human mind would itself BE a
kind of Artificial General Intelligence. In a sense its the least
interesting kind, because its just another human mind, and we
already have plenty of those. But, a human mind that can be
experimented with freely (with its permission, one hopes), and
whose dynamics can be studied with arbitrary precision, would
be incredibly helpful to scientists aiming to build a science of the
mind. Mind uploading would be followed reasonably shortly by
mind science, which would lead to a wider variety of AGIs,
deviating further and further from close emulation of human
intelligence. I personally doubt this is the way AGI will initially be
achieved, but I think its a perfectly reasonable route definitely
worth someones close attention!
249
16
AGI Against Aging
I wrote this essay originally in 2009 for Life Extension
Magazine, after some discussions about the power of AI for
longevity research with Life Extension Foundation founder/chief
exec Bill Falloon (who is a great guy) but it was too long for
them and I never got around to writing a shorter version. But I
posted it online and it got a lot of positive responses. Ive given
a load of talks on this theme as well.
2009 was far from the start of my work on AI and biology, though
I first plunged into the area in 2001, shortly after my first AI
company Webmind Inc. disappeared. I founded Biomind LLC in
2002, with a vision of applying advanced AI tech to genomics
and proteomics so as to enable humanity to more rapidly cure
diseases and end aging. By 2009 I had already gotten enough
practical sense for the biology and longevity research spaces, to
feel comfortable prognosticating and pontificating about the best
research directions for the field to follow.
Since writing AGI Against Aging in 2009, Ive done a fair bit
more work in the AI for longevity research arena. For one
thing, my work with Genescient Corp. applying AI to analyze
data from their long-lived flies, briefly discussed in AGI Against
Aging, has gone a lot further by now in fact Im right in the
middle of writing up a technical paper on some of our findings.
Also, a few supplements are now on the market, providing novel
combinations of herbs derived via a combination of my teams
AI analysis with human biological insight -- that address various
ailments (one for inflammation, one for brain aging, and one for
aging in general).
And my OpenCog / Biomind colleagues and I are gradually
moving toward implementation of more of the advanced AI ideas
described here. Just recently (mid-2014), weve finally started
loading various biological ontologies and datasets into
250
AGI Against Aging
OpenCogs Atomspace and experimenting with pattern-finding

on them therein; it looks like in late 2014 and early 2015 well
finally be trying out OpenCogs probabilistic logic algorithms on
biological data (going beyond what Ive done before with AI-forbiology, which has mainly been centered on use of the MOSES
automated program learning algorithm for various sorts of
genetics-data classification and clustering).
But all this more recent work will have to go into a new essay,
yet to be written. What we have here is my basic vision of how
proto-AGI and AGI can help longevity research.
The goal of eluding death long predates science. It has taken
hundreds of forms throughout history, pervading nearly all
cultures and eras. The ancient Chinese, for instance, had Taoist
Yoga, a very complex discipline defining a life-long series of
practices that, if adhered to precisely, purportedly resulted in
physical immortality. Part of this teaching was that by refraining
from ejaculation for his entire life, a man could store his
essential energy in a space by the top of his head, until he
accumulated enough to create an eternal fetus that would grow
into his deathless self. (We havent tracked down any analogous
method for women, by the way.)
Today, with the advent of molecular biology and systems
biology, it looks more and more likely that death is a problem
thats solvable by science without philosophically controversial
approaches like uploading, and in a manner that works equally
well for both me and women. By merely fixing problems in the
human body, much as one upgrades the machinery of an
automobile. Modern biology views the human organism as a
complex machine, which is modifiable and repairable like any
other machine, and there is plenty of data to back up the power
of this perspective. Scientists now have a short list of biological
and biochemical factors suspected to collectively underlie aging.
For each of these likely culprits, there is a pharma firm or a
maverick scientist working on its cure. It is plausible that, within
251
decades, not centuries nor millennia, bioscience will have made

the very concept of getting old obsolete.
Among the more ambitious initiatives out there, Aubrey de
Greys
SENS
(Strategies
for
Engineering
Negligible
Senescence) plan involves seven initiatives addressing seven
different causes of aging:
Aging
Damage
Discovery
SENS Solution
Cell loss,
tissue atrophy
1955
http://www.sens.org/research/introduction-to-sensresearch/cell-loss-and-atrophy
Nuclear
[epi]mutations
(only cancer
matters)
1959, http://www.sens.org/research/introduction-to-sens1982 research/nuclear-mutations
Mutant
mitochondria
1972
http://www.sens.org/research/introduction-to-sensresearch/mitochondrial-mutations
Death-resistant
cells
1965
http://www.sens.org/research/introduction-to-sensresearch/deathresistant-cells
Tissue
stiffening
1958, http://www.sens.org/research/introduction-to-sens1981 research/extracellular-crosslinks
Extracellular
aggregates
1907
http://www.sens.org/research/introduction-to-sensresearch/extracellular-junk
Intracellular
aggregates
1959
http://www.sens.org/research/introduction-to-sensresearch/intracellular-aggregates
Table: De Greys Seven Pillars of Aging

Each of these initiatives has numerous uncertainties involved
with it, and at present theres no way to tell how rapidly any of
them will meet with success. But the work proceeds apace.
In another complementary and equally exciting approach,
evolutionary biologist Michael Rose has used decades of
selective breeding to create a strain of fruit flies living 5 times as
long as normal. Via analyzing their genetics and comparing
them to many other lines of flies, Rose is now studying what
252
AGI Against Aging
makes them live so long, and how to draw lessons from them
about what pharmaceutical or nutritional substances may cure
human disease in order to prolong a healthy human life. His
overall approach to aging involves building a huge integrated
database of biological knowledge about multiple organisms, and
progressively growing this database in synchrony with
experimental evolution on flies and other model organisms.
Unlike de Grey, Rose doesnt even like to talk about aging; he
prefers to talk about disease, figuring that aging is essentially a
combination of one disease on top of another on top of another.
De Greys initiatives may resolve some of the problems
associated with aging, Roses effort may resolve others, and
there are many other researchers out there. Through their
collective activity, the average human lifespan is going to
progressively increase, as the science advances. But what might
cause a rapid acceleration in the progress toward increasing
healthy lifespans? Might it be the application of AGI technology
to life extension research? Narrow AI has already proved itself
valuable for the study of aging related diseases, so this certainly
seems a plausible conjecture.
The Biology of Cell Senescence

Before digging into potential solutions to aging, Ill say a bit about
the nature of the problem. People die as they age for many
reasons: heart disease, neurodegenerative disease, immune
disease... The list is disturbingly long. In addition to these wellknown diseases, another well-known cause of aging is cell
senescence: the fact that cells are, in many cases,
42
preprogrammed to die. Even though its just part of the story, a
quick look at cell senescence will give us a glimpse at the
complexity of the biological processes underlying aging.
42
Note that senescence on the organism level just means biological

changes due to aging (aka getting old) and cellular senescence
(cells getting old) is just one part of this process.
253
A healthy body is not a constant pool of cells, but rather a hotbed

of continual cellular reproduction. There are only a few
exceptions, such as nerve cells, which do not reproduce, but
simply persist throughout an organisms lifespan, slowly dying
off. In youth, newly formed cells outnumber dying cells; but then
from about 25 on, things begin to go downhill, and the number of
newly formed cells is less than the number of cells that die. Little
by little, bit by bit, cells just stop reproducing.
Most types of human cells have a natural limit to the number of
cell divisions they will undergo. This number, usually around 50
or so, is called the Hayflick limit, named after Leonard Hayflick,
the researcher who discovered it in the mid-1960s. Once a cells
Hayflick limit is reached, the cell becomes senescent, and
eventually, it dies.
This may have the sound of inevitability about itbut things start
seeming different when one takes a look at our one-celled
cousins, such as amoebas and paramecia. These creatures
reproduce asexually by dividing into two equal halvesneither
half sensibly classified as parent or child. This means that
essentially, the amoebas alive today are the same ones alive
billions of years ago. These fellows qualified for Social Security
when you could still walk from New York to Casablanca, and yet
theyre still alive today, apparently not having aged one bitcells
untroubled by the Hayflick limit. This nasty business of aging
seems to have come along with multicellularity and sexual
reproductiona fascinating twist on the sex and death
connection that has fascinated so many poets and artists.
Unlike in asexually reproducing creatures, cells in multicellular
organisms fall into two categories: germ-line cells which become
sperm or egg for the next generation; and soma cells that make
up the body. The soma cells are the ones that die, and the
standard answer to Why? is Why not? The disposable soma
theory argues that, in fact, our soma cells die because its of no
value to our DNA to have them keep living forever. It is, in
essence, a waste of energy. Throughout most of the history of
254
AGI Against Aging
macroscopic,
sexually-reproducing
organisms,
immortal
organisms would not have had an evolutionary advantage.
Rather, there was an evolutionary pressure toward organisms
that could evolve faster. If a species is going to evolve rapidly,
its valuable for it to have a relatively quick turnover from one
generation to the next.
There doesnt seem to be any single cellular grim reaper
process causing soma cell senescence. Rather, it would appear
that there several distinct mechanisms, all acting in parallel and
in concert.
There are junk molecules, accumulating inside and outside of
cells, simply clogging up the works. Then there are various
chemical modifications that impair the functioning of molecular
components such as DNA, enzymes, membranes and proteins.
Of all these chemical reactions, oxidation has attracted the most
attention, and various anti-oxidant substances are on the market
as potential aging remedies. Another major chemical culprit is
cross-linking, the occasional formation of unwanted bridges
between protein molecules in the DNAbridges which cannot be
broken by the cell repair enzymes, interfering in the production of
RNA by DNA. Cross-linkages in protein and DNA can be caused
by many chemicals normally present in cells as a result of
metabolism, and also by common pollutants such as lead and
tobacco smoke.
As time passes, signaling pathways and genetic regulatory
networks within cells can be altered for the worse, due to subtle
changes in cellular chemistry. The repair mechanisms that would
normally correct such errors appear to slow down over time.
Telomeres, the ends of chromosomes, seem to get shorter
each time a cell divides, causing normally suppressed genes to
become activated and impair cell function. Finally, the brain
processes that regulate organism-wide cell behavior decline over
time, partly as a result of ongoing cell death in the brain.
255
The really frustrating thing about all these phenomena is that

none of them are terribly different from other processes that
naturally occur within cells, and which cells seem to know quite
well how to cure and repair. It would seem that cells have just
never bothered to learn how to solve these particular problems
that arise through aging, because there was never any big
evolutionary advantage to doing so. It seems that we may die not
because it would be so terribly hard to engineer immortal cells,
but because it was not evolutionarily useful to our DNA to allow
us to live forever.
Long Life Through Calorie Restriction?

Shifting from the biological mechanisms up to the whole-humanorganism level, one fascinating idea that has emerged from antiaging research is calorie restriction (CR) put simply, the idea
that if you eat about 70% of what youd ordinarily want, youll live
longer. You need to eat a healthy diet, rich in vitamins and
proteins, but low in calories.
Exactly how and why it works at the molecular level isnt yet
known but the reality of the phenomenon is beyond dispute.
CR has been tested extensively in various nonhuman mammals.
For instance, mice normally dont live over 39 months, but caloric
restriction has produced mice with 56 months lifespan. This
corresponds proportionally to a 158 year-old human. And these
long-lived mice arent old and crustytheyre oldster/youngsters,
keen-minded, strong-bodied and healthy. Studies on monkeys
are currently underway, though this naturally will take a while
due to monkeys relatively long lives.
But before you start restricting your calories too heavily, you
should know that the evidence suggests it wont be quite as
effective in humans as in mice. Roughly speaking, it seems that:
the larger the organism, the less the effect of caloric restriction.
So it increases the lifespan of nematode worms a lot, mice a fair
bit, dogs only a little, and humansin his hypothesisby just a
few years.
256
AGI Against Aging
A few things about the biological mechanisms are known. CR

increases the ability of the body to repair damaged DNA, and it
decreases the amount of oxidative (free radical) damage in the
body. It increases the levels of repair proteins that respond to
stress, it improves glucose-insulin metabolism, and for some
reason not fully understood, it delays age-related immunological
decline as well. Basically, among other possible factors, many of
the well-known mechanisms of senescence set in more slowly if
the body has to process less food over its lifetime. The relation of
this line of thinking with anti-aging pharmacology has yet to be
thoroughly investigatedit may well be there are medications
that work most effectively in coordination with a caloric restriction
diet.
Among the biological actors underlying CRs impact on aging are
the sirtuins, proteins that are addressed by natural substances
like resveratrol (found in red wine, and also available now as a
nutritional supplement) and also by new drugs under
development by Sirtris, a pharma startup recently acquired by
GlaxoSmithKline for $730 million. But the sirtuins are certainly
not the whole story; our own AI-based analysis of some data
related to CR in mice has suggested a host of other biological
actors are involved.
The Paleo Lifestyle

A more palatable, and more thoroughly scientifically grounded,
alternative to the CR diet is the Paleolithic lifestyle advocated
by an increasing set of aging researchers. In essence this
involves eating the foods that were commonly eaten by humans
before the emergence of civilization, and generally trying to
emulate the physical patterns of the pre-civilized lifestyle as
closely as possible.
Evolutionary biologist Michael Rose (my colleague in some of my
recent work on the genomics of longevity) argues that the Paleo
lifestyle may be particularly effective for people over the age of
40 or so, because genes frequently have different effects at
257
different life-stages, and their effects at later ages (after the most
typical reproductive age) are generally adapted by evolution at a
much slower pace. So, our genes effects on later life-stages are
likely more hunter-gatherer-ish than our genes effects on earlier
life-stages making a Paleo approach even more beneficial for
those of us 40 or over.
Its worth noting that Michael is in this age range himself, and he
does walk the walk as well as talk the talk he keeps to the
Paleo diet and lifestyle pretty strictly, and he looks and feels
great! I myself have been keeping a sort of quasi-Paleo diet and
lifestyle, and have noticed some definite improvements in my
feeling of physical well-being also. Whereas when I tried the CR
diet for a while a few years back, I found myself feeling
pleasingly light-headed and positive, but also relatively devoid of
drive and energy and Well Sometimes annoyingly hungry!!!
Aubrey de Grey and the Seven Pillars of

Senescence
Lifestyle and diet solutions have a lot to recommend them
especially the fact that theyre something we can do right now to
help extend our lives, rather than waiting for scientific advances!
But theres a limit to what they can do. To really eliminate the
plague of involuntary death, new science and technology will be
needed though exactly what kind isnt yet entirely clear.
Aubrey de Grey, as noted above, has put forth some fairly
specific proposals in this regard. Of all the senescence
researchers in the world, no other has done as much as Aubrey
to raise the profile of the aging problem in the public eye. But as
well as spreading the good word about the desirability and
viability of radical human life extension, hes also proposed some
specific scientific ideas. We dont always agree with his
proposed solutions to particular sub-problems of aging, but we
do find him invariably rational, insightful, and intellectually
adventurous.
258
AGI Against Aging
De Greys buzzword is SENS, which stands for Strategies for

Engineered Negligible Senescencea very carefully constructed
scientific phrasing for what weve loosely been calling here antiaging research. The point of the term is that its not merely
slowing down of aging that hes afterits the reduction of
senescence to a negligible level. Hes not trying to achieve this
goal via voodoo, were trying to achieve it via engineering
mostly biological engineering, though nanoengineering is also a
possibility. De Greys website gives a very nice overview of his
ideas, together with a number of references.
As part of his effort to energize the biology research community
about SENS, de Grey launched a contest called the Methuselah
mouse prizea prize that yields money to the researcher that
produces the longest-lived mouse of species Mus musculus. In
fact there are two sub-prizes: one for longevity, and a
rejuvenation prize, given to the best life-extension therapy
thats applicable to an already partially-aged mouse. There is a
complicated prize structure, wherein each researcher who
produces the longest-lived mouse ever or the best-ever mouselifespan rejuvenation therapy receives a bit of money each week
until his record is broken.
De Greys idea is that, during the next decade or so, it should be
possible to come pretty close to defeating senescence within
miceif the research community puts enough focus on the area.
Then, porting the results from mouse to human shouldnt take all
that much longer (biological research is regularly ported from
mice to humans, as they are an unusually suitable testbed for
human therapiesthough obviously far from a perfect match). Of
course, some techniques will port more easily than others, and
unforeseen difficulties may arise. However, if we manage to
extend human lives by 30 or 40 years via partly solving the
problem of aging, then well have 30 or 40 extra years in which to
help biologists solve the other problems.
Theory-wise, de Grey agrees with the perspective weve given
abovehe doesnt believe theres one grand root cause of
259
senescence, but rather that its the result of a whole bunch of

different things going wrong, mainly because human DNA did not
evolve in such a way as to make them not go wrong. On his
website he gives the table shown above of the seven causes of
senescence, showing for each one the date that the connection
between this phenomenon and senescence first become wellknown to biologistsand also showing, for each one, the
biological mechanism that he believes will be helpful for
eliminating that particular cause.
Seven basic causesare these really all there is? De Grey
opines that, The fact that we have not discovered another major
category of even potentially pathogenic damage accumulating
with age in two decades, despite so tremendous an improvement
in our analytical techniques over that period, strongly suggests
that no more are to be foundat least, none that would kill us in
a presently normal lifetime. Lets hope hes right, though of
course the possibility exists that as we live longer new effects will
be discoveredbut if weve put enough resources into the antiaging program, we should be able to combat those as well.
De Greys particular breakdown into seven causes is slightly
arbitrary in some ways, and others would do the breakdown
differently; but his attempt to impose a global order on the
panoply of aging-related biological disasters is appealing, and
reflects a lot of deep thinking and information integration.
One of these Seven Pillars of Aging is something that has
come up in my own work applying AI to analyze Parkinsons
disease, which Ill discuss below: mutant mitochondria. A deeper
look at this case is interesting for what it reveals about the
strength and potential weaknesses of de Greys engineeringbased approach. The term engineering in the SENS acronym is
not a coincidencede Grey came to biology from computer
science, and he tends to take a different approach from
conventional biologists, thinking more in terms of mechanical
repair solutions. Whether his approach will prove the best or not
remains to be seen. Were not biologists enough to have a
260
AGI Against Aging
strong general intuition on this point, but its often a good bet to
say that variety amongst approaches, not a single orthodoxy, will
yield the best results. The mainstream molecular biology
community seems to think de Greys proposed solutions to his
seven problems reveal a strange taste; but this doesnt mean
very much, as the mainstreams scientific taste may well be
mortally flawed. Science, like any human endeavor, has its
fashions and trends. What is seen as weird science today may
be a commonplace field of study in a decade.
Regarding mitochondrial DNA damage, de Greys current
proposal is to fix it in a rather direct way, by replacing the flawed
proteins produced by the flawed mitochondrial DNA. This could
work because there is already an in-built biological mechanism
that carries proteins into mitochondria: the TIM/TOM complex,
which carries about 1000 different proteins produced from
nuclear DNA into the mitochondria.
What de Grey proposes is to make copies of the 13 proteincoding genes in the mitochondrial genome, with a few simple
modifications to make them amenable to the TIM/TOM
mechanism, and then insert them into the nuclear chromosomes.
In this way theyll get damaged much more slowly, because the
nuclear chromosomes are much more protected from mutations
than mitochondrial genes.
Sensible enough, no?
On the other hand, I recall a conversation over dinner a few
years back, somewhere in northern Virginia, between Aubrey
and Rafal Smigrodzki, the biologist who got me involved in
Parkinsons research when he was at University of Virginia.
Rafals worry about moving mitochondrial DNA into the nucleus
is that its ordinary operations may depend on other things that
are happening outside the nucleus, so maybe after its moved it
wont be able to do its thing properly. In other words, maybe
Aubreys engineering approach overlooks too much of the
261
complexity of the biological networks involved. So far as I can

tell, the jury is still out on this.
Rafal, however, has his own angle on the problem he left UVA
to work for
GENCIA Corporation, a Charlottesville firm
developing mitochondrial gene replacement therapy, via a
novel technique called protofection. Protofection allows the
removal of bad fragments of mitochondrial DNA, and their
replacement by good ones. If this could be successfully done in
living human brains and if mitochondrial DNA damage is
indeed a major cause of Parkinsons symptoms then potentially
Parkinsons could be cured via gene therapy that simply repairs
the flawed regions of mitochondrial DNA.
Whether de Greys approach or protofection or something
totally different is the best approach, no one truly knows right
now. Unfortunately, neither currently proposed approach is being
amply funded at the moment. We cant know for sure unless the
research is done, but the research cant be done without funding.
Which is why de Greys publicity and education efforts are so
extremely valuable.
Similarly, each of de Greys other six categories of aging-related
damage is amenable to a number of different approaches we
just need to do the experiments and see which one works better.
Radical Life Extension: Whats the Real

