Hans Moravec

CHILDREN^
The Future of Robot and *"
Human Intelligence
 Digitized by the Internet Archive
in 2010

http://www.archive.org/details/mindchildrenfutuOOmora
Mind Children
MIND CHILDREN
The Future of Robot
and Human Intelligence

Hans Moravec

HARVARD UNIVERSITY PRESS

Cambridge, Massachusetts
London, England
Copyright © 1988 by Hans Moravec
All rights reserved
Printed in the United States of America
10 98765432
Library of Congress Cataloging-m-Publication Data

Moravec, Hans P.

Mind children : the future of robot and human intelligence /

Hans Moravec.
p. cm.
Bibliography: p.
Includes index.
ISBN 0-674-57616-0 (cloth)
ISBN 0-674-57618-7 (paper)
1. Artificial intelligence. 2. Robotics. I. Title.

Q335.M65 1988 88-21343

006.3 — dcl9 CIP
In memory of my father, who taught me to tinker

To my mother, ivho taught me to read

To Ella, who made me complete

 Contents

Prologue 1

1 Mind in Motion 6

2 Powering Up 51

3 Symbiosis 75

4 Grandfather Clause 100

5 Wildlife 125

6 Breakout 147

Appendix 1 Retinas and Computers 163

Appendix 2 Measuring Computer Power 169

Appendix 3 The Outer Limits of Computation 178

Bibliography 197

Acknowledgments and
Illustration Credits 203

Index 207
Figures

Intelligence on Earth 18

Walking Machine 27
Five Legs 28
Three Fingers 30
Autonomous Navigation 33
Object Finding 35
A General-Purpose Robot 38
The Retina 54
Comparative Computational Power and Memory 61

A Century of Computing 64

ENIAC 76

Magic Glasses (Early Model) 87

Robot Proxy 88
Unreal Estate —The Road to Point Reyes 92

A Robot Bush 103

The Corpus Callosum 113

Computer Virus Blowup 130

Selfish Martians 142

Spacetime Pyramid 156

Large Spacetime Pyramids from Small 157

I Think, Therefore I Am 180

Two-Slit Experiment 184
Two Slits and Waves 185
Robot Pals 202
Mind Children
 Prologue

spiraling
E NGAGED for billions of years in a relentless,

arms race with one another, our genes have finally out-
smarted themselves. They have produced a weapon so powerful it
will vanquish the losers and winners alike. This device is not the
—
hydrogen bomb widespread use of nuclear weapons would merely
delav the immensely more interesting demise that has been engi-
neered.
What awaits is not oblivion but rather a future which, from our
present vantage point, is by the words "postbiological"
best described
or even "supernatural." It which the human race has
is a world in
been swept away by the tide of cultural change, usurped by its own
artificial progeny. The ultimate consequences are unknown, though

many intermediate steps are not only predictable but have already
been taken. Today, our machines are still simple creations, requiring
the parental care and hovering attention of any newborn, hardly
worthy of the word "intelligent." But within the next century they
will mature into entities as complex as ourselves, and eventually into
something transcending everything we know in whom we can take —
pride when they refer to themselves as our descendants.
Unleashed from the plodding pace of biological evolution, the
children of our minds will be free to grow to confront immense and
fundamental challenges in the larger universe. We humans will benefit
for a time from their labors, but sooner or later, like natural children,
they will seek their own fortunes while we, their aged parents, silently
fade away. Very little need be lost in this passing of the torch — it will
be in our artificial offspring's power, and to their benefit, to remember
almost everything about us, even, perhaps, the detailed workings of
individual human minds.
 Mi)id Children

The process began about 100 million years ago, when certain gene
lines hitupon a way to make animals with the ability to learn some
behaviors from their elders during life, rather than inheriting them
all at conception. It was compounded 10 million years ago when

our primate ancestors began to rely on tools made of bones, sticks,

and stone, and accelerated again with the harnessing of fire and the
development of complex languages about 1 million years ago. By the
time our species appeared, around 100 thousand years ago, cultural
evolution, the juggernaut our genes had unwittingly constructed, was
rolling with irresistible momentum.
Within the last 10 thousand years, changes within the human gene
pool have been inconsequential in comparison with the snowballing
advances in human culture. We have witnessed first an agricultural
revolution, followed by the establishment of large-scale bureaucratic
governments with the power to levy taxes for their support, the
development of written languages, and the rise of leisure classes

with time and energy to devote to intellectual concerns. In the

last thousand years or so, inventions beginning with movable type
printing have greatly speeded the flow of cultural information, and
thus its evolutionary pace.
With the coming of the industrial revolution 200 years ago, we
entered the final phase, one in which artificial substitutes for human
body functions such as lifting and transporting have become ever more
economically attractive — indeed, indispensable. Then, 100 years ago,
with the invention of practical calculating machines, we were able for
the first time to artificially duplicate some small but vexing functions
of the human mind. The computational power of mechanical devices
has risen a thousandfold every 20 years since then.
We are very near to the time when virtually no essential human
function, physical or mental, will lack an artificial counterpart. The
embodiment of this convergence of cultural developments will be
the intelligent robot, a machine that can think and act as a human,
however inhuman it may be in physical or mental detail. Such
machines could carry on our cultural evolution, including their own
construction and increasingly rapid self-improvement, without us, and
without the genes that built us