Bottleneck?
Aging is comprehensible but complex so many different
aspects, each addressable by so many human methods. The
human mind boggles when confronted with so many complex,
interlocking networks. And this, in fact, is a major problem and
one of the reasons why aging hasnt been cured yet. Biologists
have gathered a lot of data -- but the human brain ultimately was
not evolved for the integrative analysis of a massive number of
complexly-interrelated, high-dimensional biological datasets. We
desperately try to cast biological data in a form our human brains
262
AGI Against Aging
can understand effectively: We create data visualizations to ease

the application of the 30% of our brain that is customized for
visual processing; we develop vocabularies and ontologies to
better apply the large portion of our brain customized for
linguistics. But there is no portion of the brain customized for
generating hypotheses by analyzing biological data. At this
stage, the weakest link in the biomedical research pipeline is our
human brains lack of ability to holistically understand the mass
of data that has been (and is, every day, being) collected, and
use this understanding to design new experiments leading to
new understanding. Its a somewhat radical assertion, but we
contend that the human brain is the primary bottleneck along our
path to radical life extension.
There are three evident solutions to this problem: improve the
human brain, augment it with external tools, or replace it with
something better.
The former is an exciting possibility, which will surely be possible
at some point, but neuroscience and neuroengineering are
currently a long way from enabling robust human cognitive
enhancement. Furthermore, advancing neuroengineering is
largely a biology problem which means that a major bottleneck
along the path to its achievement is precisely the problem were
talking about, the limitations of the human brain at grappling with
masses of biological data.
External tools for biological data analysis are critical and
fortunately they are now plentiful, but it is increasingly clear that
the sorts of tools we have created are not sufficient to allow us to
grapple with the patterns in the data weve collected.
Contemporary bioinformatics analysis and visualization software
represents a noble, yet ultimately inadequate attempt to work
around the shortcomings of the human brain.
To see this, consider the commonplace observation that most
geneticists focus their research on a handful of genes, or at very
most a handful of biological pathways. This cognitive strategy on
263
the part of researchers makes sense because the human brain

can handle only so much information. There are some genes,
p53 for example, about which so much information is known that
very few human scientists today have it all in their heads. On the
other hand, its also well known that the human body is a highly
complex system whose dynamics are dominated by subtle
nonlinear interactions between different genes and pathways. So
the correct way to analyze biological data is not to focus on
individual genes and pathways, but to take more of a holistic,
systems-biology approach.
Can software tools help with this? It turns out the answer is yes
but only to a limited extent. While not as commonly utilized as
they should be, there do exist statistical and machine learning
approaches to analyzing biological data, which take a holistic
approach and extract global patterns from huge datasets.
Unfortunately, though, these software programs only go so far;
they produce results that still need to be interpreted by human
biologists, whose expertise is invariably limited in scope, due to
the limitations of human memory.
Visualization tools help a lot here as well, but also have fairly
strict limitations: the human eye can only take in so much
information at one time. It evolved for scanning the African veldt,
not the intricacies of biomolecular systems. Even if you had a
holographic simulation of some portion of the human body at
some particular scale, this still wouldnt allow the human
perceptual system to see the whole, to grasp all the
mathematically evident patterns in the data being visualized.
From a scientific perspective, it would be ideal if we could simply
replace human biologists with AI systems customized for
biological data analysis systems with the human capability for
insight and interpretation (or even more so), but more memory
and more capability for quantitative precision, and patternanalysis capability tuned for biological data rather than
recognizing predators on the veldt. Unfortunately, this kind of AI
scientist does not exist at present. There are serious research
264
AGI Against Aging
programs underway with the aim of producing this kind of

software; and an increasing confidence in the AI field that this is
43
indeed an achievable goal . But life extension research, and
biological research in general, cannot afford to wait for computer
scientists to produce powerful AI theres too much urgency
about moving ahead with solving medical problems causing
human suffering, right now. So what we need is a combination of
narrow-AI applied to biomedical problems, and development of
advanced AGI with a goal toward eventually doing a radically
better job.
AI Against Aging: Tales from the Research Front

Having sketched a general picture of the state of longevity
biology these days, Ill now get a bit more personal, and discuss
some of the specific work Ive done applying narrow AI not AGI
yet to help unravel the biology of aging. As well as uncovering
some intriguing aspects of longevity biology, this work has
helped me to much more fully understand exactly how AGI would
be able to help us figure out how to extend life.
The application of narrow AI technology to biological data in
general and even aging-related data in particular is a
relatively mainstream pursuit now, being carried on by a host of
researchers at various universities, companies, and government
labs. My own work with my colleagues at Biomind LLC
constitutes only a small part of this overall exciting research
thrust. But it will give you a concrete flavor for what the field is
like these days just before the dawn of the age of AGI.
43
Goertzel BN, Coelho LS, Pennachin C, Goertzel I, Queiroz M,

Prosdocimi F, Lobo F. Learning Comprehensible Classification Rules
from Gene Expression Data Using Genetic Programming and Biological
Ontologies. In: 7th International FLINS Conference on Applied Artificial
Intelligence: 2006; Genova, Italy; 2006.
265
Overall, my work in narrow AI based biomedical informatics so

far has focused on a sort of midway point between the
approaches of better tools and replace the humans. As you
already know if youve been paying any attention to the earlier
parts of this book, I believe that the creation of a powerful AI
scientist is a real possibility within the next few decades, maybe
even within the next decade and Im spending a lot of my time
working specifically toward this goal. But, via my role in the
bioinformatics firm Biomind LLC, which is working with the NIH
and has worked with the CDC and various academic biomedical
labs, Im also acutely aware that there is important biomedical
data being generated right now about important human
problems, and weve got to deal with it as best we can. So the
approach my colleagues and I are taking in the bio space is an
incremental one: as our ambitious AI scientist is gradually
created (and its a long-term research project, as one might
expect), we are utilizing the various modules of the overall AI
system to analyze biological datasets. Of course, the AI modules
are not as powerful as a full-scale AI scientist would be, but our
experience has shown that they can still provide insights beyond
what human scientists can achieve unaided, or using
conventional tools. In this way AI and biomedical science can
progress together: The more progress we make toward the AI
scientist, the more powerful the insights generated by the partial
versions of the system.
AI Uncovers the Role of Mitochondrial DNA in Parkinsons and
Alzheimers Disease
One of the most exciting chapters so far in my exploration of the
application of AI to bioscience, involved work we did in 2005
analyzing data regarding the genetic roots of Parkinsons
disease. In this case, the result of the AI analysis was a powerful
statistical validation of the hypothesis that Parkinsons is caused
by mitochondrial mutations. These results seem reasonably
likely to lead to a practical diagnostic test for Parkinsons, and if
266
AGI Against Aging

44
the work being done at Gencia on protofection works out, they

may ultimately form the foundation of a mitochondrial gene
therapy based cure.
Over a million Americans have Parkinsons disease. Yet in spite
of years of effort by medical researchers, tracking down the
genetic roots of the disorder has proved devilishly difficult. The
DNA one usually hears about lies in the nucleus of a cell, the
cells center. In many cases the genetic roots of disease can be
traced down to mutations in the nuclear DNA, called SNPs or
Single-Nucleotide Polymorphisms. Biomind had a significant
success with this sort of analysis when analyzing SNP data
regarding Chronic Fatigue Syndrome: the AI was able to tease
out patterns of mutational combination that provided the first real
evidence that CFS is at least partially a genetically-founded
45
disease . While this sort of approach has not proved workable
for Parkinsons, a variation proved dramatically successful.
Mitochondria, the cell's energy-producing engines, also contain a
small amount of DNA. What the AI has told us is that the right
place to look for the genetic roots of Parkinsons is in the
mutations in the mitochondrial DNA. Our software identified a
particular region of a particular gene on the mitochondrial
genome that appears to be strongly associated with Parkinsons
46
disease .
Much smaller, lesser known and lesser studied than its nuclear
cousin, the mitochondrial genome is nonetheless vital to cellular
44
Smigrodzki RM, Khan SM. Mitochondrial microheteroplasmy and a

theory of aging and age-related disease. Rejuvenation Res. 2005
Fall;8(3):172-98.
45
Goertzel BN, Pennachin C, de Souza Coelho L, Maloney EM, Jones
JF, Gurbaxani B. Allostatic load is associated with symptoms in chronic
fatigue syndrome patients. Pharmacogenomics 2006; 7:485-494.
46
Smigrodzki R, Goertzel B, Pennachin C, Coelho L, Prosdocimi F,

Parker WD, Jr. Genetic algorithm for analysis of mutations in
Parkinson's disease. Artificial Intelligence in Medicine 2005; 35:227241.
267
function in humans and other animals. The human mitochondrial

genome only contains seven genes, whereas the nuclear
genome contains around 30,000 at last count. But these seven
genes carry out a lot of valuable functions. If they stop working
properly, serious problems can arise. In 1999, Dr. Davis Parker,
together with Russell H. Swerdlow and scientists from San Diego
firm MitoKors published work suggesting that defects in the
mitochondrial genome may be correlated with Parkinsons
disease. As a babys mitochondrial DNA comes entirely from its
mother, these results suggest that Parkinsons may be passed
maternally but that its defects can skip generations, making the
emergence of the disease appear random.
The work Parker and Swerdlows team did involved clever
manipulations of embryonic human nerve cells. They removed
the mitochondrial DNA from the embryonic nerve cells and
replaced it with other DNA: Sometimes from healthy people and
sometimes from Parkinsons patients. What resulted was the
nerve cells receiving the mitochondrial DNA from Parkinsons
patients began acting like nerve cells on MPTP. Low complex I
activity, meaning insufficient energy obtained from mitochondria
and eventually leading to Parkinsons-like sluggishness.
These results were fascinating and suggestive but where were
the actual mutations? All this showed was that the problem lay
somewhere in the mitochondrial genome. The question was
where. Which mutations caused the problem?
To answer this question, Parker and colleagues sequenced
mitochondrial DNA drawn from the nerve cells of a number of
Parkinsons patients, as well as a number of normal individuals,
and looked for patterns. But to their surprise, when in 2003 they
set about seriously analyzing this data, they found no simple,
consistent pattern. There were no specific genetic mutations
common to the Parkinsons patients that were not common to
samples taken from healthy subjects.
268
AGI Against Aging
Enter artificial intelligence. Dr. Rafal Smigrodzki, one of Parkers

collaborators, was familiar with my AI research work and
suggested that perhaps my AI technology might be able to find
the patterns in the mitochondrial DNA data.
To make a long story short, it worked. Appropriately enough, the
solution turned out to be an AI software technique called genetic
algorithms, which simulates the process of evolution by natural
selection beginning with a population of random solutions to a
problem, then gradually evolving better solutions via letting the
fittest solutions combine with each other to form new ones, and
making small mutations to the fittest solutions. In this case,
what the software was evolving was potential patterns
distinguishing Parkinsons patients from healthy subjects based
on the sequences of amino acids in their mitochondrial DNA.
This kind of data analysis is highly exploratory and is never
guaranteed to yield a solution but in this case things worked
out happily, and a variety of different data patterns were
discovered.
The trick, it turns out, is that while there are no specific mutations
corresponding to Parkinsons disease, there are regions and
combinations of regions -- of the mitochondrial genome that tend
to be mutated in Parkinsons patients. There are many different
rules of the form If there are mutations in this region of this
mitochondrial gene and that region of that mitochondrial gene,
then the person probably has Parkinsons disease. While it took
some advanced AI technology to find these patterns, once
discovered, the patterns are very easy for humans to
understand. The patterns were validated by subsequent
47
biological analysis on additional patients .
47
Smigrodzki R, Goertzel B, Pennachin C, Coelho L, Prosdocimi F,

Parker WD, Jr. Genetic algorithm for analysis of mutations in
Parkinson's disease. Artificial Intelligence in Medicine 2005; 35:227241.
269
Yet more excitingly, weve done further work (to be published

shortly) with Dr. Parker on comparable data regarding
Alzheimers disease, showing patterns that are similar in nature
but different in detail. Once again, although the crucial idea to
look at the mitochondria in the first place was provided by human
biological intuition, the human brain was unable to detect the
relevant patterns in the mitochondrial mutation data, even
augmented with cutting-edge statistical tools. But AI found the
relevant patterns, which are then easily validated via further
biological experiments.
AI Helps Unravel the Genetic Mechanisms Underlying the
Efficacy of Calorie Restriction for Life Extension
As well as helping to understand and diagnose (and, ultimately,
cure) aging-related diseases like Parkinsons and Alzheimers, AI
technology can help us better understand, refine and design
methods for extending the maximum lifespan of organisms. One
recent example of this is a study my colleagues and I recently
48
published in Rejuvenation Research , pertaining to the genetic
mechanisms underlying the impact of calorie restriction diets on
maximum lifespan. The exact mechanism by which calorie
restriction works remains incompletely understood (though there
are plenty of theories!), but our AI-based analysis revealed a
central role for several genes whose involvement in CRs
efficacy was not previously known. These results suggest a
number of specific biological experiments, and we are in
discussions with biology research labs regarding the best way to
carry out these experiments. These experiments of course will
produce new data to be analyzed via AI algorithms, and will likely
provide information on how various elements of the many
existing theories of CRs efficacy combine to provide the true
explanation. Through this sort of iterative interaction between AI
48
Goertzel B, Coelho L, Mudado M. Identifying the Genes and Genetic

Interrelationships Underlying the Impact of Calorie Restriction on
Maximum Lifespan: An Artificial Intelligence Based Approach, to
appear in Rejuvenation Research.
270
AGI Against Aging
analysis, human judgment and laboratory experiments, we can

progress much faster than would be possible without AI in the
picture.
In our application of AI to CR, we initially fed our AI system three
datasets that other researchers had posted online, based on
their work studying mice on calorie restriction diets. We then
merged these three datasets into a single composite dataset for
the purpose of conducting a broader-based analysis, using AI
technology rather than the standard statistical methods that the
researchers had originally used on their datasets.
Along with providing a large amount of other information, this
analysis resulted in a list of genes that the AI found to be
important for CRs impact on lifespan. An essential point here is
that the AI was capable of teasing out nonlinear interactions
between different genes and gene products. The genes that the
AI points out as important for CR and its impact on aging are
important, not necessarily in terms of their individual actions, but
most often largely in terms of their interactions with other genes.
The AI also provided a map of gene interrelationships (shown in
Figure 1), suggesting which inter-gene interactions are most
important for the effect of CR on life extension. In particular, our
graphical analysis revealed that the genes Mrpl12, Uqcrh and
Snip1 play central roles in the effects of CR on life extension,
interacting with many other genes (which the analysis
enumerates) in carrying out their roles. This is the first time that
the genes Snip1 and Mrpl12 have been identified as important in
the aging context.
To double-check the validity of these results we obtained from
analyzing three datasets at once, we then ran the same AI
processes all over again, but throwing a fourth dataset into the
mix. Much to our relief the results were largely the same
suggesting that the AI is producing real biological insights, not
just some kind of data processing artifacts.
271
Broadly, the biological interpretation of these analytical results

suggests that the effects of CR on life extension are due to
multiple factors, including factors identified in prior theories of
49
50
51
aging, such as the hormesis , development , cellular and free
52
radical theories. None of these individual theories stands out as
obviously correct, based on the patterns of gene-combination
effects identified by the AI system. But genes with predicted
involvement according to many of these theories play a role,
along with other genes not highlighted by any prior theories or
experiments.
49
Sinclair DA. Toward a unified theory of caloric restriction and

longevity regulation. Mech Ageing Dev 2005; 126:987-1002.
50
de Magalhaes JP, Church GM. Genomes optimize reproduction:
aging as a consequence of the developmental program. Physiology
(Bethesda) 2005; 20:252-259.
51
Shay JW, Wright WE. Hallmarks of telomeres in ageing research. J

Pathol 2007; 211:114-123.
52
Harman D. Aging: a theory based on free radical and radiation

chemistry. J Gerontol 1956; 11:298-300.
272
AGI Against Aging
273
Unraveling the Mystery of the Methuselah Flies

Perhaps my most exciting current research project involving AI
and biology data has to do with Michael Roses Methuselah flies,
mentioned above: Fruit flies that have been bred by directed
evolution, over the last 30 years, to live 5x longer than ordinary
fruit flies. Simply by setting up a situation where longer-lived
flies are more likely to breed with each other, and letting it
operate for many, many generations, a new strain of flies was
created. This is a miracle, and a puzzle: because these
Methuselah flies were created via directed evolution rather than
genetic engineering or other direct methods, we dont know
how what it is that makes them live so long. Now we need to
reverse engineer what directed evolution did, and understand
what combination of genetic mutations occurred to create the
long-lived flies, and why these mutations had the impact they
did. This is not a completely simple matter, because evolution is
messy: the Methuselah flies are bound to have a lot of
inessential differences from regular flies along with the
functionally critical ones, and the inessential and critical ones are
going to be complexly bound up with each other. Traditional
statistical analysis methods can identify some genes that are
important to understanding the difference between the
Methuselah flies and ordinary flies, but, they cant unravel the
genomic, proteomic and metabolomic interrelationships.
Even without a full understanding of what keeps them ticking,
analysis of the Methuselah flies has borne some fruit (sorry!).
Genescient, the company that now has rights to the IP of the
Methuselah flies, has used the Methuselah flies to find some
substances that can be fed to normal flies to make them live
much longer than usual. Furthermore, this research has led to
insights regarding nutraceuticals for promoting longevity in
humans. But these results are minor compared to what could be
achieved if the essential cause of the Methuselah flies longevity
were understood. Not all biologists agree that understanding
aging in fruit flies will help us understand human aging, but
theres a strong argument to be made. To the extent that aging is
274
AGI Against Aging
a basic property of cellular function, it is likely to be the same

process across many organisms and indeed, Genescient has
done studies showing that the genes most significant in
characterizing the Methuselah flies tend to be ones that also
relate to human diseases.
Aubrey de Greys engineering approach to combating aging
focuses on the symptoms of aging, which occur at various levels
throughout the body (as a single example, he proposes to use
certain bacteria to clean up gunk that appears between the
bodys cells increasingly with age). While this approach may
have great value, theres also something to be said for trying to
fix the basic cellular processes underlying aging. Perhaps if
these processes are fixed then many of the symptoms will
disappear on their own. Ultimately, de Greys approach and
Genescients approach may lead to complementary therapies.
So far we have only applied AI to a small set of fly data, but we
have already found some interesting conclusions. The general
role of AI here is to identify which genes are important for the
Methuselah flies longevity, and how these genes combine with
each other and based on this understanding, to figure out
which pathways can be impacted with pharmaceuticals or
nutraceuticals to cause ordinary flies to live longer. AI can also
select among the various relevant genes and pathways to
estimate which ones are most likely to lead to human aging
therapies. As in our previous examples, the AI is far from
autonomous here; it is serving as a helper to human biologists
and data analysis. But there is a lot of data and the biology is
complex, so the latter can use all the help they can get!
I cant recount the details results of my work with Genesicent
here due to intellectual property concerns, but I can review the
basic sorts of things were finding. For instance, we have found
one gene that seems to be very important to fly longevity, and
that produces a certain enzyme known in humans due to its
deficiency in people with a certain monogenetic disease
involving central nervous system malfunction. Another gene
275
emerging as important is a tumor suppressor gene (and the

relation between cancer tumor suppression and aging is very
well known), which plays a role in the Methuselah flies in
combination with several particular genes related to metabolism.
None of these findings, in itself, tells you why the Methuselah
flies live so long but they point research in specific directions,
some of which would not have been conceived based on this
data without the AI-based analysis results.
AI That Reads Biological Research Articles

So far weve been discussing the use of AI to analyze
quantitative biological datasets. But theres another fact that
must also be considered, which is that the vast majority of
biomedical knowledge online right now exists only in textual
format. Most datasets arent placed online, and as big as the
biological databases are, most knowledge that could be placed
in there, actually hasnt been, either because no one has gotten
around to it, or because researchers prefer to keep their data
proprietary.
For example, at Biomind weve done a lot of work with the Gene
Ontology, which is an outstanding database that categorizes
genes by function. If you look up apoptosis in the Gene
Ontology, youll find a few dozen genes that have been
categorized as being associated with apoptosis
preprogrammed cell death. But the catch is, if you browse

through the journal literature online, youll find even more. The
Gene Ontology cant keep up. This is a tribute to the rapid pace
of biomedical research these days, but its also an indication of
one direction biomedical software has got to go in: Weve got to
write computer programs that can grab the information directly
from the texts where its been published! This is a domain of
research called Bio-NLP bio natural language processing.
Once a sufficiently powerful AI scientist is created, Bio-NLP
wont be necessary, as the AI will simply recognize all the
relevant patterns in the data directly, without need for human
276
AGI Against Aging
insight. But were not there yet. So at the present time, the best
strategy for AI data analysis is to incorporate all available
sources of information, including direct experimental data and
text humans have produced based on interpreting that data.
In 2006, I co-organized the sixth annual Bio-NLP workshop, as
part of the annual HTL-NAACL Computational Linguistics
conference. At previous Bio-NLP workshops, nearly all the work
presented had pertained to fairly simple problems, such as
recognizing gene and protein names in research papers (a task
made more difficult than it should be by the presence of multiple
naming conventions among biologists). But starting in 2006 we
saw more and more researchers creating software with the
capability to recognize relationships between biological entities,
as expressed in natural language text; and this trend has
intensified subsequently. The latest Bio-NLP software (see
53
Rzhetskys work
for an impressive example) takes in a
research paper and tells you which genes, proteins, chemical
and pathways are mentioned, and how they are proposed by the
authors to relate to each other (which genes are in which
pathways, which enzymes catalyze which reactions, which
genes upregulate which others, etc.). This is a far cry from full
understanding of the contents of research papers, but its
definitely a start.
AI Based Logical Inference Based on Information Automatically
Extracted from PubMed Abstracts
The paper I presented at Bio-NLP 2006 regarded a research
prototype called BioLiterate, which we built for the NIH Clinical
Center in 2005. What the BioLiterate prototype did was extract
relationships from various biomedical research abstracts, and try
to glue them together using logical reasoning. So, for example, if
one paper said that p38 map kinase inhibition prevents bone
53
Krauthammer M, Rzhetsky A. GENIES: a natural-language

processing system for the extraction of molecular pathways from journal
articles. Bioinformatics 2001;17(Suppl 1):S7482 (0)
277
loss, and another paper said the DLC inhibits p38, then the
software would put A and B together, deciding (using logical
reasoning) that maybe DLC prevents bone loss (the actual
sentences the AI used in these inferences, found in PubMed
abstracts, are shown in the figure above). The logical inference
was provided by the Probabilistic Logic Networks module of the
54
Novamente Cognition Engine . BioLiterate was a prototype,
rather than a robust and deployable software solution, but it
made its point: If you build a Bio-NLP system and then use the
right sort of rules to pipe its output into a computational
reasoning system, you get an automated biological hypothesis
making system.
The Holistic Biobase

The work weve discussed above, applying AI to bioinformatics,
has already led to exciting results. A growing community of other
researchers, doing similar work, is finding similar successes.
Continuing this general approach, applying AI technology to
various datasets in isolation or in small groups, there is little
doubt that an ongoing stream of comparable results can be
obtained, providing a significant and worthwhile acceleration to
the advancement of bioscience.
But, we could do a lot better. The real future of bioscience, we
suggest, lies in the simultaneous analysis of a lot more than the
four datasets we considered in our calorie restriction study. We
need to feed dozens, hundreds, thousands, tens and hundreds
of thousands of datasets simultaneously into the same AI system
along with all the biological texts online and let the AI go to
town hunting down the patterns that are concealed therein. AI
can detect far more patterns in such a data-store than the human
mind.
54
Ikle M, Goertzel B, Goertzel I, Heljakka A. Probabilistic Logic

Networks: A Comprehensive Framework for Uncertain Inference,
Springer, May 2008.
278
AGI Against Aging
Right now, the mass of available data is terrifyingly underutilized,

due to the limitations of the human brain and the corresponding
processes of the scientific community (which are adapted to the
limitations of the human brain). Human scientists analyze
individual datasets, or small collections of datasets, using brains
that evolved for solving other sorts of problems (aided by
statistical, visualization and in rare cases AI tools); and then
these humans write papers summarizing their results. Of course,
the papers written about a certain dataset ignore nearly all the
information in that dataset, focusing on the particular patterns
that the researchers noticed (which are often the ones they were
looking for in the first place, based on their prior knowledge and
biases). Then, researchers read the papers other researchers
have written, and use the conclusions in these papers to guide
the analysis of new datasets. The multiple datasets that have
been collected are brought together indirectly only via human
beings reading and writing papers, each of which contains an
extremely partial view into the data on which its based. This is a
dramatic, tragic loss of information compared to what would
happen if the datasets were actually analyzed collectively in a
serious way.
What were suggesting is the creation of a Holistic Biobase a
massive data repository containing all the biomedical information
on the Web today including quantitative data, relational data,
textual information in articles and abstracts... Everything. The
data in this repository should then be analyzed using powerful AI
systems that are able to study the data as a whole, identifying
complex patterns not amenable to direct human analysis nor
conventional statistics. These software systems will help humans
make better discoveries, and in some cases they will surely
make new discoveries on their own suggest new experiments,
propose new hypotheses, make connections that no human
could make due to our limited ability to store and analyze
information in our brains.
The Holistic Biobase should ideally be an open information
resource, so that any scientist with statistical or AI tools and a bit
279
of savvy can crunch the data in their own way. A decent, if

55
partial, model for the Holistic Biobase is Freebase , which is an
open online database containing various sorts of information of
general interest. In principle, one could just load biological
datasets into Freebase, but in practice this isnt likely to be the
best approach, for several reasons. Freebase is a traditional
relational database, which is not the most natural data structure
for AI purposes (a graph database would be preferable). And
more critically, it doesnt solve the problems of metadata
standardization and data normalization, which are perhaps the
main obstacles standing in the way of constructing the variety of
mega-bio-database Im envisioning.
If the Holistic Biobase concept sounds overambitious and
fanciful, please remember that the Human Genome Project once
sounded very much the same. A few decades ago the synthetic
organism project of Venters lab at the J. Craig Venter Institute
in Rockville, Maryland would also have sounded sciencefictionally speculative. And how many people would have labeled
the notion of a Google-scale database of online documents
implausible or insane, just one or two decades ago? Biology and
computer science both are in the midst of phases of rapid
advance, which opens up possibilities that could barely have
been conceived of before.
As a very simple example of the value the Holistic Biobase would
have, lets turn back to the calorie restriction data analysis
project mentioned above. Were excited with the results we
achieved based on our four-dataset analysis but its easy to
see how much more powerful the results could be if we had a
massive integrative data repository at our disposal. For example,
calorie restriction is connected with energy metabolism, a
connection we as humans can exploit by interpreting the results
of calorie restriction data in the context of our own knowledge
55
Freebase: an open, shared database of the worlds knowledge.

http://www.freebase.com
280
AGI Against Aging
about energy metabolism pathways. But what if we integrated

masses of raw data regarding energy metabolism in various
aging-related contexts into the analysis and looked at this data
together with the calorie restriction data? Who knows what
might turn up? Bodies are complex systems, and the effect of
calorie restriction on life extension is surely not a phenomenon
best understood in isolation. And of course there are a dozen
other pathways that should be considered along with energy
metabolism.
What kinds of AI algorithms will be able to grapple with the
Holistic Biobase in a really effective way? We dont have much
experience doing this kind of massive-scale biological data
analysis, but the experience we do have gives us significant
guidance. There have already been some commercial products
pushing in this direction for instance Silicon Genetics GeNet
database (for microarray data) and associated MetaMine
statistical datamining package. But GeNet/Metamine handles
only standard statistical methods, and applies only to microarray
data. On the other hand, the methods weve been using in
Biomind to date are more advanced analytically and are oriented
toward combined analysis of multiple types of data. However,
they have not yet been tailored for massive-scale data analysis.
Our strong suspicion is that to handle the Holistic Biobase, new
methods will be needed. Current applications of AI to
bioinformatics have focused on the application of machine
learning algorithms for pattern recognition essentially,
algorithms that look at one or more datasets and explicitly scan
them for patterns using complex algorithms. To handle larger
numbers of data and yet preserve the capability for analytical
sophistication, a paradigm shift will be required and this
paradigm shift ties in naturally with the trends of development in
the AI field itself. What is needed is the fusion of bioinformatics
data analysis with automated reasoning. More specifically:
Automated probabilistic reasoning, since biological data is
riddled with uncertainties. Automated reasoning allows an AI
system to study a handful of datasets, derive results regarding
281
the patterns in these datasets, and then extrapolate these

patterns to see what they imply about other datasets. This step
of inferential extrapolation allows far more scalable analysis than
machine-learning pattern-recognition methods alone. My own
team is currently pursuing this vision via integrating our
OpenBiomind bioinformatic AI software with our OpenCog
general-AI platform, which includes a powerful probabilistic
56
inference framework called Probabilistic Logic Networks . This
will allow us, for example, to massively extend our calorie
restriction data analysis project, to include numerous datasets
drawn from studies of different but allied biological phenomena.
While this vision goes a fair bit beyond current practice, there are
some contemporary projects with smaller but vaguely similar
ambitions. One example is a project called ImmPort this is a
program funded by the National Institute of Health, specifically
the National Institute for Allergies and Infectious Diseases, which
Biomind is involved with via a subcontract to Northrop-Grumman
IT. ImmPort is still in the making, but what its going to be, when
its finished, is a Web portal site for NIH-funded immunologists.
Biomind has played a small role in this to integrate
bioinformatics analysis technologies into the portal, both
innovative machine learning techniques and more standard
methods. The most exciting part of ImmPort is probably its
potential to enable massive data integration. When an
immunologist uploads data into ImmPort, it will automatically be
put in a standard format, so it can be automatically analyzed in
the same way as all the other datasets that were uploaded
and, most excitingly, so it can be analyzed in terms of the
patterns that emerge when you put it together with all the other
datasets. This is something thats hardly ever being done right
now the application of bioinformatic technology to look for
patterns spanning dozens or hundreds or thousands of datasets.
56
Ikle M, Goertzel B, Goertzel I, Heljakka A. Probabilistic Logic

Networks: A Comprehensive Framework for Uncertain Inference,
Springer, May 2008.
282
AGI Against Aging
However, the scope of AI currently envisioned within ImmPort is

restricted to machine-learning algorithms; extension to more
powerful automated inference methods is beyond the scope of
the project.
Projects like ImmPort are definitely a step in the right direction
but only a step. Even if every immunologist on the planet were
to upload their data into ImmPort, and even if ImmPort were to
incorporate inference-based data analysis, the restriction to
immunological data alone would still constitute a huge limitation.
The immune system is not an island, it is intricately connected
with nearly all other body systems. As an example, in its work
with the CDC, Biomind found that Chronic Fatigue Syndrome is
most likely a complex interaction between immune, endocrine,
autonomic nervous and other functions. What we need is not just
a holistic immunology database but a holistic biology database,
and with a focus on powerful cross-dataset AI analysis as well as
statistical and machine learning methods. Furthermore, there is
as yet no ImmPort analogue for data directly related to life
extension.
Our hope is that over the next decade the ideas described here
will become boring and mainstream, and the value of massive,
sophisticated, AI-based cross-dataset analysis will move from
outrageous to obvious in the consensus view. Until that time, we
will continue to be slowed down in our quest to extend human life
and cure human disease by the limitations of our human brains
at analyzing the relevant biological data.
And a post-script from 2014: The ideas in the above essay

have NOT yet become boring and mainstream. But they are not
as outrageous-sounding as they were in 2008, thats for sure.
And we are gradually making progress toward a Holistic
Biobase, via integrating various biological knowledge bases and
datasets into the OpenCog Atomspace. The vision articulated in
AI Against Aging seems to be slowly moving toward realization
283
though, as is often the case with technology Im interested in,

much more slowly than one would hope, due to funding and
attention issues.
284
17
AGI and Magical Thinking
In March 2010, while the nuclear reactor disaster in Fukushima
Japan was all over the news, a friend asked me to write him a
few paragraphs on the topic of AGI and nuclear disasters. How,
he asked me, might disasters like Fukushima be prevented if we
already had powerful artificial general intelligence?
I actually didnt think this was a great idea for an article, because
it seemed too obvious to me that problems like Fukushima could
easily be avoided if one had radically superhuman intelligence at
ones disposal (although, new and currently unforeseen
problems might well arise!). But he insisted, and said he had a
particular group of colleagues he wanted to distribute my
thoughts to So I went ahead and wrote him the paragraphs.
And having done that, I figured I might as well post the result as
a brief item in H+ Magazine.
The H+ post received a fair bit of negative attention, even to the
point of getting slammed by Forbes Magazine blogger Alex
Knapp (whom I emailed with a bit afterwards, and who turned out
to be a really nice guy with a good knowledge of science).
Overall the discussion with the articles critics was interesting
and highlighted the difficulties of getting people even highly
intelligent and educated people to really understand the power
that advanced AGI would have.
First, the little article in question
285
Could AGI Prevent

Future Nuclear Disasters?
Ben Goertzel; March 23, 2011
In the wake of a tragedy like the nuclear incidents were
currently seeing in Japan, one of the questions that rises to
the fore is: What can we do to prevent similar problems in
the future?
This question can be addressed narrowly, via analyzing
specifics of nuclear reactor design, or by simply resolving to
avoid nuclear power (a course that some Western nations
may take, but is unlikely to be taken by China or India, for
example). But the question can also be addressed more
broadly: What can we do to prevent unforeseen disasters
arising as the result of malfunctioning technology, or
unforeseen interactions between technology and the natural
or human worlds?
Its easy to advocate being more careful, but careful
attention comes with costs in both time and money, which
means that in the real world care is necessarily
compromised to avoid excessive conflict with other
practically important requirements. For instance, the
Japanese reactor designs could have been carefully
evaluated in scenarios similar to the one that has recently
occurred; but this was not done, most likely because it was
judged too unlikely a situation to be worth spending scarce
resources on.
What is really needed, to prevent being taken unawares by
freak situations like what were seeing in Japan, is a
radically lower-cost way of evaluating the likely behaviors of
our technological constructs in various situations, including
those judged plausible but unlikely (like a magnitude 9
286
Could AGI Prevent Future Nuclear Disasters?
earthquake). Due to the specialized nature of technological

constructs like nuclear reactors, however, this is a difficult
requirement to fulfill using human labor alone. It would
appear that the development of advanced artificial
intelligence, including Artificial General Intelligence (AGI)
technology, has significant potential to improve the situation.
An AI-powered artificial nuclear scientist would have been
able to take the time to simulate the behavior of Japanese
nuclear reactors in the case of large earthquakes, tidal
waves, etc. Such simulations would have very likely led to
improved reactor designs, avoiding this recent calamity plus
many other possible ones that we havent seen yet (but may
see in the future).
Of course, AGI may also be useful for palliating the results of
disasters that do occur. For instance, cleanup around a
nuclear accident area is often slowed down due to the risk of
exposing human workers to radiation. But robots can already
be designed to function in the presence of high radiation;
whats currently underdeveloped is the AI to control them.
And, most of the negative health consequences of radiation
from a nuclear accident such as the recent ones are longterm rather than immediate. Sufficiently irradiated individuals
will have increased cancer risk, for example. However, the
creation of novel therapies based on AI modeling of
biological systems and genomic data, could very plausibly
lead to therapies remedying this damage. The reason
relatively low levels of radiation can give us cancer is
because we dont understand the body well enough to
instruct it how to repair relatively minor levels of radiationincurred bodily damage. AGI systems integratively analyzing
biomedical data could change this situation in relatively short
order, once developed.
Finally, the creation of advanced intelligences with different
capabilities than the human mind, could quite plausibly lead
to new insights, such as the development of alternative
287
power sources without the same safety risks. Safe nuclear

fusion is one possibility, but there are many others; to take
just one relatively pedestrian example, perhaps intelligent
robots capable of operating easily in space would perfect
some of the existing designs for collecting solar energy from
huge solar sails.
There is no magic bullet for remedying or preventing all
disasters, but part of the current situation seems to be that
the human races ability to create complex technological
systems has outstripped its ability to simulate their behavior,
and foresee and remedy the consequences of this behavior.
As the progress of technology appears effectively
unstoppable, the most promising path forward may be to
progressively (and, hopefully, rapidly) augment the human
mind with stronger and stronger AI.
OK -- now, I admit this wasnt my strongest article ever. It was
hastily typed in a few minutes, more of a quickie blog post than a
real magazine article. But to me, its weakest point was its
obviousness. Alex Knapps Forbes critique, however, took a
quite different tack, accusing me of unscientific magical
thinking. In his own words:
You can show the reliance on magical thinking with just a
few quick word changes. For example, Im going to change
the title of the article to Could Djinn Prevent Future Nuclear
Disasters?, then make just a handful of word changes to the
paragraphs quoted:
What is really needed, to prevent being taken unawares by
freak situations like what were seeing in Japan, is a
radically lower-cost way of evaluating the likely behaviors of
our technological constructs in various situations, including
those judged plausible but unlikely (like a magnitude 9
earthquake). Due to the specialized nature of technological
constructs like nuclear reactors, however, this is a difficult
requirement to fulfill using human labor alone. It would
288
appear that finding magic lamps that hold Djinn has

significant potential to improve the situation.
A Djinn would have been able to take the time to simulate
the behavior of Japanese nuclear reactors in the case of
large earthquakes, tidal waves, etc. Such simulations would
have very likely led to improved reactor designs, avoiding
this recent calamity plus many other possible ones that we
havent seen yet (but may see in the future).
Very funny, right? Three cheers for Alexs sleight of word!
But, as I pointed out in my reply to Alex: What if someone in
1950 had forecast the Internet and everything it can do in 2011?
They could have written The Internet will help you navigate your
car; the Internet will let you select millions of books to read from
any house or mobile phone; the Internet will let you reserve
flights and monitor global weather conditions; the Internet will
help spark revolutions in Third World countries; the Internet will
allow people to get college educated from their homes.
And some clever wordsmith could have spoofed them by
replacing Internet with Djinn, and accusing them of magical
thinking.
The Djinn will help you navigate your car; the Djinn will let you
select millions of books to read from any house or mobile phone;
the Djinn t will let you reserve flights and monitor global weather
conditions; the Djinn will help spark revolutions in Third World
countries; the Djinn will allow people to get college educated
from their homes
And most people in 1950 would have laughed along knowingly,
at the fools who thought some mythical hypothetical construct
called the Internet could do all those things.
289
Of course the Internet has limitations, which would have been

hard to foresee in 1950. But nevertheless, it does a lot that would
have seemed like magic in 1950.
As Arthur C. Clarke said, Any sufficiently advanced technology
is indistinguishable from magic.
My thinking about AGI may be wrong but its not magical!
Now, I suppose Alexs main point (very loosely paraphrasing)
was that just listing amazing things AGI can do isnt very useful,
and it would be better to describe a path to actually making these
things work. And I agree with that. A detailed discussion of how
to get from current tech to robots capable of solving or
preventing nuclear disasters would have been a better article!
I never wrote that better article (though I would if I had some
evidence the folks in charge of preventing or mitigating nuclear
disasters had some propensity to listen to me), but Alexs critique
did prompt me to do a little reading on the current state of
robotics for nuclear disaster management, which I posted in a
P.S. to my original brief article.
I found a very nice article in the IEEE spectrum, describing in
detail some of the reasons why current robot technology is only
of limited use for helping with nuclear disasters. The basic
reasons come down to
290
lack of radiation shielding (a plain old engineering

problem rather than directly an AGI problem though
hypothetically an AGI scientist could solve the problem,
Im sure human scientists can do so also).
relative physical ineptitude at basic tasks like climbing

stairs and opening doors (problems to be solved by a
combination of engineering advances and intelligent
control (ultimately AGI) advances)
the need for tele-operation which is awkward when

robots are moving around inside shielded buildings and
so forth. This is purely an AGI problem the whole goal
of applying AGI to robotics is to let them operate
autonomously
There are many approaches in the research community aimed at

creating AGI robots of this sort, and if things go well one of these
may lead to robots capable of providing serious aid to nuclear
disasters within a, say, 5-25 year timeframe. As well as my own
OpenCog approach, one can point to a load of others say,
57
Juyang Wengs developmental robotics project at Michigan
58
State University, the European IM-CLEVER
(Intrinsically
Motivated Cumulative Learning Versatile Robots) project which
aims to make a robot capable of autonomously learning various
practical tasks, initially childlike ones and then growing more
sophisticated. These exemplify the kind of work that I think will
eventually lead to robots capable of fixing nuclear disasters
which of course is just the most obvious of the many ways that
AGI may help to prevent or mitigate nuclear disasters.
What looks at first glance like magical thinking, turns out in fact
to be rational extrapolation from the current work of serious
scientists at major universities. Exactly how long it will take to get
from here to there is impossible to know rigorously, and the
experts disagree but theres a big difference between well
almost surely get there with science and engineering, though we
dont know for sure when and magical thinking. Nuclear
disaster recovery robots are not like femtotech something
thats probably feasible but will require radical new science, and
might end up being impossible according to portions of physical
law that are now poorly understood. Theyre more like Drexlerstyle nanotech clearly feasible via physical law, and reasonably
seen as an extrapolation of current research, though not close
57
http://www.cse.msu.edu/~weng/
58
http://www.im-clever.eu/
291
enough to current knowledge as to succumb to straightforward

schedule estimation.
Another comment on my brief nuclear disaster piece was posted
on an email list by AI researcher Steve Richfield, who argued
that:
The disaster is the child of two incredible simple human
errors:
1. A simple toilet-filling mechanism connected to an outside
inlet for use in disasters would have provided a way to
keep the nuclear fuel submerged despite technical
failures. No one would ever install such a simple thing
because it would be an admission of engineering
weakness.
2. Spent fuel is kept on site rather than being reprocessed
for purely political reasons no one wants the
reprocessing done in their own back yard.
An AGI would be faced with these same two pressures. Only
through the exercise of very un-democratic power could it
get past them.
As I have pointed out on other forums, the obvious,
reasonable, expected things that an AGI should and would
mandate are COMPLETELY socially unacceptable to just
about everyone.
While insightful in some ways, I also found these comments a bit
amusing. First of all, even if true, these points dont obviate the
potential value of AGI repair robots. And what about the
possibility that an AGI could find politically acceptable technical
solutions, which humans havent been able to conceive given the
resources allocated to the problem? Heck, just inventing a
cheap material much stronger than those currently used, which
could then be deployed to store the waste, would be a big help. I
bet this could be done in short order by an AGI scientist with
sensors and actuators at the nano-scale, capable of
292
manipulating nano-fibers as readily as we stack blocks or tie

knots in rope.
What I saw in the various reactions to my simple, blithe little
piece on AGI and nuclear disasters, is the difficulty that even
very educated people have in coming to terms with the potential
that advanced AGI has. Viewing AGI through the lense of our
contemporary limitations, its all too easy to miss its potential
power just as (to use some far weaker phenomena as
analogies) nearly everybody failed to foresee the transformative
power of the Internet or mobile phones, even mere years before
their advent.
293
19
Plenty More Room at the Bottom
(Beyond Nanotech to Femtotech)
This H+ Magazine article emerged in 2010, as a result of
discussions I was having with my good friend Hugo de Garis.
Hugo began his scientific career as a physicist and then verged
into AI, working e.g. on novel software/hardware architectures for
evolving neural nets in FPGA chips; he has also made a name
for himself with outrageous futurist prognostications such as his
predictions of a coming Artilect War between pro and anti
technology forces. But in 2010, after retiring from his position
as an AI professor at Xiamen University in China (where he
introduced me to my now-wife Ruiting LIan, while I was visiting
him there co-running the 2009 AGI Summer School), he returned
to his first love of mathematical physics and began speculating
seriously about technologies way smaller than nano.
The
concept had occurred to me before, but he got me to take it more
seriously than I had been, and I put a bit of thought into it, and
wrote this article.
Im pretty sure that advanced AGI at the human level and
significantly beyond can be achieved using plain ordinary
digital computers For example, server farms like the kind
powering Google and Yahoo today. Even if the human brain
does make use of some funky quantum nonlocality effects or
even weirder physics (as some claim, though theres no
evidence of this currently), I really doubt such effects are
necessary for achieving humanlike (and transhuman) general
intelligence.
But, even if my view on this is right (and the vast majority of
physicists and AI researchers, even AGI skeptics, agree with my
views on this), thats not to discount the potential for achieving
yet greater intelligence by using more exotic computing fabrics.
In order to unlock the full depths of computing power implicit in
294
physical reality in a grain of sand, as Hugo de Garis likes to put

it were almost surely going to need to go beyond traditional
digital computing, and venture into quantum computing and
beyond.
Well need to delve deep into nanotech and quite possibly
beyond, into the world of femtotech, as yet almost wholly
uncharted But in no way ruled out by known physics.
Not long ago, nanotechnology was a fringe topic; now it's a
flourishing engineering field, and fairly mainstream. For example,
while writing this article, I happened to receive an email
advertisement for the Second World Conference on
59
Nanomedicine and Drug Delivery, in Kerala, India . It wasn't so
long ago that nanomedicine seemed merely a flicker in the eyes
60
of Robert Freitas and a few other visionaries !
But every metric system geek knows nano's not so small, really.
As youll recall from the last chapter, a nanometer is 109
meters the scale of atoms and molecules. A water molecule is
a bit less than one nanometer long, and a germ is around a
thousand nanometers across. On the other hand, a proton has a
diameter of a couple femtometers where a femtometer, at
1015 meters, makes a nanometer seem positively gargantuan.
Now that the viability of nanotech is widely accepted (in spite of
some ongoing heated debates about the details), it's time to ask:
What about femtotech? Picotech or other technologies at the
scales between nano and femto seem relatively uninteresting,
because we don't know any basic constituents of matter that
exist at those scales. But femtotech, based on engineering
structures from subatomic particles, makes perfect conceptual
sense, though it's certainly difficult given current technology.
59
60
http://www.nanomedicine.macromol.in/
http://www.nanomedicine.com/]
295
The nanotech field was arguably launched by Richard

Feynman's 1959 talk There's Plenty of Room at the Bottom. As
Feynman wrote there,
It is a staggeringly small world that is below. In the year
2000, when they look back at this age, they will wonder why
it was not until the year 1960 that anybody began seriously
to move in this direction.
Why cannot we write the entire 24 volumes of the
Encyclopedia Brittanica on the head of a pin?
But while Feynman's original vision was focused on the nanoscale (the head of a pin, etc.), it wasn't intrinsically restricted to
this level. There's plenty of room at the bottom, as he said and
the nano-scale is not the bottom! There's plenty more room
down there to explore.
One might argue that, since practical nanotech is still at such an
early stage, it's not quite the time to be thinking about femtotech.
But technology is advancing faster and faster each year, so it
makes sense to think a bit further ahead than contemporary
hands-on engineering efforts. Hugo de Garis has been talking to
me about femtotech for a while, and has touched on the topic in
various lectures and interviews; and he convinced me that the
topic is worth looking at in spite of our current lack of knowledge
regarding its practical realization. After all, when Feynman gave
his Plenty of Room at the Bottom lecture, nanotech also
appeared radically pie-in-the-sky.
There are many possible routes to femtotech, including some fun
ones I won't touch here at all like micro black holes and BoseEinstein condensation of squarks. I'll focus here largely on a
particular class of approaches to femtotech based on the
engineering of stable degenerate matter not because I think
this is the only interesting way to think about femtotech, but
merely because one has to choose some definite direction to
explore if one wants to go into any detail at all.
296
Physics at the Femto Scale

To understand the issues involved in creating femtotech, you'll
first need to recall a few basics about particle physics.
In the picture painted by contemporary physics, everyday objects
like houses and people and water are made of molecules, which
are made of atoms, which in turn are made of subatomic
particles. There are also various subatomic particles that don't
form parts of atoms (such as photons, the particles of light, and
many others). The behavior of these particles is extremely weird
by everyday-life standards with phenomena like non-local
correlations between distant phenomena, observer-dependence
of reality, quantum teleportation and lots of other good stuff. But I
won't take time here to review quantum mechanics and its
associated peculiarities, just to run through a few facts about
subatomic particles needed to explain how femtotech might
come about.
Subatomic particles fall into two categories: fermions and
bosons. These two categories each contain pretty diverse sets of
particles, but they're grouped together because they also have
some important commonalities.
The particles that serve as the building blocks of matter are all
fermions. Atoms are made of protons, neutrons and electrons.
Electrons are fermions; and so are quarks, which combine to
build protons and neutrons. Quarks appear to occur in nature
only in groups, most commonly groups of 2 or 3. A proton
contains two up quarks and one down quark, while a neutron
consists of one up quark and two down quarks; the quarks are
held together in the nucleus by other particles called gluons.
Mesons consist of 2 quarks a quark and an anti-quark. There
are six basic types of quark, beguilingly named Up, Down,
Bottom, Top, Strange, and Charm. Out of the four forces
currently recognized in the universe electromagnetism, gravity
and weak and strong nuclear forces quarks are most closely
associated with the strong nuclear force, which controls most of
297
their dynamics. But quarks also have some interaction with the
weak force, e.g. the weak force can cause the transmutation of
quarks into different quarks, a phenomenon that underlies some
kinds of radioactive decay such as beta decay.
Stylistic depiction of a proton, composed of two Up quarks and

one Down quark
On the other hand, bosons are also important for example

photons, the particle-physics version of light, are bosons.
Gravitons, the gravity particles proposed by certain theories of
gravitation, would also be bosons.
The nucleus of an atom contains protons and neutrons. The

electrons are arranged in multiple shells around the nucleus, due
298
to the Pauli exclusion principle. Also note, this sort of solar

system model of particles as objects orbiting other objects is just
a crude approximation.
The carbon atom, whose electrons are distributed across two

shells.
Finally, just one more piece of background knowledge before we

move on to femtotech. Fermions, unlike bosons, obey the Pauli
exclusion principle which says that no two identical fermions
can occupy the same state at the same time. For example, each
electron in an atom is characterized by a unique set of quantum
numbers (the principle quantum number which gives its energy
level, the magnetic quantum number which gives the direction of
orbital angular momentum, and the spin quantum number which
gives the direction of its spin). If not for the Pauli exclusion
principle, all of the electrons in an atom would pile up in the
lowest energy state (the K shell, the innermost shell of electrons
orbiting the nucleus of the atom). But the exclusion principle
implies that the different electrons must have different quantum
states, which results in some of the electrons getting forced to
have different positions, leading to the formation of additional
shells (in atoms with sufficient electrons).
299
Degenerate Matter as a Possible Substrate for

Femtotech
One can view the Pauli exclusion principle as exerting a sort of
pressure on matter, which in some cases serves to push
particles apart. In ordinary matter this Pauli pressure is minimal
compared to other forces. But there is also degenerate matter
matter which is so extremely dense that this Pauli pressure or
degeneracy pressure, preventing the constituent particles from
occupying identical quantum states, plays a major role. In this
situation, pushing two particles close together means that they
have effectively identical positions, which means that in order to
obey the Pauli exclusion principle, they need to have different
energy levels creating a lot of additional compression force,
and causing some very odd states of matter to arise.
For instance, in ordinary matter, temperature is correlated to
speed of molecular motion. Heat implies faster motion, and
cooling something down makes its component molecules move
more slowly. But in degenerate matter, this need not be the
case. If one repeatedly cools and compresses a plasma,
eventually one reaches a state where it's not possible to
compress the plasma any further, because of the exclusion
principle that won't let us put two particles in the same state
(including the same place). In this kind of super-compressed
plasma, the position of a particle is rather precisely defined but
according to a key principle of quantum theory, Heisenberg's
uncertainty principle, you can't have accurate knowledge of both
the position and the momentum (movement) of a particle at the
same time. So the particles in a super-compressed plasma must
therefore have highly uncertain momentum i.e. in effect, they're
moving around a lot, even though they may still be very cold.
This is just one example of how degenerate matter can violate
our usual understanding of how materials work.
At the present time, degenerate matter is mostly discussed in
astrophysics, in the context of neutron stars, white dwarf stars,
and so forth. It has also been popular in science fiction for
300
example, in the Star Trek universe, neutronium (matter formed

only from large numbers of neutrons, stable at ordinary gravities)
is an extremely hard and durable substance, often used as
armor, which conventional weapons cannot penetrate or dent at
all. But so far neutronium has never been seen in reality.
Strange matter defined as matter consisting of an equal
number of up, down and strange quarks is another kind of
degenerate matter, with potential applications to femtotech,
which I'll return to a little later.
As a substrate for femtotech, degenerate matter appears to have
profound potential. It serves as an existence proof that, yes, one
can build stuff other than atoms and molecules with subatomic
particles. On the other hand, there is the problematic fact that all
the currently known examples of degenerate matter exist at
extremely high gravities, and derive their stability from this
extreme gravitational force. Nobody knows, right now, how to
make degenerate matter that remains stable at Earth-level
gravities or anywhere near. However, neither has anybody
shown that this type of degenerate matter is an impossibility
according to our currently assumed physical laws. It remains a
very interesting open question.
Bolonkin's Fantastic Femtotech Designs

If you type femtotech into a search engine, you'll likely come up
with a 2009 paper by A.A. Bolonkin, a former Soviet physicist
now living in Brooklyn, entitled Femtotechnology: Nuclear
61
Matter with Fantastic Properties . Equations and calculations
notwithstanding, this is an explicitly speculative paper but the
vision it presents is intriguing.
Bolonkin describes a new (and yet unobserved) type of matter
he calls AB-matter, defined as matter which exists at ordinary
61
http://www.scipub.org/fulltext/ajeas/ajeas22501514.pdf
301
Earthlike gravities, yet whose dynamics are largely guided by

Pauli exclusion principle based degeneracy pressure. He
explores the potential of creating threads, bars, rods, tubes, nets
and so forth using AB-matter. He argues that
this new AB-Matter has extraordinary properties (for
example, tensile strength, stiffness, hardness, critical
temperature, superconductivity, supertransparency and zero
friction.), which are up to millions of times better than
corresponding properties of conventional molecular matter.
He shows concepts of design for aircraft, ships,
transportation, thermonuclear reactors, constructions and so
on from nuclear matter. These vehicles will have
unbelievable possibilities (e.g., invisibility, ghost-like
penetration through any walls and armor, protection from
nuclear bomb explosions and any radiation flux).
All this sounds exciting indeed! And the parallels between
Bolonkin's diagrams and Drexler's diagrams in Nanosystems are
obvious. But nowhere in Bolonkin's fascinating thought-piece
does he address the million-dollar question of how and why he
thinks such structures could be made stable.
I discussed this with Steve Omohundro, a fellow AI researcher
and futurist thinker who started his career as a physicist, and
Steve very articulately expressed the same common sense
nuclear physics worries I experienced on reading Bolonkin's
paper:
A standard model for a nucleus is the "liquid drop" model
and it gives pretty good predictions. Basically it treats the
nucleus as a liquid with a pretty high surface tension. The
nucleons in the center are energetically very happy because
they are surrounded by other nucleons attracted by the
strong interaction. The nucleons on the surface are not so
energetically happy because they interact with fewer other
nucleons than they might otherwise. This creates a high
effective "surface tension" for the nuclear liquid. That's what
302
makes nuclei want to be spherical. And when they get too

big they become unstable because the surface area is
relatively larger and electrostatic repulsion overcomes the
nuclear
attraction.
All of Bolonkin's proposed femtostructures seem unstable to
me. His femto rods or whiskers are like streams of water
which are subject to instabilities that cause them to break
into a sequence of droplets. Imagine one of his rods
periodically squeezing inward and outward keeping the
volume fixed. If the surface area is decreased the
perturbation will be increased and eventually break the rod
into droplets.
Even if they weren't subject to that instability, there would be
tremendous tensile force trying to pull the two ends of a rod
together and turning it into a ball (which has a smaller
surface area than the same volume cylinder). I didn't see any
suggestions for what he wants to use to counteract that
tensile force.
Like me, Steve has a tendency to be open-minded about wildsounding future possibilities. But open-mindedness must be
tempered with a bit of realism.
I'm hoping for a sequel from Bolonkin containing at least back-ofthe-envelope stability arguments. But Im afraid it may not be
forthcoming. After a bunch of hunting, in early 2011 I finally
managed to get in touch with Alexander Bolonkin and have a
brief conversation about femtotech. He turned out to be a really
nice guy, and an old friend of my late great friend Valentin
Turchin (the Russian dissident, physicist, computer scientist and
systems theorist with whom I collaborated on supercompilation in
1999-2001, and to whose memory I dedicated my book A
Cosmist Manifesto). Like Turchin, Bolonkin had been a Soviet
dissident, and he had arrived in New York sometime after Val
Turchin, so that Val had played the role of helping him settle into
America. He showed me a host of his futurist writings, which
echoed many ideas current in the transhumanist community.
303
Similar to Val Turchin, he had been thinking and writing about

these topics (superintelligent robots, cryonics, nanotech,
femtotech, immortality on Earth, uploading, etc. etc.) since the
1960s. Not that many modern American techno-futurists realize
that the Soviet scientific community got there way before the
West! And not only in terms of science-fictional speculation, but
also in terms of deep scientific and engineering theory like
femtotechnology!
About femtotech, Bolonkin kindly shared with me some of his
thoughts, such as:
If we conventionally join the carbon atom to another carbon
atom a lot of times, we then get the conventional piece of a
coil. If we joint the carbon atom to another carbon atom by
the indicated special methods, we then get the very strong
single-wall nanotubes, graphene nano-ribbon (super-thin
film), armchair, zigzag, chiral, fullerite, torus, nanobud and
other forms of nano-materials. That outcome becomes
possible because the atomic force (van der Waals force,
named for the Dutch physicist Johannes Diderik van der
Waals, 1837-1923, etc.) is NON-SPHERICAL and active in
the short (one molecule) distance. The nucleon nuclear force
also is NON-SPHERICAL and they may also be active about
the one nucleon diameter distance. That means we may
also produce with them the strings, tubes, films, nets and
other geometrical constructions.
You can get a stable AB-matter if you put the nucleus in line
and calculate the electrostatic (long distance) force which
creates the tensile stress and keep this line as string.
It was evident hed done a lot of back-of-the-envelope
calculations beyond what hed written in his one brief paper on
the topic and I also got the feeling he wasnt feeling any great
urgency to write up the details in any formal way, for others to
peruse and continue his work. I can well understand this, since I
myself have probably written up only 25% or so of the interesting
304
technical ideas Ive worked out in various scribblings and notes

here and there. Writing stuff up just isnt nearly as fun as figuring
it out in the first place, and often takes more time time that
could be spent figuring out more and more cool stuff! It would be
great for some young transhumanist physicist to spend a few
months carefully picking his brain on the topic, and maybe write
a joint opus on the theory of femtotech. If I werent so busy
pushing toward AGI, Id be sorely tempted to spend a few
months that way myself!
Might Dynamic Stabilization Work on Degenerate

Matter?
And in spite of his initial skeptical reaction, after a little more
thought Steve had a rather interesting brainstorm:
I just had a thought about how to stabilize degenerate
femtomatter: use dynamic stabilization. The classic example
is the shaking inverted pendulum. An upside down pendulum
is unstable, falling either left or right if perturbed. But if you
shake the base at a sufficiently high frequency, it adds a
"pondermotive"
pseudopotential which stabilizes the unstable fixed point.
62
Here's a video of a guy who built one
The same approach can stabilize fluid instabilities. If you turn
a cup of fluid upside down, the perfectly flat surface is an
unstable equilibrium. The Rayleigh-Taylor instability causes
ripples to grow and the fluid to spill out. But, I remember
seeing a guy years ago who put a cup of oil in a shaking
apparatus and was able to turn it upside down without it
spilling. So the oscillations were able to stabilize all the fluid
modes at once. I wonder if something similar might be used
to stabilize degenerate matter at the femto scale?
62
http://www.youtube.com/watch?v=rwGAzy0noU0
305
A fascinating idea indeed! Instead of massive gravity or massive

heat, perhaps one could use incredibly fast, low-amplitude
vibrations to stabilize degenerate matter. How to vibrate
subatomic particles that fast is a whole other matter, and surely a
difficult engineering problem but still, this seems a quite
promising avenue. It would be interesting to do some
mathematics regarding the potential dynamic stabilization of
various configurations of subatomic particles subjected to
appropriate vibrations.
An inverted pendulum kept vertical via dynamic stabilization. The

rod would rotate and fall down to one side or another if it weren't
vibrating. But if it's vibrated very fast with low amplitude, it will
remain upright due to dynamic stabilization. Con Conceivably a
similar phenomenon could be used to make stable degenerate
matter, using very fast femtoscale vibrations.
Of course, such metaphorical ideas must be taken with a grain of

salt. When I think about the liquid drop model of the nucleus,
I'm somewhat reminded of how the genius inventor Nikola Tesla
intuitively modeled electricity as a fluid. This got him a long way
compared to his contemporaries, leading him to develop AC
power and ball lightning generators and all sorts of other
amazing stuff yet it also led to some mistakes, and caused him
to miss some things that are implicit in the mathematics of
electromagnetism but not in the intuitive metaphorical "electricity
as fluid" model. For instance Tesla's approach to wireless power
306
transmission was clearly misguided in some respects (even if it

did contain some insights that haven't yet been fully
appreciated), and this may have been largely because of the
limitations of his preferred fluid-dynamics metaphor for electricity.
Where degenerate matter is concerned, metaphors to liquid
drops and macroscopic shaking apparatuses may be very helpful
for inspiring additional experiments, but eventually we can
expect rigorous theory to far outgrow them.
The bottom line is, in the current state of physics, nobody can
analytically solve the equations of nuclear physics except in
special simplified cases. Physicists often rely on large-scale
computer simulations to solve the equations in additional cases
but these depend on various technical simplifying assumptions,
which are sometimes tuned based on conceptual assumptions
about how the physics works. Intuitive models like nucleus as
water droplet are based on the limited set of cases in which
we've explored the solutions of the relevant equations using
analytical calculations or computer simulations. So, based on the
current state of the physics literature, we really don't know if it's
possible to build stable structures of the sort Bolonkin envisions.
But there are surely worthwhile avenues to explore, including
Steve's intriguing suggestion.
Gell-Mann Gives Femtotech A Definite Maybe

A few weeks ago, while at an event in San Francisco, I was
thrilled to have the opportunity to discuss femtotech with Murray
Gell-Mann who is not only a Nobel Prize winning physicist, but
also one of the world's ultimate gurus on quarks, since he
invented and named the concept and worked out a lot of the
theory of their behavior. I knew my friend Hugo de Garis had
briefly discussed femtotech with Gell-Mann a decade and a half
previously, but that he hadn't expressed any particular thoughts
on the topic. I was curious if Gell-Mann's views on the topic had
perhaps progressed a bit.
307
To my mild disappointment, Gell-Mann's first statement to me

about femtotech was that he had never thought about the topic
seriously. However he went on to say that it seemed to be a
reasonable idea to pursue. As a mathematician and AI guy
dabbling in physics, this was a relief to me at least the great
physicist didn't laugh at me!!
When I probed Gell-Mann about degenerate matter, he spent a
while musing about the possible varieties of degenerate matter in
which the ordinary notion of quark confinement is weakened.
Confinement is the property that says quarks cannot be
isolated singularly, and therefore cannot be directly observed,
but can only be observed as parts of other particles like protons
and neutrons. At first it was thought that quarks can only be
observed in triplets, but more recent research suggests the
possibility of weak confinement that lets you observe various
aspects of individual quarks in an isolated way. Quark-gluon
63
plasmas , which have been created in particle accelerators
using very high temperatures (like, 4 trillion degrees!), are one
much-discussed way of producing almost unconfined quarks.
But Gell-Mann felt the possibilities go far beyond quark-gluon
plasmas. He said he thought it possible that large groups of
quarks could potentially be weakly-confined in more complex
ways, that nobody now understands.
So after some fun discussion in this vein, I pressed Gell-Mann
specifically on whether understanding these alternative forms of
weak multi-quark confinement might be one way to figure out
how to build stable degenerate matter at Earth gravity...
His answer was, basically, definitely maybe.
Then we changed the topic to AI and the Singularity, where I'm
on firmer ground and there he was a little more positive,
actually. He said he thought it was crazy to try to place a precise
63
http://www.bnl.gov/rhic/physics.asp
308
date on the Singularity, or to estimate anything in detail about it

in advance but he was sympathetic to the notion of
accelerating technological change, and very open to the idea
that massively more change is on the way. And, contra his fellow
physicist Roger Penrose, he expressed doubt that quantum
computing (let alone femtocomputing) would be necessary for
achieving human-level AI. Even if the human brain somehow
uses strange quantum effects in some particulars, he felt, digital
computers should most likely be enough to achieve human-level
intelligence.
A few moments later at the same event, I asked a young Caltech
physics postdoc the same questions about degenerate matter
and femtotech and he gave a similar answer, only mildly more
negative in tone. He said it seemed somewhat unlikely that one
could make room-temperature stable structures using
degenerate matter, but that he couldn't think of any strong
reason why it would be impossible....
Currently, it seems, where degenerate matter based femtotech is
concerned nobody knows.
Strange Matter and Other Strange Matters

64
Gell-Mann's comments reminded me of strangelets strange

hypothetical constructs I first found out about a few years ago
when reading about some strange people who had the strange
65
idea that the Large Hadron Collider might destroy the world by
unleashing a strange chain reaction turning the Earth into
strangelets. Fortunately this didn't happen and it seems at
least plausible that strangelets might pose a route to stable
degenerate matter of a form useful for femtotech.
64
http://www.wisegeek.com/what-is-a-strangelet.htm
http://public.web.cern.ch/public/en/lhc/safetyen.html
65
309
A strangelet is (or would be, if they exist at all, which is unknown)

an entity consisting of roughly equal numbers of up, down and
strange quarks. A small strangelet would be a few femtometers
across, with around the mass of a light nucleus. A large
strangelet could be meters across or more, and would then be
called a strange star or a quark star.
In a (hypothetical) strange star, quarks are not confined in the
traditional sense, but may still be thought of weakly confined in
some sense (at least that's Gell-Mann's view)
So far, all the known particles with strange quarks like the
Lambda particle are unstable. But there's no reason to believe
that states with a larger number of quarks would have to suffer
66
67
from this instability. According to Bodmer and Witten's
strange matter hypothesis, if enough quarks are collected
together, you may find that the lowest energy state of the
collective is a strangelet, i.e. a state in which up, down, and
strange quarks are roughly equal in number.
66
http://prd.aps.org/abstract/PRD/v4/i6/p1601_1
67
http://prd.aps.org/abstract/PRD/v30/i2/p272_1
310
The hypothetical chain reaction via which strangelets eat the Earth
So, where does the End of the World come in? There are some
interesting (albeit somewhat speculative) arguments to the effect
that if a strangelet encounters ordinary matter, it could trigger a
chain reaction in which the ordinary matter gets turned into
strangelets, atom by atom at an accelerating pace. Once one
strangelet hits a nucleus, it would likely turn it into strange
matter, thus producing a larger and more stable strangelet,
which would in turn hit another nucleus, etc. Goodbye Earth,
hello huge hot ball of strange matter. This was the source of the
worries about the LHC, which did not eventuate since when the
LHC was utilized no strangelets were noticeably produced.
One of the many unknowns about strangelets is their surface
tension nobody knows how to calculate this, at present. If
strangelets' surface tension is strong enough, large stable
strangelets should be possible and potentially, strangelets with
complex structure as femtotech requires.
And of course, nobody knows what happens if you vibrate
strangelets very very fast with small amplitude can you
311
produce stable strangelets via dynamic stabilization? Could this

be a path to viable femtotechnology, even if stable strangelets
don't occur in nature. After all, carbon nanotubes appear not to
occur in nature either.
The Future of Femtotech

So what's the bottom line is there still more room at the
bottom?
Nanotech is difficult engineering based on mostly known physics.
Femtotech, on the other hand, pushes at the boundaries of
known physics. When exploring possible routes to femtotech,
one quickly runs up against cases where physicists just don't
know the answer.
Degenerate matter of one form or another seems a promising
potential route to femtotech.
Bolonkin's speculations are
intriguing, as are the possibilities of strangelets or novel weakly
confined multi-quark systems. But the issue of stability is a
serious one; nobody yet knows whether large strangelets can be
made stable, or whether degenerate matter can be created at
normal gravities, nor whether weakly confined quarks can be
observed at normal temperatures, etc. Even where the relevant
physics equations are believed known, the calculations are too
hard to do given our present analytical and computational tools.
And in some cases, e.g. strangelets, we run into situations where
different physics theories held by respected physicists probably
yield different answers.
Putting my AI futurist hat on for a moment, I'm struck by what a
wonderful example we have here of the potential for an only
slightly superhuman AI to blast way past humanity in science
and engineering. The human race seems on the verge of
understanding particle physics well enough to analyze possible
routes to femtotech. If a slightly superhuman AI, with a talent for
physics, were to make a few small breakthroughs in
computational physics, then it might (for instance) figure out how
312
to make stable structures from degenerate matter at Earth

gravity. Bolonkin-style femtostructures might then become
plausible, resulting in femtocomputing and the slightly
superhuman AI would then have a computational infrastructure
capable of supporting massively superhuman AI. Can you say
Singularity? Of course, femtotech may be totally unnecessary
in order for a Vingean Singularity to occur (in fact I strongly
suspect so). But be that as it may, it's interesting to think about
just how much practical technological innovation might ensue
from a relatively minor improvement in our understanding of
fundamental physics.
Is it worth thinking about femtotech now, when the topic is
wrapped up with so much unresolved physics? I think it is, if for
no other reason than to give the physicists a nudge in certain
directions that might otherwise be neglected. Most particle
physics work even experimental work with particle accelerators
seems to be motivated mainly by abstract theoretical interest.
And there's nothing wrong with this understanding the world is
a laudable aim in itself; and furthermore, over the course of
history, scientists aiming to understand the world have spawned
an awful lot of practically useful by-products. But it's interesting
to realize that there are potentially huge practical implications
waiting in the wings, once particle physics advances a little more
if it advances in the right directions.
So, hey, all you particle physicists and physics funding agency
program managers reading this article (and grumbling at my
oversimplifications: sorry, this is tough stuff to write about for a
nontechnical audience!), please take note why not focus some
attention on exploring the possibility of complexly structured
degenerate matter under Earthly conditions, and other possibly
femtotech-related phenomena?
Is there still plenty more room at the bottom, after the nanoscale
is fully explored for the construction of super-advanced AGI
and other future technologies? It seems quite possibly so but
we need to understand what goes on way down there a bit
313
better, before we can build stuff at the femtoscale. Fortunately,

given the exponentially accelerating progress we're seeing in
some relevant areas of technology, the wait for this
understanding and the ensuing technologies may not be all that
long.
314
20
The Singularity Institute's Scary Idea
(and Why I Don't Buy It)
This essay, posted on my blog The Multiverse According to
Ben in 2009, attracted more comments than any other blog post
68
Ive made because the organization it discussed, SIAI,
possessed a great number of passionate supporters, as well as
a great number of passionate detractors. The version given
here was edited by me at some later point, but is essentially the
same as the blog post version.
I felt some reluctance about including this essay here in this
book, because SIAI has somewhat moved on from where it was
back in 2009 including a name change (to MIRI, Machine
Intelligence Research Institute) and a bunch of new
management.
However, in the end I decided to include it,
because in spite of MIRIs greater professionalism, the basic
SIAI perspective I critiqued in my essay is still there at the heart
of MIRI; and in various ways the debate that this essay
represents is still ongoing.
As I write these words in 2014,
Oxford philosopher Nick Bostroms book Superintelligence has
recently been published and in large part, it consists of a more
academic and sober presentation of a minor variant of the line of
thinking I call here The Singularity Institutes Scary Idea .
In a 2012 online dialogue with new MIRI executive Luke
Muehlhauser, I delved into these same issues in an updated
way, including responses to some of Bostroms essays which
form the conceptual core of his Superintelligence book. That
68
http://multiverseaccordingtoben.blogspot.hk/2010/10/s
ingularity-institutes-scary-idea-and.html
315
dialogue is contained in my book Between Ape and Artilect,

and can fairly be considered the next step in the debate after the
essay youre about to read:
This essay focuses on some of the ideas of the organization that
until 2013 was called SIAI, the Singularity Institute for Artificial
Intelligence (and is now called MIRI, the Machine Intelligence
Research Institite). This is a group that, for all its flaws (some of
which Ill focus on here), is one of the more ambitious and
interestingly-focused organizations on the planet.
I have pretty severe issues with a particular line of thinking this
group has advocated, that I'll refer to here as "SIAI's Scary Idea."
Roughly, the Scary Idea posits that: If I or anybody else actively
trying to build advanced AGI succeeds, we're highly likely to
cause an involuntary end to the human race.
When SIAI rebranded as MIRI, they changed their marketing
focus and organizational structure significantly but at core, they
certainly didnt drop the Scary Idea.
Ah, and one more note to forestall (or maybe provoke, for some
readers) confusion: the Singularity Institute for AI is not, and
never was, affiliated with the other California futurist organization
called Singularity University though there are some overlaps.
Ray Kurzweil founded SU and advised SIAI; and I previously
advised SIAI and currently advise SU. Further, in 2013, as part
of rebranding itself as MIRI, SIAI sold the Singularity Summit
conference series to Singularity University. Small Singularitarian
world!!
As the organization was called SIAI for most of the time I
interacted with them, and when I wrote the initial version of this
chapter (as a blog post), and as SIAI/MIRI looks ugly, Im going
to keep referring to SIAI throughout this chapter.
316
The Singularity Institute's Scary Idea (and Why I Don't Buy It)
Since this chapter is going to be somewhat critical, I first want to

clarify that I'm not in disagreement with the existence of SIAI as
an institution, nor with the majority of their activities. I only
disagree with certain positions habitually held by some SIAI
researchers, and by the community of individuals heavily
involved with SIAI.
Actually, SIAI has been quite good to me, overall. I enjoyed all
the Singularity Summits, which they've hosted, very much (well, I
didnt attend the 2011 and 2012 Summits, but I spoke at all the
previous ones). I feel these Summits have played a major role in
the advancement of society's thinking about the future, and I've
felt privileged to speak at them. I applaud SIAI for having
consistently being open to Summit speakers whose views are
strongly divergent from those commonly held by the SIAI
community.
Additionally, to their credit, in 2008, SIAI and my company
69
Novamente LLC seed-funded the OpenCog open-source AGI
project, which was based on software code spun out from
Novamente. The SIAI/OpenCog relationship diminished
substantially when Tyler Emerson passed the leadership of SIAI
along to Michael Vassar, but it was very instrumental in getting
OpenCog off the ground. I've also enjoyed working with Michael
Vassar on the Board of Humanity+, of which I am Chair and he is
a Board member.
When SIAI was helping fund OpenCog, I took the title of
"Director of Research" of SIAI, but I never actually directed any
research there apart from OpenCog. The other SIAI research
was always directed by others, which was fine with me. There
were occasional discussions about operating in a more unified
manner, but it didn't happen. All this is perfectly ordinary in a
small start-up type organization. And once SIAI decided
OpenCog was no longer within its focus, after a bit of delay I
69
http://opencog.org/
317
decided it didn't make sense for me to hold the Director of

Research title anymore, since as things were evolving, I wasn't
directing any SIAI research. I remain as an Advisor to SIAI,
which is going great. Given all this history, I feel a bit like by
writing this chapter Im biting the hand that once fed me - or at
least, that once fed OpenCog, in its infancy. But, one of the
positive attributes of SIAI is the emphasis it places on rationality
and devotion to truth. So in this sense, a chapter like this, which
rationally questions the truth of some of SIAIs frequent
assertions, is very much in the SIAI spirit.
And so, without further ado
SIAI's Scary Idea (Which I Don't Agree With)

The collection of concepts I call SIAI's Scary Idea has been
worded in many different ways by many different people. In the
following paragraph Ive tried to word it in a way that captures the
idea fairly if approximately, and won't annoy too many people:
SIAIs Scary Idea: Progressing toward advanced AGI without a
design for "provably non-dangerous AGI" (or something closely
analogous, often called "Friendly AI" in SIAI lingo) is highly likely
to lead to an involuntary end for the human race.
One of the issues with the Scary Idea is that its rarely clarified
what "provably" really means. A mathematical proof can only be
applied to the real world in the context of some assumptions, so
one reasonable interpretation of "a provably non-dangerous AGI"
is "an AGI whose safety is implied by mathematical arguments
together with assumptions that are believed reasonable by some
responsible party?" Of course, this begs the question of who
responsible party is. Perhaps "the overwhelming majority of
scientists," or some self-appointed group of experts, such as
SIAI itself?
If you read the preceding chapters of this book you know that,
although I don't agree with the Scary Idea, I do agree that the
318
development of advanced AGI has significant risks associated

with it. There are also dramatic potential benefits associated with
it, including the potential of protection against risks from other
technologies (like nanotech, biotech, narrow AI, etc.). So the
development of AGI has difficult cost-benefit balances
associated with it, just like the development of many other
technologies.
I also agree with Nick Bostrom, a host of Science Fiction writers,
and many others, that AGI is a potential "existential risk."
Meaning that, in the worst case, AGI could wipe out humanity
entirely. I think nanotech, biotech, and narrow AI could also do
so, along with a bunch of other things.
I certainly don't want to see the human race wiped out! I
personally would like to transcend the legacy human condition
and become a transhuman superbeing. And I would like
everyone else to have the chance to do so, if they want to.
However, even though I think this kind of transcendence will be
possible, and will be desirable to many, I wouldn't like to see
anyone forced to transcend in this way. I would like to see the
good old fashioned human race continue, if there are humans
who want to maintain their good old fashioned humanity, even if
other options are available.
But SIAI's Scary Idea goes way beyond the mere statement that
there are risks as well as benefits associated with advanced AGI,
and that AGI is a potential existential risk.
Finally, I note that most of the other knowledgeable futurist
scientists and philosophers who have come into close contact
with SIAI's perspective, also don't accept the Scary Idea.
Examples include Robin Hanson and Ray Kurzweil.
Obviously, Im not opposed to anyone having radical ideas that
their respected peers mostly don't accept! I totally get that. My
own approach to AGI is somewhat radical, and most of my
friends in the AGI research community, while they respect my
319
work and see its potential, aren't quite as enthused about it as I

am. Radical positive changes are often brought about by people
who clearly understand certain radical ideas well before anyone
else "sees the light." However, my own radical ideas are not
telling whole research fields that if they succeed they're bound to
kill everybody. So, in that way, they are in a much different
situation than the Scary Idea!
What is the Argument for the Scary Idea?

Although an intense interest in rationalism is one of the
hallmarks of the SIAI community, I still have not yet seen a clear
logical argument for the Scary Idea laid out anywhere. SIAI
Media Director Michael Anissimov says hes working on a book
that will present such an argument, along with other topics. But
for the moment, if one wants a clear argument for the Scary
Idea, one basically has to construct it oneself.
As far as I can tell from discussions and the available online
material, some main ingredients of peoples' reasons for believing
the Scary Idea are ideas like:
1. If one pulled a random mind from the space of all
possible minds, the odds of it being friendly to humans
(as opposed to, e.g., utterly ignoring us, and being
willing to repurpose our molecules for its own ends) are
very low.
70
2. Human value is fragile , as well as complex, so if you

create an AGI with a roughly-human-like value system,
then this may not be good enough, and it is likely to
rapidly diverge into something with little or no respect for
human values.
3. "Hard takeoffs" (in which AGIs recursively self-improve
and massively increase their intelligence) are fairly likely
70
http://lesswrong.com/lw/y3/value_is_fragile/
320
once AGI reaches a certain level of intelligence; and

humans will have little hope of stopping these events.
4. A hard takeoff, unless it starts from an AGI designed in a
"provably Friendly" way, is highly likely to lead to an AGI
system that doesn't respect the rights of humans to
exist.
Note, Im not directly quoting any particular thinker associated
with SIAI here. I'm merely summarizing, in my own words, ideas
that I've heard and read very often from various individuals
associated with SIAI.
If you put the above points all together, you come up with a
heuristic argument for the Scary Idea. Roughly, the argument
goes something like: If someone builds an advanced AGI without
a provably Friendly architecture, probably it will have a hard
takeoff, and then probably this will lead to a superhuman AGI
system with an architecture drawn from the vast majority of
mind-architectures that are not sufficiently harmonious with the
complex, fragile human value system to make humans happy
and keep humans around.
The line of argument makes sense, if you accept the premises.
But, I don't. I think the first of the above points is reasonably
plausible, though I'm not by any means convinced. I think the
relation between breadth of intelligence and depth of empathy is
a subtle issue which none of us fully understands (yet). It's
possible that with sufficient real-world intelligence tends to come
a sense of connectedness with the universe that militates against
squashing other sentients.
But I'm not terribly certain of this, any more than I'm terribly
certain of its opposite. I agree much less with the final three
points listed above, and I haven't seen any careful logical
arguments for these points. I doubt human value is particularly
fragile. Human value has evolved and morphed over time and
will continue to do so. It already takes multiple different forms. It
321
will likely evolve in future in coordination with AGI and other

technology. I think it's fairly robust.
I think a hard takeoff is possible, though I don't know how to
estimate the odds of one occurring with any high confidence. I
think it's very unlikely to occur until we have an AGI system that
has very obviously demonstrated general intelligence at the level
of a highly intelligent human. And I think the path to this "hard
takeoff enabling" level of general intelligence is going to be
somewhat gradual, not extremely sudden.
I don't have any strong sense of the probability of a hard takeoff,
from an apparently but not provably human-friendly AGI, leading
to an outcome likable to humans. I suspect this probability
depends on many features of the AGI, which we will identify over
the next years and decades, via theorizing based on the results
of experimentation with early-stage AGIs.
Yes, you may argue: the Scary Idea hasn't been rigorously
shown to be true but what if it IS true?
OK but ... pointing out that something scary is possible, is a very
different thing from having an argument that it's likely.
The Scary Idea is certainly something to keep in mind, but there
are also many other risks to keep in mind, some much more
definite and palpable. Personally, I'm a lot more worried about
nasty humans taking early-stage AGIs and using them for
massive destruction, than about speculative risks associated
with little-understood events like hard takeoffs.
Is Provably Safe or "Friendly" AGI A Feasible

Idea?
The Scary Idea posits that if someone creates advanced AGI
that isn't somehow provably safe, it's almost sure to kill us all.
322
But not only am I unconvinced of this, I'm also quite unconvinced

that a "provably safe" AGI is even feasible. The idea of provably
safe AGI is typically presented as something that would exist
within mathematical computation theory or some variant thereof.
So, right off the bat, that's one obvious limitation of the idea, as
mathematical computers don't exist in the real world, and realworld physical computers must be interpreted in terms of the
laws of physics, and humans' best understanding of the "laws" of
physics seem to radically change from time to time. So, even if
there were a design for provably safe real-world AGI, based on
current physics, the relevance of the proof might go out the
window when physics next gets revised.
Also, there are always possibilities like: the alien race that is
watching us and waiting for us to achieve an IQ of 333, at which
point it will swoop down upon us and eat us, or merge with us.
We can't rule this out via any formal proof, and we can't
meaningfully estimate the odds of it either. Yes, this sounds
science-fictional and outlandish; but is it really more outlandish
and speculative than the Scary Idea?
A possibility that strikes me as highly likely is that, once we have
created advanced AGI, and have linked our brains with it
collectively, most of our old legacy human ideas (including
physical law, aliens, and Friendly AI) will seem extremely limited
and ridiculous.
Another issue is that the goal of "Friendliness to humans" or
"safety" or whatever you want to call it, is rather nebulous and
difficult to pin down. Science fiction has explored this theme
extensively. So, even if we could prove something about "smart
AGI systems with a certain architecture that are guaranteed to
achieve goal G," it might be infeasible to apply this to make AGI
systems that are safe in the real-world, simply because we don't
know how to boil down the everyday intuitive notions of "safety"
or "friendliness" into a mathematically precise goal G like the
proof refers to.
323
This is related to the point Eliezer Yudkowsky makes that "value

is complex." Actually, human value is not only complex, it's
nebulous and fuzzy, and ever-shifting, and humans largely grok
it by implicit procedural, empathic and episodic knowledge,
rather than explicit declarative or linguistic knowledge.
Transmitting human values to an AGI is likely to be best done via
interacting with the AGI in real life, but this is not the sort of
process that readily lends itself to guarantees or formalization.
But, setting those worries aside, is the computation-theoretic
version of provably safe AI even possible? Could one design an
AGI system and prove, in advance, that, given certain
reasonable assumptions about physics and its environment, it
would never veer too far from its initial goal (e.g. a formalized
version of the goal of treating humans safely, or whatever)?
I very much doubt one can do so, except via designing a
fictitious AGI that can't really be implemented because it uses an
infeasible amount of computational resources. In fact, Ive put a
fair amount of thought into an AGI design like this, which I call
GOLEM. Ill briefly describe GOLEM in a later chapter, because
I think its an interesting intellectual experiment. However,
suspect it's too computationally wasteful to be practically
feasible, at least for now (pre-Singularity).
I strongly suspect that to achieve high levels of general
intelligence using realistically limited computational resources,
one is going to need to build systems with a nontrivial degree of
fundamental unpredictability to them. This is what neuroscience
suggests, it's what my concrete AGI design work suggests, and
it's what my theoretical work on GOLEM and related ideas
suggests. And none of the public output of SIAI researchers or
enthusiasts has given me any reason to believe otherwise, yet.
Practical Implications
The above discussion of SIAI's Scary Idea may just sound like
fun science-fictional speculation, but this chapter originated out
324
of a frustrating practical situation that arose as a direct result of

the Scary Idea. In 2010 I wrote an entry on my blog about the
OpenCog project, and the comments field on the blog post got
totally swamped with SIAI-influenced people saying stuff in the
vein of: Creating an AGI without a proof of Friendliness is
essentially equivalent to killing all people! So I really hope your
OpenCog work fails, so you don't kill everybody!!!
On the most extreme side, one amusing/alarming quote from a
commentator (probably not someone directly affiliated with SIAI)
was "if you go ahead with an AGI when you're not 100% sure
that it's safe, you're committing the Holocaust." There were many
other comments in the same rough vein, and also a number of
other similar comments directed to me privately via email.
If one fully accepts SIAI's Scary Idea, then one should not work
on practical AGI projects, ever. Nor should one publish papers
on the theory of how to build AGI systems. Instead, one should
spend one's time trying to figure out an AGI design that is
somehow provable-in-advance to be a Nice Guy. For this
reason, SIAI's research group is not currently trying to do any
practical AGI work.
Actually, so far as I know, my "GOLEM" AGI design (mentioned
above) is closer to a "provably Friendly AI" than anything the
SIAI research team has come up with. (At least anything they
have made public.) I definitely agree that AGI ethics is a very
important problem, but I doubt the problem is most effectively
addressed by theory alone. I think the way to achieve to a useful
real-world understanding of AGI ethics is going to be to
Build some early-stage AGI systems, e.g. artificial

toddlers, scientists' helpers, video game characters,
robot maids and butlers, etc.
Study these early-stage AGI systems empirically, with a

focus on their ethics as well as their cognition.
325
Attempt to arrive at a solid theory of AGI intelligence and

ethics based on a combination of conceptual and
experimental-data considerations.
Have humanity collectively plot the next steps from

there. Based on the theory we find, maybe we go ahead
and create a superhuman AI capable of hard takeoff, or
maybe we pause AGI development because of the risks.
Maybe we build an "AGI Nanny" to watch over the
human race and prevent AGI or other technologies from
going awry. Whatever choice we make then, it will be
made based on far better knowledge than we have right
now.
So what's wrong with this approach? Nothing, really. If you hold

the views of most AI researchers or futurists. There are plenty of
disagreements about the right path to AGI, but wide and implicit
agreement that something like the above path is sensible.
But, if you adhere to SIAI's Scary Idea, there's a big problem with
this approach because, according to the Scary Idea, there's
too huge of a risk that these early-stage AGI systems are going
to experience a hard takeoff and self-modify into something that
will destroy us all. But I just don't buy the Scary Idea.
I do see a real risk that, if we proceed in the manner I'm
advocating, some nasty people will take the early-stage AGIs
and either use them for bad ends, or proceed to hastily create a
superhuman AGI that then does bad things of its own volition.
These are real risks that must be thought about hard, and
protected against as necessary, but they are different from the
Scary Idea, and not really so different from the risks implicit in a
host of other advanced technologies.
Conclusion
Although I think SIAI is performing a useful service by helping
bring these sorts of ideas to the attention of the futurist
326
community (alongside the other services they're performing, like

the wonderful Singularity Summits), that all said, I think the Scary
Idea is potentially a harmful one. At least, it would be a harmful
one, if more people believed it. Luckily, such a viewpoint is
currently restricted to a rather small subset of the futurist
community.
Many people die each day, and many others are miserable for
various reasons -- and all sorts of other advanced and potentially
dangerous technologies are currently under active development.
My own view is that unaided human minds may well be unable to
deal with the complexity and risk of the world that human
technology is unleashing. I actually suspect that our best hope
for survival and growth through the 21st century is to create
advanced AGIs to help us on our way: To cure disease, develop
nanotech and better AGI, invent new technologies, and help us
keep nasty people from doing destructive things with advanced
technology.
I think that to avoid actively developing AGI, out of speculative
concerns like the Scary Idea, would be an extremely bad idea.
That is, rather than "if you go ahead with an AGI when you're not
100% sure that it's safe, you're committing the Holocaust," I
suppose my view is closer to "if you avoid creating beneficial AGI
because of speculative concerns, then you're killing my
grandma!" (Because advanced AGI will surely be able to help us
cure human diseases and vastly extend and improve human
life.)
So perhaps I could adopt the slogan: "You don't have to kill my
grandma to avoid the Holocaust!" But really, folks. Well, you
get the point.
Humanity is on a risky course altogether, but no matter what I
decide to do with my life and career (and no matter what Bill Joy
or Jaron Lanier or Bill McKibben, etc., write), the race is not
327
going to voluntarily halt technological progress. It's just not

happening.
We just need to accept the risk, embrace the thrill of the amazing
time we were born into, and try our best to develop nearinevitable technologies like AGI in a responsible and ethical way.
To me, responsible AGI development doesn't mean fixating on
speculative possible dangers and halting development until illdefined, likely-unsolvable theoretical/philosophical issues are
worked out to everybody's (or some elite group's) satisfaction.
Rather, it means proceeding with the work carefully and openly,
learning what we can as we move along, and letting experiment
and theory grow together, as they have been doing quite
successfully for the last few centuries, at a fantastically
accelerating pace.
328
21
Does Humanity Need an AI Nanny?
This H+ Magazine article, from 2011, is one of my lines of
thought that I feel most ambivalent about. Its a case where my
logical mind pushes in one direction, and my emotions in
another. Emotionally, Im a freedom-loving American anarchist
freak by nature, and I hate the idea of being constantly observed
and monitored by critical eyes, especially eyes connected to
actuators with the power to stop me from doing what I want.
Logically, though, it seems to me that given the increasing
dangers coming along with increasingly advanced technology,
universal surveillance or sousveillance are probably the only
rational courses, if we want to survive long enough to transcend
to the next stages of intelligence evolution.
The article attracted a bit of attention, including getting me
invited to participate in an episode of Michio Kakus popular
science TV show, for which I was filmed talking about the AI
Nanny in the office of a company doing video surveillance of
various offices sitting in front of a bunch of monitors showing
the outputs of security cameras.
Not my usual work
environment! . But meeting Kaku was cool; and we discussed
string theory and various more out-there physics ideas, on the
van ride to the filming location.
One possible solution to the acute risks posed by rapidly
advancing technology development is to build a powerful yet
limited AGI system, with the explicit goal of keeping things on the
planet under control while we figure out the hard problem of how
to create a probably positive Singularity. That is: to create an AI
Nanny.
329
The AI Nanny would forestall a full-on Singularity for a while,

71
restraining it into what Max More has called a Surge , and
giving us time to figure out what kind of Singularity we really
want to build and how. Its not entirely clear that creating such an
AI Nanny is plausible, but Ive come to the conclusion it probably
is. Whether or not we should try to create it, though thats a
different question.
The AI Nanny in Detail

Let me be more precise about what kind of Nanny Im thinking
of. Imagine an advanced AGI system with
General intelligence somewhat above the human level,

but not too dramatically so maybe, qualitatively
speaking, as far above humans as humans are above
apes
Interconnection to powerful worldwide

systems, online and in the physical world
Control of a massive contingent of robots (e.g. service

robots, teacher robots, etc.) and connectivity to the
worlds home and building automation systems, robot
factories, self-driving cars, and so on and so forth
A cognitive architecture featuring an explicit set of goals,

and an action selection system that causes it to choose
those actions that it rationally calculates will best help it
achieve those goals
A set of preprogrammed goals including the following

aspects:
A strong inhibition against modifying its preprogrammed

goals
surveillance
71
http://strategicphilosophy.blogspot.com/2009/06/howfast-will-future-arrive-how-will.html
330
A strong inhibition against rapidly modifying its general

intelligence
A mandate to cede control of the world to a more

intelligent AI within 200 years
A mandate to help abolish human disease, involuntary

human death, and the practical scarcity of common
humanly-useful resources like food, water, housing,
computers, etc.
A mandate to prevent the development of technologies

that would threaten its ability to carry out its other goals
A strong inhibition against carrying out actions with a

result that a strong majority of humans would oppose, if
they knew about the action in advance
A mandate to be open-minded toward suggestions by intelligent,

thoughtful humans about the possibility that it may be
misinterpreting its initial, preprogrammed goals
There you have it: an AI Nanny.
Obviously, this sketch of the AI Nanny idea is highly simplified
and idealized a real-world AI Nanny would have all sort of
properties not described here, and might be missing some of the
above features, substituting them with other related things. My
point here is not to sketch a specific design or requirements
specification for an AI Nanny, but rather to indicate a fairly
general class of systems that humanity might build.
The nanny metaphor is chosen carefully. A nanny watches over
children while they grow up, and then goes away. Similarly, the
AI Nanny would not be intended to rule humanity on a
permanent basis only to provide protection and oversight while
we grow up collectively; to give us a little breathing room so we
can figure out how best to create a desirable sort of Singularity.
A large part of my personality rebels against the whole AI Nanny
approach Im a rebel and a nonconformist; I hate bosses and
331
bureaucracies and anything else that restricts my freedom. But,

Im not a political anarchist because I have a strong suspicion
that if governments were removed, the world would become a lot
worse off, dominated by gangs of armed thugs imposing even
less pleasant forms of control than those exercised by the US
Army and the CCP and so forth. Im sure government could be
done a lot better than any country currently does it but I dont
doubt the need for some kind of government, given the realities
of human nature. Well, maybe maybe the need for an AI
Nanny falls into the same broad category. Perhaps, like
government, an AI Nanny is a relatively offensive thing, that is
nonetheless a practical necessity due to the unsavory aspects of
human nature.
We didnt need government during the Stone Age because
there werent that many of us, and we didnt have so many
dangerous technologies.
But we need government now.
Fortunately, these same technologies that necessitated
government, also provided the means for government to operate.
Somewhat similarly, we havent needed an AI Nanny so far,
because we havent had sufficiently powerful and destructive
technologies. And, these same technologies that apparently
necessitate the creation of an AI Nanny, also appear to provide
the means of creating it.
The Argument for an AI Nanny

To recap and summarize, the basic argument for trying to build
an AI Nanny is founded on the premises that:
1. Its impracticable to halt the exponential advancement of
technology (even if one wanted to)
2. As technology advances, it becomes possible for
individuals or groups to wreak greater and greater
damage using less and less intelligence and resources
332
3. As technology advances, humans will more and more

acutely lack the capability to monitor global technology
development
and
forestall
radically
dangerous
technology-enabled events
4. Creating an AI Nanny is a significantly less difficult
technological problem than creating an AI or other
technology with a predictably high probability of
launching a full-scale positive Singularity
5. Imposing a permanent or very long term constraint on
the development of new technologies is undesirable
The fifth and final premise is normative; the others are empirical.
None of the empirical premises are certain, but all seem likely to
me. The first three premises are strongly implied by recent
social and technological trends. The fourth premise seems
commonsensical based on current science, mathematics and
engineering.
These premises lead to the conclusion that trying to build an AI
Nanny is probably a good idea. The actual plausibility of building
an AI Nanny is a different matter, of course I believe it is
plausible, but of course, opinions on the plausibility of building
any kind of AGI system in the relatively near future vary all over
the map.
333
Complaints and Responses

I have discussed the AI Nanny idea with a variety of people over
the last year or so, and have heard an abundance of different
complaints about it but none have struck me as compelling.
Its impossible to build an AI Nanny; the AI R&D is too
hard. But is it really? Its almost surely impossible to build
and install an AI Nanny this year; but as a professional AI
researcher, I believe such a thing is well within the realm of
possibility. I think we could have one in a couple decades if
we really put our collective minds to it. It would involve a host
of coordinated research breakthroughs, and a lot of largescale software and hardware engineering, but nothing
implausible according to current science and engineering.
We did amazing things in the Manhattan Project because we
wanted to win a war how hard are we willing to try when
our overall future is at stake?
It may be worth dissecting this hard R&D complaint into two
sub-complaints:
AGI is hard: building an AGI system with slightly

greater than human level intelligence is too hard;
Nannifying an AGI is hard: given a slightly

superhuman AGI system, turning into an AI Nanny is too
hard.
Obviously both of these are contentious issues.

Regarding the AGI is hard complaint, at the AGI-09 artificial
72
intelligence research conference, an expert-assessment survey
was done, suggesting that a least a nontrivial plurality of
professional AI researchers believes that human-level AGI is
72
http://hplusmagazine.com/2010/02/05/how-long-tillhuman-level-ai/
334
possible within the next few decades, and that slightlysuperhuman AGI will follow shortly after that.
Regarding the Nannifying an AGI is hard complaint, I think its
validity depends on the AGI architecture in question. If one is
talking about an integrative, cognitive-science-based, explicitly
73
74
goal-oriented AGI system like, say, OpenCog or MicroPsi or
75
LIDA , then this is probably not too much of an issue, as these
architectures are fairly flexible and incorporate explicitly
articulated goals. If one is talking about, say, an AGI built via
closely emulating human brain architecture, in which the
designers have relatively weak understanding of the AGI
systems representations and dynamics, then the nannification
is hard problem might be more serious. My own research
intuition is that an integrative, cognitive-science-based, explicitly
goal-oriented system is likely to be the path via which advanced
AGI first arises; this is the path my own work is following.
Its impossible to build an AI Nanny; the surveillance technology
is too hard to implement. But is it really? Surveillance tech is
advancing bloody fast, for all sorts of reasons more prosaic than
the potential development of an AI Nanny. Read David Brins
76
book The Transparent Society , for a rather compelling
argument that before too long, well all be able to see everything
everyone else is doing.
Setting up an AI Nanny, in practice, would require a world
government. OK, yes it would Sort of. It would require
either a proactive assertion of power by some particular party,
creating and installing an AI Nanny without asking everybody
73
http://opencog.org/
74
http://micropsi.com/publications/Bach_AGI_2011.pdf
75
http://en.wikipedia.org/wiki/LIDA_%28cognitive_archit
ecture%29
76
http://www.davidbrin.com/transparent.htm
335
elses permission; or else a degree of cooperation between the

worlds most powerful governments, beyond what we see today.
Either route seems conceivable. Regarding the second
cooperative path, its worth observing that the world is clearly
moving in the direction of greater international unity, albeit in fits
and starts. Once the profound risks posed by advancing
technology become more apparent to the worlds leaders, the
required sort of international cooperation will probably be a lot
easier to come by. Hugo de Gariss most recent book Multis and
77
Monos riffs extensively on the theme of emerging world
government.
Building an AI Nanny is harder than building a self-modifying,
self-improving AGI that will retain its Friendly goals even as it
self-modifies.
Yes,
someone
really
made
this
counterargument to me; but as a scientist, mathematician and
engineer, I find this wholly implausible. Maintenance of goals
under radical self-modification and self-improvement seems to
pose some very thorny philosophical and technical problem
and once these are solved (to the extent that theyre even
solvable) then one will have a host of currently-unforeseeable
engineering problems to consider. Furthermore there is a huge,
almost surely irreducible uncertainty in creating something
massively more intelligent than oneself. Whereas creating an AI
Nanny is merely a very difficult, very large scale science and
engineering problem.
If someone creates a new technology smarter than the AI
Nanny, how will the AI Nanny recognize this and be able to nip it
in the bud? Remember, the hypothesis is that the AI Nanny is
significantly smarter than people. Imagine a friendly, highly
intelligent person monitoring and supervising the creative
projects of a room full of chimps or intellectually challenged
individuals.
77
http://www.amazon.com/Multis-MulticulturedMonocultured-Towards-Creation/dp/0882801627/
336
Why would the AI Nanny want to retain its initially preprogrammed goals, instead of modifying them to suit itself
better? for instance, why wouldnt it simply adopt the goal of
becoming an all-powerful dictator and exploiting us for its own
ends? But why would it change its goals? What forces would
cause it to become selfish, greedy, etc.? Lets not
anthropomorphize. Power corrupts, and absolute power
corrupts absolutely is a statement about human psychology, not
a general law of intelligent systems. Human beings are not
architected as rational, goal-oriented systems, even though
some of us aspire to be such systems and make some progress
toward behaving in this manner. If an AI system is created with
an architecture inclining it to pursue certain goals, theres no
reason why it would automatically be inclined to modify these
goals.
But how can you specify the AI Nannys goals precisely? You
cant right? And if you specify them imprecisely, how do you
know it wont eventually come to interpret them in some way that
goes against your original intention? And then if you want to
tweak its goals, because you realize you made a mistake, it
wont let you, right? This is a tough problem, without a perfect
solution. But remember, one of its goals is to be open-minded
about the possibility that its misinterpreting its goals. Indeed,
one cant rule out the possibility that it will misinterpret this metagoal and then, in reality, closed-mindedly interpret its other goals
in an incorrect way. The AI Nanny would not be a risk-free
endeavor, and it would be important to get a feel for its realities
before giving it too much power. But again, the question is not
whether its an absolutely safe and positive project but rather,
whether its better than the alternatives!
78
What about Steve Omohundros Basic AI Drives ? Didnt

Omohundro prove that any AI system would seek resources and
78
http://selfawaresystems.com/2007/11/30/paper-onthe-basic-ai-drives/
337
power just like human beings? Steves paper is an instant

classic, but his arguments are mainly evolutionary. They apply to
the case of an AI competing against other roughly equally
intelligent and powerful systems for survival. The posited AI
Nanny would be smarter and more powerful than any human,
and would have, as part of its goal content, the maintenance of
this situation for 200 years (200 obviously being a somewhat
arbitrary number inserted for convenience of discussion). Unless
someone managed to sneak past its defenses and create
competitively powerful and smart AI systems, or it encountered
alien minds, the premises of Omohundros arguments dont
apply.
What happens after the 200 years is up? I have no effing
idea, and thats the whole point. I know what I want to happen
I want to create multiple copies of myself, some of which remain
about like I am now (but without ever dying), some of which
gradually ascend to godhood via fusing with uber-powerful AI
minds, and the rest of which occupy various intermediate levels
of transcension. I want the same to happen for my friends and
family, and everyone else who wants it. I want some of my
copies to fuse with other minds, and some to remain distinct. I
want those who prefer to remain legacy humans, to be able to do
so. I want all sorts of things, but thats not the point the point is
that after 200 years of research and development under the
protection of the AI Nanny, we would have a lot better idea of
whats possible and what isnt than any of us do right now.
What happens if the 200 years pass and none of the hard
problems are solved, and we still dont know how to launch a fullon Singularity in a sufficiently reliably positive way? One
obvious possibility is to launch the AI Nanny again for a couple
hundred more years. Or maybe to launch it again with a different,
more sophisticated condition for ceding control (in the case that
it, or humans, conceive some such condition during the 200
years).
338
What if we figure out how to create a Friendly self-improving

massively superhuman AGI only 20 years after the initiation of
the AI Nanny then wed have to wait another 180 years for the
real Singularity to begin! Thats true of course, but if the AI
Nanny is working well, then were not going to die in the interim,
and well be having a pretty good time. So whats the big deal?
A little patience is a virtue!
But how can you trust anyone to build the AI Nanny? Wont
they secretly put in an override telling the AI Nanny to obey
them, but nobody else? Thats possible, but there would be
some good reasons for the AI Nanny developers not to do that.
For one thing, if others suspected the AI Nanny developers had
done this, some of these others would likely capture and torture
the developers, in an effort to force them to hand over the secret
control password. Developing the AI Nanny via an open,
international, democratic community and process would diminish
the odds of this sort of problem happening.
What if, shortly after initiating the AI Nanny, some human sees
some fatal flaw in the AI Nanny approach, which we dont see
now. Then wed be unable to undo our mistake. Oops.
But its odious!! Yes, its odious. Government is odious too,
but apparently necessary. And as Winston Churchill said,
Democracy is the worst possible form of government, except for
all the others. Human life, in many respects, is goddamned
odious. Nature is beautiful and cooperative and synergetic and
also red in tooth and claw. Life is wonderful, beautiful and
amazing and tough and full of compromises. Hell, even physics
is a bit odious some parts of my brain find the Second Law of
Thermodynamics and the Heisenberg Uncertainty Principle
damned unsatisfying! I wouldn't have written this chapter in my
early 20s, because back then I was more steadfastly oriented
toward idealistic solutions but now, in my mid-40s, Ive pretty
well come to terms with the universes persistent refusal to
behave in accordance with all my ideals. The AI Nanny scenario
is odious in some respects but human life is odious sometimes,
339
and we need to be thinking about realistic solutions It stands to

reason that the transcendence of humanity may involve some
measure of human odiousness as well as a rich dose of human
beauty!
A Call to Brains
This chapter is not supposed to be a call to arms to create an AI
Nanny. As Ive said above, the AI Nanny is not an idea that thrills
my heart. It irritates me. I love freedom, and Im also impatient
and ambitious I want the full-on Singularity yesterday,
goddammit!!!
But I do think its worth seriously considering whether some form
of AI Nanny might well be the best path forward for humanity
the best way for us to ultimately create a Singularity according to
our values. At very least, its worth fleshing out the idea fairly
carefully, and weighing it against alternatives.
So this is more of a call to brains, really. Id like to get more
people thinking about what an AI Nanny might be like, and how
we might engineer one. And Id like to get more people thinking
actively and creatively about alternatives.
Perhaps you dislike the AI Nanny idea even more than I do. But
even so, consider: Others may feel differently. You may well
have an AI Nanny in your future anyway. And even if the notion
seems unappealing now, you may well enjoy it quite
tremendously when it actually happens.
Oh Brave New World
340
22
AGI, Consciousness, Spirituality, Life, the
Universe, and Everything
My futurist friend Giulio Prisco has been striving for some time to
bring together transhumanism and religion. His Turing Church
blog pursues this theme, and he has also organized a series of
online conferences. Conversations with Giulio formed much
of my inspiration for writing A Cosmist Manifesto though that
book ended up philosophical rather than religious in nature in
the end I dont quite share Giulios vision that futurism should, or
will, end up having all that much in common with traditional
religions. In any case, the following lengthy H+ Magazine article
(edited a bit for inclusion here) was also inspired by my ongoing
dialogue with Giulio, and represents an attempt by my 2012 self
to clarify the various philosophical and spiritual perspectives at
play regarding future technologies, the Singularity, and so forth.
Artificial General Intelligence is a technical topic, but its also
much more than that. It has big implications about humanitys
place in the universe both what humanitys place is logically
and scientifically speaking, and what we want it to be
normatively. It raises a host of dizzying possibilities regarding our
possible futures, including the potential of fusing our minds with
AGIs cyborgically, and becoming more than human in various
senses (some quite strong). It even raises the possibility of
understanding the universe and finding answers to the Big
Questions of Life, the Universe and Everything with the help of
engineered transhuman minds, to aid us in venturing where the
human mind is too weak to travel. And then the possibilities
highlighted by folks like Hans Moravec, Hugo de Garis and the
SIAI, of human extinction, of our replacement by transhuman
AGIs who dont care anymore for humans than humans do for
flies or bacteria.
341
One critical aspect of the relation between AGI and the Big
Picture has to do with consciousness. If one thinks AGIs can be
conscious in the sense that humans are or maybe even more
broadly and richly conscious then that certainly informs ones
sense of AGIs place in the Cosmos. On the other hand if one
things AGIs are intrinsically just feeling-less, experience-less
tools, then the story becomes quite different.
There are many, many ways of fitting AGI into ones view of
79
Cosmos and consciousness; a few possibilities are as follows :
79
Physical Monism. There is a physical universe with a

kind of independent, absolute existence (though of a
commonsensically peculiar sort, as quantum theory has
taught us), and talking of anything besides physical
entities and configurations thereof is basically just
babbling nonsense.
Informational Monism: The universe basically consists

of information, and talking about anything besides
information isnt really possible.
People are
conglomerations of information; and AGIs will ultimately
be more complex and capable conglomerations of
information. Soul and consciousness are merely
conceptual
constructs,
i.e.
conglomerations
of
information created by other conglomerations of
information to communicate with themselves and each
other. Physical entities like rocks and electrons are
also best thought of as conglomerations of information,
since we only know them via the information we perceive
about them. In a sense, this means the universe is
essentially some kind of computer (though a massively
parallel one).
Note that the labels in this list are ones Ive adopted for my own
convenience, rather than necessarily because theyre the most standard
ones.
342
AGI, Consciousness, Spirituality, Life, the Universe, and Everything
Quantum Informational Monism: Like the above, but

on quantum theory. The universe basically consists of
quantum information, thus is essentially a quantum
computer.
Cognitive Vitalism: Human-level AGI is impossible

because intelligence requires some kind of soul or
other special entity, that humans have but engineered
entities will never have
Experiential Vitalism: Superhumanly capable AGI is

possible, but even so, these AGIs will in a sense be
zombies without the same kind of consciousness (aka
soul?) that people have
Panpsychism: Every element of the physical or

informational realms is in a sense already mental, with
some sort of spark of awareness; indeed it makes no
sense to talk about matter separate from mind. The
consciousness associated with a complex system like a
brain, body, AGI software system or robot is just a
particular manifestation of the universal mental aspect of
reality.
Spiritual Monism: Since the material of the universe is

spiritual already, as part of its intrinsic nature, each
configuration of the universes matter/information is
already associated with its own variety of soul.
Superhuman AGI is possible, and such AGIs will have
the same kind of special soul-stuff as humans do
perhaps manifesting it more intensely in some ways, as
they develop. Fusing with superhuman AGIs, uploading,
or improving our brains in various ways, may be part of
our spiritual growth quest.
Cosmism: The universe itself is a massively complex

intelligent system, and humans and AGIs are small parts
of it, and as systems become more intelligent they have
more and more potential to harmonize with the (currently
343
somewhat mysterious to us) ambient intelligence of the

Cosmos
The Implicate Order: There is a sort of implicate

information structuring the universe, which is different
from explicitly observable information and physical
entities have both implicate and explicate aspects.
Conscious experience has to do with both aspects, and
to the extent there is an intelligence to the whole
universe it would also have both implicate and explicate
aspects. One idea Ive been experimenting with is to
model the implicate order as a kind of cosmist
questioning process.
Of course this is not intended as a complete list of possible

perspectives And its also worth noting that these are not
mutually exclusive categories, by any means! Some individuals
or schools of thought may adhere to more than one of these
perspectives.
Sometimes I hear people express the nave thought that AGI
research is naturally aligned with a particular philosophical
perspective, usually the views I refer to above as physical or
informational monism. But I dont think any such natural
alignment exists. Ive observed AGI researchers and AGIfocused transhumanist thinkers to take a great variety of
perspectives.
In this chapter Ill briefly run through these views Ive listed, with
more emphasis on breadth than depth. I dont want to turn this
into a book on cosmic philosophy! But I do want to indicate the
wonderful diversity of ways in which the notion of AGI is already
being interwoven with various human thought and belief
systems. One can expect the weaving of AGI into these various
human conceptual networks to get even richer once advanced
AGI is actually created and interacts with us on a regular basis!
Ill spend relatively little time on those perspectives that are
already extremely well-known due to their commonality, and
344
dwell more on those views I find most interesting! Also, note that
two of my own favorite perspectives appear at the end of the
chapter (Cosmism and the Implicate Order My other favorites
of the bunch being Informational Monism and Panpsychism). So
if you don't read to the end youll miss the good stuff!
Physical Monism
What I refer to as physical monism is the view that the physical
world is in some sense absolutely real and out there and existent
and everything else is a bunch of malarkey. Intelligence, mind
and consciousness and experience are viewed as patterns or
configurations (or whatever) of physical entities like particles or
waves.
To my mind, this is a potentially useful (though rather limited)
perspective for getting through everyday life, but somewhat
deficient as a real intellectual theory.
After all, how do we know about this supposedly absolutely
existent physical world? Either we take it on faith, like a
prototypical religious persons belief in God Or we infer it from
various observations, i.e. various pieces of information. But if its
the latter then isnt the more fundamental reality associated
with the information, rather than with the physical world whose
existence we infer from the information?
A classic expression of this view is G.E. Moores observation that
when you kick a rock, you know its real, and philosophical
babbling becomes irrelevant in the light of this direct evidence of
the rocks reality. But what is it that you actually know is real,
when you kick the rock? The definite experienced reality is
attached to the feelings coming to you from your foot (i.e.
feelings that you have learned to attach to the concept/percept
network called foot), and the sensations coming into your eye
when you look at your rock and the foot. Actually there might be
no rock present in physical reality at all your brain might be
345
connected to an apparatus that makes you think youre kicking a

rock.
And this leads up to the next perspective in my list, which I find
deeper and more interesting
Informational Monism
The view of the world as bottoming out in some absolute
physical reality seems hopelessly nave to me -- but the view of
the world as bottoming out in information seems much less so. I
dont quite adhere to this perspective myself, but nor do I know
of any rational, scientific arguments against it. AGI researcher
Joscha Bach, whom I interviewed above, has described the
perspective well:
We grow up with the illusion of a direct access to an outside
world, and this intuition is reflected in the correspondence
theory of truth: our concepts derive their meaning from their
correspondence to facts in a given reality. But how does this
correspondence work? According to our current
understanding of biology, all access to the world is mediated
through a transfer of data, of bits (i.e., discernible
differences) expressed by electrical impulses through
sensory and motor nerves. The structure of the world, with
percepts, concepts, relations and so on, is not part of this
data transfer, but is constructed by our minds. It represents
encodings over the regularities found in the data patterns at
the mind's interface. In theory, all the data entering my mind
over my lifetime could be recorded as a finite (but very long)
vector of bits, and everything that I consider to be knowledge
of the world is a more concise re-encoding of parts of this
vector.
Of course, even the concept that the world is mediated
through sensory nerves is not directly accessible. It is an
encoding, too (it just happens to be the best encoding that
we found so far). And because we cannot know the "real"
346
structure behind the data patterns at our mind's systemic

interface, the correspondence theory of truth does not work,
outside the abstract and pure realm of mathematics.
But what about information itself: how can we know about
bits, data and encodings? Fortunately, these are
mathematical concepts. They can be defined and operated
upon outside of any empirical world. For instance, the theory
of Natural Numbers does not need anything material (like
piles of apples and oranges) to work: It follows automatically
from the completely abstract and theoretical set of Peano's
axioms. But Natural Numbers can be used to encode some
aspect of apples and oranges--their cardinality, and their
behavior with respect to addition and subtraction.
The space of mathematics is self-contained, but to explore it,
we need a certain kind of information processing system,
which is a mathematical entity too. Minds are part of the
class of information processing systems that can perform
mathematics (at least to some extend), but minds can do
even more: they can conceptualize a world, reflect, plan,
imagine, anticipate, decide, dream, interpret themselves as
persons, be in emotional states, attach relevance to
concepts and so on. The concept of mind is the one we
attach to ourselves, we use it to encode that part of the
information vector that we consider to be us.
With respect to our understanding of it, the abstract theory of
minds is still in its infancy. Our common-sense
understanding of what it takes to be a mind is good enough
to use it as an encoding concept (i.e., to use it as a
conceptual framework that allows the interpretation of parts
of the world, as people, self, mental states, emotions,
motives, beliefs and so on). Yet, our concept of minds is
incomplete, muddled and likely inconsistent. It is possible,
however, taken enough time, effort and brain power, to
define all the aforementioned capabilities of minds down to
the nitty-gritty detail that would make it the mathematical
347
equivalent of a theory of Natural Numbers: a formal and

complete theory of what it takes to be a mind. We can
express this theory, for instance, as a computer program.
And this project is exactly what Artificial Intelligence is about.
I think this makes far more sense than physical realism, or
mind/matter dualism. It has led some thinkers to posit that the
universe is basically a giant computer. While I dont quite
embrace this myself, I do think the model of the universe as a
giant computer has a lot to teach us.
Joscha and I have argued about this extensively face-to-face,
and as I told him in our conversations, I agree with him that
insofar as its investigable using science, the universe may
be considered as consisting solely of information. The
difference between our perspectives is that I dont think all
aspects of the universe are scientifically investigable.
Science, by its nature, is about gathering finite sets of finiteprecision observations i.e. the whole corpus of scientific
knowledge consists of some finite set of bits. But theres no
reason to believe that the whole universe consists of some finite
set of bits. We may not be able to measure anything else via the
means science, but this just means that only information is
scientifically existent, not that scientific existence is the only
kind of existence!
When we last discussed the matter, he was willing to concede
this logical point, but also said (to paraphrase more or less
loosely) he felt there was no point in attempting to talk or think
about forms of existence beyond the scientifically measurable
realm. And this difference of perspective between us doesnt
affect our ability to collaborate on AGI work at all, since our AGI
engineering efforts do in fact concern the explicit manipulation of
scientifically measurable entities!
348
Quantum Informational Monism

A variant on the above perspective holds that the universe
basically consists of quantum information. This seems to be, for
example, the view of Seth Lloyd, quantum computing pioneer
and author of the fabulous book Programming the Universe. In
the same sense that informational monism deems the world a
classical computer, this perspective deems the world a quantum
computer.
In the quantum informational monist view, classical information is
an approximation to quantum information, relevant in situations
of minimal quantum coherence. From the classical informational
monist view, on the other hand, quantum theory itself is just a
computational explanation of some classical information thats
been gathered from measuring instruments (which are
themselves known via information gathered via other measuring
instruments, either artificial ones or the biological ones we call
our senses). I.e., quantum theory is a classical computer
program for getting from some classical information to some
other classical information.
Joscha Bach has summarized the informational monist view of
quantum computation as follows:
Everything that enters the human mind can be reduced to a
finite number of discernible differences. In this sense, the
most primitive encoding of the universe, as presented to an
individual observer, would be an astronomically large, finite
vector of bits. Obviously, our minds can do better than that:
much of the structure found in that vector conforms to a
three-dimensional space, filled with objects. These objects
may influence other objects in their vicinity, by touching
them, radiating upon them and so on. Thus, we arrive at
classical physics as a way of encoding the universe.
Classical physics, when expressed rigorously, is a selfcontained mathematical theory; objects in classical physics
are entities with properties and modes of interaction that are
349
reducible to calculations. As it turns out, classical physics

can not only be reduced to computation, it is in turn powerful
enough to explain how computation is possible (i.e.,
information processing in general): using classical physics,
we can conceive of a computer that can simulate classical
physics.
When we look upon our basic universe input vector more
closely, however, we find that the model of classical physics
is flawed: it is not totally consistent with the available data.
What we take to be objects does not only interact locally, but
occasionally over distances in space and time, and what we
take to be a defined state of a microscopic object can only
be described as a space of possible states, which it occupies
all at once. If we care about this discrepancy, we have to
abandon the notion of the classical universe, and adopt a
theory that accommodates the observations, i.e., quantum
mechanics.
Even the quantum mechanical universe is just our way of
encoding the finite string of bits we began with, so of course
it is computational. But quantum computation allows to
accommodate computers that perform some kinds of
computation vastly more efficient, which means that a
computer designed along classical principles might be too
slow to practically simulate a complex quantum mechanical
system.
At this point, it is not entirely clear if a computational theory
of the mind would have to be formulated along the lines of
quantum computation, or if classical computation is
sufficient. But there is practically no evidence that the
information processing performed by the neurons in our
brains would somehow crucially depend on non-local or
quantum-superpositional effects, or that people can
somehow perform computations that would require quantum
computing. Thus, even though classical computers are too
limited for detailed low-level simulations of our universe, they
350
are likely probably perfectly adequate for a detailed low-level

simulation of a mind.
For Joscha, in other words, its bits that are primary. Classical
physics is one way of explaining bits, and quantum physics is
another way of explaining bits. Either classical or quantum
physics is ultimately expressible as a mathematical construct, i.e.
as a set of formulas for manipulating symbols, used to predict
future bits from past bits. So in Joschas informational monist
view, the ultimately reality is the bits and not the classical or
quantum physical explanation of bit patterns. On the other hand,
from a quantum computationalist view, if the assumption of the
quantum model of reality is the simplest way to explain the
stream of observed bits, then this merits the assignation of some
fundamental reality to the quantum constructs involved in the
quantum model of reality.
Experiential Vitalism
The perspectives Ive considered so far in this chapter are
basically scientific in nature, focused on things like physics and
information theory. But there are also some serious AGI
researchers who take quite different views, drawing more from
the religious or spiritual side of life.
Im not a religious person by nature or culture according to
facile high-level classifications of religious belief, Ive veered
between atheist, agnostic and spiritual but not religious in
various forms at various times of my life. However, Ive always
been fascinated by religious peoples beliefs and attitudes,
feeling that they do address important aspects of the universe
that science (so far, at least) has given us little insight about.
After all, no matter how much you value math and science, they
cant tell you everything. Mathematical derivations start with
assumed axioms; and as David Hume was the first to carefully
argue) science requires some inductive bias or it cant infer
anything specific from observed data. Everybodys got to
351
assume something to get by in the world, whether they realize it

or not. So its interesting to look at the great variety of things that
different people and cultures assume.
Even those who, like Zen masters or Sufis, like to talk about
assuming nothing even these folks still implicitly act as if
theyre assuming something, as they go about their lives. When
the Zen master opens his refrigerator to get some water, he is
implicitly assuming there will be water in there instead of for
example, a lion, in which case he might bring his gun with him to
the fridge rather than a cup. And the various implicit assumptions
he makes throughout his life network together and evolve in
complex ways, just like with everybody else. Both the implicit
and the explicit assumptions people make are quite interesting.
However, I do find it interesting to understand what truly religious
people think about AGI and other transhumanist topics. In an
attempt to fulfill my curiosity on this, last year I did some H+
Magazine interviews with a few religious and spiritual types
about AGI, transhumanism and religion. I wont include the
whole interviews here because that would bring us too far afield;
but Ill give some brief highlights that may give you a sense for
how some others are thinking about this
For starters: Changle Zhou, the dean of the Cognitive Science
Department at Xiamen University where Im an adjunct
research professor, and where a team of students are working
on some OpenCog-related AI software projects is also an
experienced Zen Buddhist practitioner. He reports with a grin
that he stopped his formal Zen study some years back, when his
Zen master declared him Enlightened! Hes also the author of a
book, in Chinese, about the relation between Zen and Science. It
was both a privilege and a source of considerable entertainment
for me to interview him about his views on the relation between
AI and Zen
Consciousness, as I understand it, has three properties: selfreferential, coherent and qualia. Even if a robot becomes a
352
zombie that acts just like it possesses consciousness, it wont

really possess consciousness because the three properties of
consciousness cannot all be realized by the reductive, analytic
approach, whereas AI is based on this sort of approach.
In Zen we say that SUCHNESS is the nature of everything. For a

conscious human being, its Suchness is its consciousness; but
for material things, their Suchness is their properties such as
shape, substance or weight but not consciousness.
The Chinese word for "suchness" is "

" or "
", i.e.
"buddhahood", also known as "the foundational " or "
"
or "Alayavijnana" etc.... It is the name for the nature of
everything.
The awareness (enlightenment) of Zen is beyond all concepts,

but all the approaches we use for building robots and AI
systems, and all the behaviors of robots and AI systems, are
based on concepts
Perhaps these brief excerpts give you the essence of our
dialogue, and our disagreement. He believes that super-capable
robots and other amazing futurist technologies are likely to exist
but he sees any robots that human build, as essentially
extensions of human intelligence. He sees natural intelligences
like humans, as possessing a certain special quality Suchness
that mechanical objects like robots will never possess.
I call this sort of perspective experiential vitalism. That is: the
belief that, while the practical functions of intelligence can be
carried out by digital computers or other engineered systems,
353
theres some essential crux of consciousness that human

mind/brains possess that these other systems never will.
Cognitive Vitalism
A related perspective on AGI and the cosmos, fairly common in
the general population though not so much among scientists, is
that human brains contain some kind of special quality, which
lies outside the domain of empirical science and is responsible
for some key aspects of human intelligence.
Perhaps the most articulate advocates of this perspective that
Ive encountered id Selmer Bringsjord, an AI researcher and
logician who is also a devout Christian. Where Changle says
Suchness, Selmer says Soul but their perspectives seem
closely related. However, theres a key difference because
Selmer also argues that some functional aspects of human-level
intelligence are reliant on the Soul, which digital computers will
never possess. This, he believes, is the core reason why the AI
field has not yet led to software with human-level general
intelligence.
Selmer goes on to associate the soul with non-Turing
hypercomputation, a kind of computing that cannot be achieved
via physical computers built according to any kind of currently
known physics, and that also cannot be measured using
empirical science as currently understood. That is: Science as
we know it consists of the collection of scientific data which
consists of finite sets of bits (two other ways to phrase finite sets
of bits would be finite amounts of information or finite sets of
finite-precision numbers), and then the extrapolation of these to
predict the outcomes of future experiments, which will also take
the form of finite sets of bits. There is no scientific experiment,
conductable within the contemporary understanding or practice
of science, that would be able to identify a hypercomputable
process, or distinguish it from a conventional computing process.
In that sense, hypercomputing is a non-empirical concept
beyond the realm of measurement!
354
And here the philosophy gets tricky.

If hypercomputers are not measurable, in what sense can we
know about them at all? What sense does it make to talk about
them? Well, they can be described using mathematics!
But then, what is this mathematics? In practice its just the
marking-down of mathematical symbols on paper or computer
screens i.e. its a game played by humans with finite sets of
bits.
But in the Platonist philosophy of mathematics, mathematical
entities are envisioned to have some fundamental reality going
beyond the notations that we make to indicate them. So if one
accepts a Platonist view that math constructs have their own
reality beyond empirical science or human communication, then
one may say that hypercomputable entities exist in this abstract
Platonic math-space, and that Soul also exists in (or at least is
better represented in terms of, compared to any representation
in terms of empirical science) this abstract Platonic math-space.
And digital computers that we build, which manipulate finite bit
sets based on scientific theories inferred from scientific data
thats comprised of finite bit sets, live in a much smaller and
more impoverished region of abstract Platonic math-space, not
touching the region needed to talk meaningfully about human
Soul, or human-level intelligence.
One of the more frustrating, and gutsier, things about this
perspective is that it ultimately places the understanding of
human-level general intelligence outside the domain of science
though not necessarily outside the domain of mathematics. It
suggests that we can intuitively apprehend human-level general
intelligence using our trans-scientific hypercomputation
capabilities, but can never test these apprehensions
scientifically. It also raises the possibility that maybe we could
somehow build an AGI using intuition rather than science. I.e., if
the world has hypercomputable aspects, and our minds do also,
then maybe the hypercomputable aspects of our minds could
355
intuitively tell us how to shape the hypercomputable aspects of

the world, and we could make a physical AGI system in a
manner not determined by any finite bit sets of measurements.
But so far as I know Selmer is not trying to pursue any kind of
recondite Soul-guided AGI engineering of this nature. Rather, his
concrete AGI work is focused on the implementation of logic
systems in digital computers and on seeing exactly how far
one can push this sort of methodology before it runs into the
fundamental limits that he believes to exist.
I dont know of any simple non-technical write-up of these
interesting notions, but if you have a bit of math and computing
background, you may enjoy Selmers book Superminds: People
Harness Hypercomputation, and More.
Panpsychism
Panpsychism occurs in various forms, but in the broad sense it
refers simply to the idea that mind is a fundamental feature of the
universe and each of its parts, rather than something that is the
exclusive property of specific kinds of systems like humans,
other higher animals, intelligent computer programs, etc.
Though not a common view in contemporary Western society,
philosophy or science, panpsychism does have a long history in
historical Western philosophy, encompassing thinkers like
Leibniz, James, Whitehead, Russell, Fechner and Spinoza. A
host of recent books treat the topic, including Skrbinas Mind that
Abides: Panpsychism in the New Millienium and Strawsons
Consciousness and its Place in Nature.
Panpsychism also has a long and rich history in Eastern
philosophy, e.g. the modern Vedantic thinker Swami Krishnanda
observes
The Vedanta philosophy concludes that matter also is a phase of
consciousness and objects of knowledge embody in themselves
356
a hidden potential of consciousness which is also the Self of the

perceiving subject, enabling experience in the subject. The
subject-consciousness (Vishayi-chaitanya) is in a larger
dimension of its own being as universality and all-pervadingness
beholds itself in the object-consciousness (Vishaya-chaitanya),
thereby reducing all possible experience to a degree of universal
consciousness. Experience is neither purely subjective nor
entirely objective; experience is caused by the universal element
inherent in both the subject and the object, linking the two terms
of the relation together and yet transcending both the subject
and the object because of its universality.
Advocates of panpsychism point out that alternative theories of
mind and consciousness are riddled with problems and
inconsistencies, whereas panpsychism is simple and coherent,
its only "problem" being that it disagrees with the intuition of
many modern Western folk. Most current theories of
consciousness involve mind and awareness somehow emerging
out of non-sentient matter, which is conceptually problematic.
Philosopher Galen Strawson has recently lamented the basic
senselessness of the notion that mental experience can emerge
from a wholly non-mental, non-experiential substrate: "I think it is
very, very hard to understand what it is supposed to involve. I
think that it is incoherent, in fact..."
Dualist theories in which the mind-realm and the matter-realm
are separate but communicating also run into difficulties, e.g. the
problem that (put crudely) the mind-realm must be utterly
undetectable via science or else in effect it becomes part of the
matter-realm. Panpsychism holds that everything in the world
has mental extent, similar to how it has spatial and temporal
extent, which is a simple proposal that doesn't give rise to any
conceptual contradictions.
Some have objected to panpsychism due to the apparent lack of
evidence that the fundamental entities of the physical world
possess any mentalistic properties. However, this lack of
evidence may easily be attributed to our poor observational
357
skills. By analogy, humans cannot directly detect the

gravitational properties of small objects, but this doesn't render
such properties nonexistent. And in appropriate states of
consciousness,
humans
can
directly
apprehend
the
consciousness of objects like rocks, chairs or particles, a fact
driven home forcefully by Aldous Huxley in his classic book The
Perennial Philosophy.
Panpsychism is not without its difficulties, e.g. the combination
problem, first raised by William James which in essence
wonders: if everything is conscious, how does the consciousness
of a whole relate to the consciousnesses of its parts? How does
the brain's consciousness come out of the consciousnessess of
its component neurons, for example?
But this doesn't seem a problem on the order of "how does
consciousness emerge from non-conscious matter", it seems
more a technical issue. A large variety of qualitatively different
part-whole relationships may exists, as physicists have noted in
the last century. Quantum mechanics has made clear that
systems are not simply the sum of their parts but can sometimes
exhibit properties that go beyond those of the parts and which
cannot be detected by examining the parts in isolation. And black
hole physics has shown us the possibility of wholes (black holes)
that totally lose most of the properties possessed by their parts
and render the parts in accessible (a black hole has only the
properties of mass, charge and spin, regardless of the other
properties possessed by the objects that combined to form the
black hole). The nature of part-whole relationships in
panpsychism certainly bears further study, but merely appears
subtle, not incoherent. And the emergent, holistic aspect of
consciousness is well known in Eastern thought, e.g. Swami
Krishananda says that:
The three states of waking, dream and sleep, through which
we pass in our daily experience, differ from one another, and
yet a single consciousness connects them, enabling the
individual to experience an identity even in the otherwise
358
differentiatedness of these states. Since consciousness links

the three states into a singleness of experience, it is
immanent in them and yet transcends them, not capable of
identity with any of them.
In short, the panpsychist view of consciousness has a long
history in both Eastern and Western philosophy, and has no
glaring conceptual problems associated with it, the main difficulty
with it being that most people in contemporary Western cultures
find it counterintuitive. At least one of the authors has found it a
useful guide for thinking about the mind, perhaps largely
because it doesn't contain any confusing inconsistencies or
incoherencies that "get in the way" of analyzing other issues
involved with machine consciousness, such as reflective
consciousness, self and will.
Spiritual Monism
Panpsychism holds that everything in the cosmos has at least a
spark of mind, of cosciousness, in it. Quite often though not
always this perspective comes along with a more religious
view, which holds that everything in the cosmos has a spark of
God in it (in some sense or another). So that ultimately,
everything is part of the Mind of God. I call this spiritual
monism. Of course this may be interpreted in many, many
different ways and it would be out of place to review them all
here, but I cant resist giving one particularly fascinating
example.
Around the same time I interviewed Changle, I also did a long
and fascinating interview with Lincoln Cannon, the leader of the
Mormon Transhumanist Association. As well as advocating a
sort of Mormon spiritual monism, he advocats the intriguing
position that Mormonism is the most transhumanist religion
because it explicitly advocates human beings improving
themselves incrementally until they effectively become gods. It
also contains the idea that God used to be an ordinary being like
us, until he self-improved and became, well, transhuman
359
What I find fascinating when talking to Lincoln, is how thoroughly

hes integrated transhumanism AGI, nanotechnology, mind
uploading and all the rest into his Mormon world view. Its not
as though he keeps Mormonism and transhumanism in different
parts of his brain, one personal and one intellectual or whatever
rather, for him theyre all parts of one big conceptual complex!
And in this conceptual complex, unlike Changles, there is plenty
of room for computer programs and robots that have the same
kind of intelligence, consciousness and spirituality as humans:
Ben
Oh, and one more thing I just cant resist asking In your
perspective could an intelligent computer program have a soul?
Could it have consciousness? Could an intelligent computer
program become a God, in the same sense that a person could?
Will AIs be able to participate in the collective deification process
on the same status as humans?
Lincoln
In Mormonism, "soul" is used to describe the combination of
spirit and body, rather than just spirit. I think computer
programs already have spirits, or actually ARE spirits.
In Mormon cosmology, God creates everything spiritually before
physically, organizing and reorganizing uncreated spirit and
matter toward greater joy and glory. All things have spirits.
Humans have spirits. Non-human animals and even the Earth
have spirits, and will be glorified according to the measure of
their creation, along with us. Many Mormons also anticipate that
the day will come when, emulating God, we learn to create our
own spirit children. Spirit is in and through all things. Recall, too,
that Mormons are philosophical materialists (not dualists), so
even spirit is matter, which God organizes as the spiritual
creation of all things. So far as I'm concerned, spirit as described
by Mormonism is information, and software engineering is
spiritual creation. We are already engaged in the early stages of
the creation of our spirit children. Taking a step back, consider
360
how this adds perspective to the problem of evil: what justifies

our development of artificial intelligence in an evil world?
Cosmism
Another take on the relation of AGI to the cosmos is contained in
the broad philosophy of Cosmism a term originated by
Konstantin Tsiolokovsky and other Russian Cosmists in the late
1800s, and then borrowed by myself and Giulio Prisco in 2010 to
denote a closely related futurist philosophy, more tailored for the
modern era. Rather than positing a fundamental theory about
the stuff of the universe or the makeup of mind, Cosmism posits
an attitude toward life, technology and the world, which includes
an attitude toward AGI and the cosmos and refinement of our
understanding of their nature and relationship.
I wrote a short book in 2010 called A Cosmist Manifesto,
presenting my views on Life, the Universe and Everything, and
drawing together thoughts about AGI and other advanced
technologies with my take on Zen Buddhism and other spiritual
philosophies. As I state at the start of the book,
By Cosmism I mean: a practical philosophy focused on
enthusiastically and thoroughly exploring, understanding and
enjoying the cosmos, in its inner, outer and social aspects
Cosmism advocates
pursuing joy, growth and freedom for oneself and all

beings
ongoingly, actively seeking to better understand the

universe in its multiple aspects, from a variety of
perspectives
taking nothing as axiomatic and accepting all ideas,

beliefs and habits as open to revision based on thought,
dialogue and experience
361
Near the beginning of the Cosmist Manifesto is the following list

of high-level principles, which was initially written by Giulio Prisco
and then edited by myself:
Ten Cosmist Convictions
1. Humans will merge with technology, to a rapidly

increasing extent. This is a new phase of the evolution of
our species, just picking up speed about now. The divide
between natural and artificial will blur, then disappear.
Some of us will continue to be humans, but with a
radically expanded and always growing range of
available options, and radically increased diversity and
complexity. Others will grow into new forms of
intelligence far beyond the human domain.
2. We will develop sentient AI and mind uploading
technology. Mind uploading technology will permit an
indefinite lifespan to those who choose to leave biology
behind and upload. Some uploaded humans will choose
to merge with each other and with AIs. This will require
reformulations of current notions of self, but we will be
able to cope.
3. We will spread to the stars and roam the universe. We
will meet and merge with other species out there. We
may roam to other dimensions of existence as well,
beyond the ones of which we're currently aware.
4. We will develop interoperable synthetic realities (virtual
worlds) able to support sentience. Some uploads will
choose to live in virtual worlds. The divide between
physical and synthetic realities will blur, then disappear.
5. We will develop spacetime engineering and scientific
"future magic" much beyond our current understanding
and imagination.
362
6. Spacetime engineering and future magic will permit

achieving, by scientific means, most of the promises of
religions and many amazing things that no human
religion ever dreamed. Eventually we will be able to
resurrect the dead by "copying them to the future".
7. Intelligent life will become the main factor in the
evolution of the cosmos, and steer it toward an intended
path.
8. Radical technological advances will reduce material
scarcity drastically, so that abundances of wealth,
growth and experience will be available to all minds who
so desire. New systems of self-regulation will emerge to
mitigate the possibility of mind-creation running amok
and exhausting the ample resources of the cosmos.
9. New ethical systems will emerge, based on principles
including the spread of joy, growth and freedom through
the universe, as well as new principles we cannot yet
imagine
10. All these changes will fundamentally improve the
subjective and social experience of humans and our
creations and successors, leading to states of individual
and shared awareness possessing depth, breadth and
wonder far beyond that accessible to "legacy humans"
This is followed by a longer list of a few dozen principles and
hypotheses, which are then elaborated in more detail but, well,
if you want to know all that, just read the Cosmist Manifesto
book!
In the Cosmist, view AGI is just one (important) part of the
ongoing process of intelligence spreading itself through the
Cosmos a process that we and our intelligent software
creations are products of, agents of, and passengers upon.
In late 2010 Giulio asked me to give a talk at an online Turing
Church workshop he organized, on the Cosmist Manifesto
363
theme I spoke off the cuff as usual, but beforehand I wrote

down some notes so Id have them there to look at, just in case I
drew a blank during the actual workshop (which didnt happen,
thankfully!). Here are the notes I made, which are somewhat
similar to what I actually said:
The relation between transhumanism and spirituality is a big
topic, which I've thought about a lot -- right now I'll just make
a few short comments. Sorry that I won't be able to stick
around for this whole meeting today, I have some family stuff
I need to do, but I'm happy to be able to participate at least
briefly by saying a few remarks.
Earlier this year I wrote a book touching on some of these
comments, called "A Cosmist Manifesto" -- I'm not going to
reiterate all that material now, just touch on a few key points.
The individual human mind has a tendency to tie itself in
what the psychologist Stanislaw Grof calls "knots" intricate
webs of self-contradiction and fear, that cause emotional
pain and cognitive confusion and serve as traps for mental
energy. Ultimately these knots are largely rooted in the
human self's fear of losing itself the self's fear of realizing
that it lacks fundamental reality, and is basically a construct
whose main goals are to keep the body going and
reproducing and to preserve itself. These are some
complicated words for describing something pretty basic, but
I guess we all know what I'm talking about.
And then there are the social knots, going beyond the
individual ones The knots we tie each other up in
These knots are serious problems for all of us and they're
an even more serious problem when you think about the
potential consequences of advanced technology in the next
decade. We're on the verge of creating superhuman AI and
molecular nanotech and brain-computer interfacing and so
forth but we're still pretty much fucked up with
364
psychological and social confusions! As Freud pointed out in

Civilization and its Discontents, we're largely operating with
motivational systems evolved for being hunter-gatherers in
the African savannah, but the world we're creating for
ourselves is dramatically different from that.
Human society has come up with a bunch of different ways
to get past these knots.
One of them is religion which opens a doorway to
transpersonal experience, going beyond self and society,
opening things up to a broader domain of perceiving, being,
understanding and acting. If you're not familiar with more
philosophical side of the traditional religions you should look
at Aldous Huxley's classic book "The Perennial Philosophy"
it was really an eye-opener for me.
Another method for getting past the knots is science. By
focusing on empirical data, collectively perceived and
understood, science lets us go beyond our preconceptions
and emotions and biases and ideas. Science, with its focus
on data and collective rational understanding, provides a
powerful engine for growth of understanding. There's a
saying that "science advances one funeral at a time" i.e.
old scientific ideas only die when their proponents die. But
the remarkable thing is, this isn't entirely true. Science has
an amazing capability to push people to give up their closely
held ideas, when these ideas don't mesh well with the
evidence.
What I see in the transhumanism-meets-spirituality
connection is the possibility of somehow bringing together
these two great ways of getting beyond the knots. If science
and spirituality can come together somehow, we may have a
much more powerful way of getting past the individual and
social knots that bind us. If we could somehow combine the
rigorous data focus of science with the personal and
collective mind-purification of spiritual traditions, then we'd
365
have something pretty new and pretty interesting and

maybe something that could help us grapple with the
complex issues modern technology is going to bring us in
the next few decades
One specific area of science that seems very relevant to
these considerations is consciousness studies. Science is
having a hard time grappling with consciousness, though it's
discovering a lot about neural and cognitive correlates of
consciousness. Spiritual traditions have discovered a lot
about consciousness, though a lot of this knowledge is
expressed in language that's hard for modern people to deal
with. I wonder if some kind of science plus spirituality hybrid
could provide a new way for groups of people to understand
consciousness, combining scientific data and spiritual
understanding.
One idea I mentioned in the Cosmist Manifesto book is some
sort of "Confederation of Cosmists", and Giulio asked me to
say a little bit about that here. The core idea is obvious
some kind of social group of individuals interested in both
advanced technology and its implications, and personal
growth and mind-expansion. The specific manifestation of
the idea isn't too clear. But I wonder if one useful approach
might be to focus on the cross-disciplinary understanding of
consciousness using science and spirituality, and also
advanced technologies like neuroscience and BCI and AGI.
My thinking is that consciousness studies is one concrete
area that truly seems to demand some kind of fusion of
scientific and spiritual ideas So maybe focusing on that in
a truly broad, cross-tradition, Cosmist way could help us
come together more and over help us work together to
overcome our various personal and collective knots, and
build a better future, and all that good stuff
Anyway there are just some preliminary thoughts, these are
things I'm thinking about a lot these days, and I look forward
to sharing my ideas more with you as my thoughts develop
366
Cosmism doesnt ask us to commit to viewing the universe as

information or quantum information or hypercomputation or Godstuff or whatever it asks for a broader sort of commitment, to
an attitude of joy, growth, choice and open-mindedness.
Adoption of this attitude may lead to a variety of different
perspectives on AGI as it and related technologies evolve.
Science in its current form, and religion and philosophy in their
current forms, may turn out to be overly limited for the task of
understanding (human or artificial) mind; if so, by actively
engaging with the world and studying and engineering things,
and by reflecting on ourselves carefully and intelligently, we will
likely be able to discover the next stage in the evolution of
collective thinking
The Implicate Order

Finally, at risk of leaving you thinking Im totally nuts, Im going to
share with you some more recent thinking Ive been doing, going
beyond the ideas in the Cosmist Manifesto in a sense. These are
half-baked ideas at this stage -- but who knows, maybe some
reader will encounter them and have and publish an idea that will
help my own thinking along.
The great quantum physicist David Bohm, when he turned more
philosophical in his later years, posited the notion of the
implicate order i.e. an aspect of the universe that implicitly
underlies all things, but isnt in itself scientifically measurable or
sensorially perceptible. The explicate order that we can see and
measure, in some sense emerges from the implicate order (and
then folds back into it, contributing to it).
There are relations between this notion and panpsychism, in that
the spark of mind implicit in something may be equated to (or at
very least associated with) the implicate aspect of that thing.
Bohm also connected the implicate order with quantum
mechanics, though in a manner that I never fully understood
from his writings. Sometimes it seemed he wanted us to look at
quantum logic as a sort of interface between the implicate and
367
explicate orders. Not that the explicate order uses classical logic
and the implicate order uses quantum logic; but rather that
quantum logic captures some aspects of the explicate/implicate
interaction that classical logic misses.
Recently Ive begun thinking about the implicate order from a
different perspective, and looking at models of the implicate
order as a logic of questions rather than a logic of answers. Im
experimenting with modeling the implicate order as something I
call QP, a questioning process not a process of questioning
anything in the everyday world, but rather a process of
questioning itself. To quote a manuscript I wrote on this a while
ago (tentatively titled ?),
If I had to summarize QP in a brief phrase of (almost)
ordinary English, I suppose Id go with something like the
process of a complex, autopoietic pattern/process-system
growing and developing via self-referentially (and joyfully,
autonomously and interconnectedly) self-questioning. Sorry
if that sounds like gobbledygook! It makes a lot of sense to
me, and I hope it will make a lot of sense to you after you
finish the book!
QP The process of questioning. Questioning everything,
including the process of questioning everything and so forth!
What Ive been studying is how one might model the universe as
something that fundamentally emerges from this kind of selfquestioning process.
Another famous quantum physicist, John Wheeler, speculated
about the possibility of deriving quantum mechanics and general
relativity theory (two great physics theories that remain ununified, leaving modern physics in a state of unacceptable
contradiction) from some sort of statistical analysis of the space
of logical propositions. So that physics would emerge from a
pre-geometry made of logic. My QP approach is actually
somewhat similar, except that Im looking at a logic of questions
368
rather than a logic of answers And thinking about the

emergence of the mental as well as physical universes.
Science, after all, is a questioning process its about
questioning the universe, and about questioning science itself.
Every scientific theory brings new ideas and tools used to
question its own self, which is why every scientific theory
ultimately leads to its own destruction/transcension. The
scientific method itself is not a constant, its constantly being
revised, because it invariably leads to relentless questioning
which undermines its own foundations.
And while institutionalized religions may seem to have more to
do with obeying than with questioning, Gnostic Christianity was
precisely about trying to question and understand everything for
oneself, to know God directly. Jainist Buddhism and Sufism
shared the same aspect Jainists were trained to greet their
every thought, belief or concern with the response Not this! Not
this!
In the QP perspective I elaborate in ?, the questioning
processes of science and gnostic religion are examples and
instantiations of the broader questioning process woven into the
fabric of existence.
As I write this, Im reminded of my Hong Kong friend Gino Yu,
who likes to throw up his hands and grin and ask his friends,
gesturing at the world around, and at himself and the rest of us:
What is this?? When I mentioned QP to Gino he simply said
Oh, thats just the Socratic method.
What is this, indeed??
AGI will not tell us what is this, in any definitive way. AGI will
get rid of some of our old questions, and replace them with new
questions!
369
The Ten Cosmist Convictions describe the future of humanity

the future of the ongoing growth and unfolding of the universe,
the ramification of the Cosmos into configurations as far beyond
current human life as we are beyond the proto-life in the oceans
of early Earth, or the lifeless molecules spinning through the void
of space prior to the formation of planets. The QP train of
thought, tries to dig deeper into this growth and unfolding
process looking at the ongoing unfolding and growth of the
universe as a process of relentless self-questioning. It looks at
the universe as a big mind, which is constantly asking itself
What is this? What am I doing? and in doing so is changing
itself, bringing about new forms like planets and proto-life and
humans and AGI systems.
But please be assured: You dont have to follow me into these
peculiar thought-domains to appreciate my AGI work, or my
projections about the future of technology! Any more than you
have to agree with my panpsychist view of consciousness to
think a completed OpenCog system will be conscious you may
have your own conception of consciousness and your own way
of applying it to OpenCog.
One thing Im quite sure of is: None of us humans really knows
whats going on!
Conclusion
So, at the end of this romp through strange ideas and
perspectives, whats the take-home point about AGI and the
Cosmos?
As you already know, Im not a Mormon Im not religious at
all Im not even really a Jew, in terms of beliefs, in spite of my
Jewish racial background. And unlike my colleague Selmer, I
dont place much stock in the hypercomputable Soul.
I feel like existing religious and spiritual notions of God and
Soul are getting at important aspects of the universe which
370
science misses (and probably will always miss, due to its

foundation in data-sets comprising finite sets of bits); yet I also
feel like theyre tangled up with a lot of superstitious beliefs and
historical cruft, which makes me want to consider them more as
general inspiration than as foundations for my thought and
understanding.
I dont know whether classical or quantum information is
ultimately a better model of the universe; I think there are a lot of
unknowns thereabouts, which will be resolved as science, math
and philosophy unfold.
I strongly gravitate toward some form of panpsychism, but Im
not exactly sure what kind. When consciousness theorist Stuart
Hameroff said to me I dont think an electron has consciousness
but I think it has some kind of proto-consciousness, I felt like I
probably disagreed -- but also wondered whether we were just
getting tangled up in webs of language. I tend to agree with
Charles Pierce, Spinoza, Galen Strawson and others that
drawing a rigid distinction between mind and matter is ultimately
logically incoherent.
I find Cosmism an agreeable attitude toward myself and the
universe, but Im also aware of its limitations. I cant help
questing toward a more fundamental understanding of the
Cosmos, even though I suspect my limited human brain isnt
going to be able to understand the Cosmos too well, and that
AGIs, brain enhancement and the like will give us much deeper
and better perspectives. Im currently somewhat bemused by
modeling the universe as a process of self-questioning, but
unsure how far Ill be able to take the idea.
Often I find myself holding back the part of my mind that wants to
spend a lot of time theorizing about such things, because I feel
its more important to work on building AGI! Understanding the
world is important to me, but the choice is between trying to
understand it directly using my human brain, or building an AGI
371
mind that will be able to help me understand it far better than my

human brain is capable of!
One thing that struck my mind when writing this chapter, was the
old saw: People can adapt to almost anything It occurred to
me that it makes sense to generalize this to: Robust conceptual
frameworks can adapt to almost anything. For example,
perspectives like Mormonism and Buddhism and panpsychism
can adapt themselves to make meaningful statements about AGI
and Singularity, notions that were completely unanticipated (and
would have been largely incomprehensible) at the time these
religions and philosophies were founded. This says a lot about
the adaptable nature of human mind and culture (which
emanates, of course, from the adaptable nature of the biological
world from which these emerged).
Ultimately, it seems, human conceptual and belief systems are
able to adapt to all sorts of new realities, including the Internet
and mobile phones and birth control, and soon including AGI and
molecular assemblers and cyborgs and what-not. Spirituality and
religion embody key aspects of human nature, such as the quest
for deep fundamental understanding of the universe, and deep
communion with the universe and with other minds and these
quests will keep on going as technology advances, not so much
colliding with scientific and technological growth as synergizing
with it. Technological revolution will foster a continuous unfolding
and expansion of spiritual understanding and experience, which
may end up taking a tremendous diversity of different forms far
stranger (and perhaps far deeper) than wild ride of ideas weve
rolled through in this chapter.
And so, as I already emphasized, I can make one statement
about the contents of this chapter with great confidence: Every
single idea posited here will appear to us rather silly and
limited, once we have expanded our world view via intense
communication and possible fusion with trans-human AGI
systems. In other words: Where the cosmos is concerned, we
humans dont understand much! We understand more than non-
372
human animals do, and more than pre-civilized humans did

But the AGIs we create will understand much, much more.
Which is not to say that any mind will ever come to a complete
and thorough understanding of intelligence and the cosmos
maybe it will, maybe it wont; thats among the many things
humanity currently has no clue about.
Just as the contradiction between quantum theory and general
relativity (gravity theory) tells us that current physics MUST be
seriously incomplete in some ways so, I feel, does the human
races wild diversity of confusingly contradictory and
complementary views on the cosmos tell us that we really dont
understand this world were living in, and our intelligences place
in it, and the roles and properties that will be played by AGIs we
create. We dont necessarily understand what our current
activities are going to lead to, any better than the cavemen who
first created language understood that their invention was going
to lead to Dostoevsky, differential calculus, Prolog, Google Talk
and Neuroscience for Dummies.
The various concepts and perspectives were currently
experimenting with (and in some cases pouring our hearts and
minds into) classical and quantum information theory,
hypercomputing, implicate orders, Mormonism and Buddhism
and Cosmism and what-not all exist within the limited scope of
current human individual and collective mind. These
perspectives may help us cope with the changes that we are
now wreaking in ourselves and the world, via creating AGI and a
host of other technologies; and they may well affect the particular
nature of the technologies we create and the future world these
technologies help create. But these perspectives will then be
subverted, and made to seem quaint and ridiculous, by the
greater intelligence to which they will lead. And yet, we have to
work on refining and elaborating our current perspectives on the
world knowing they will seem absurd and limited in hindsight,
from the perspective of greater intelligence because this
refinement and elaboration is part of the process of bringing
about said greater intelligence.
373
374

Ten Years To The Singularity

Uploaded by

Copyright:

Available Formats

Ten Years To The Singularity

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ten Years To The Singularity

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What are some of the challenges discussed regarding developing AGI?

What are some of the challenges discussed regarding developing AGI?

Ten Years

Copyright 2014 Ben Goertzel

Objection 7: Human-Level Intelligence May Require Quantum (Or More

Other AGI Scenarios................................................................................... 233!

Quantum Informational Monism ..................................................................349!

Ten Years to the Singularity If We Really Really Try

Worth noting, perhaps, is that one thing this book DOESNT

of notes that was released posthumously under that title),

Next Steps For And Beyond Humanity

However, all the change weve seen so far is trivial compared to

Ten Years to the Singularity If We Really Really Try

Next Steps For And Beyond Humanity

was pleased to have a brief appearance in the Ray Kurzweil

Ten Years to the Singularity If We Really Really Try

there will be new categories of technology that we cant now

Ten Years to the Singularity If We Really Really Try

with greater than human intelligence, which can then work

The Top Priority for Mankind

Acceleration of Biomedical Research

Ten Years to the Singularity If We Really Really Try

the point of economic practicality has proved a slow one,

The Top Priority for Mankind

this research direction seems to be the understanding of

approach. Reverse engineering the brain is a process that will

Ten Years to the Singularity If We Really Really Try

extensive, built-in capabilities. These would have to include

the language processing AI underlying search engines

the planning and scheduling AI used extensively in

grandmaster-beating chess programs

video game AI characters, playing a central role in

hundreds of millions of people

AI financial trading systems

AI data mining systems helping businesses, scientists

However, these narrow-AI programs have not progressed that

Ten Years to the Singularity If We Really Really Try

influenced very substantially by the names used to refer to

Why the Time is Ripe for AGI

In the big picture, it definitely doesnt bother me at all. The vast

2. Computer science has advanced greatly, providing all

Ten Years to the Singularity If We Really Really Try

groups of people (email lists, wikis, Web repositories of

bits and pieces that have each been written by a different

Ten Years to the Singularity If We Really Really Try

humanoids with various technical advantages over the current

non-invasive brain imaging technology that can measure the

Ten Years to the Singularity If We Really Really Try

a diagram explaining how the human mind works, and showing

better tools embodying richer understanding. 20 or 30 years from

Ten Years to the Singularity If We Really Really Try

The Internet, as it stands, is nowhere near as great as it could be

Ten Years to the Singularity If We Really Really Try

Why So Little Work on AGI?

2. Of those that do, 80% dont believe its possibleeither

Ten Years to the Singularity If We Really Really Try

Narrow AI gets plenty of money, in forms like Googles and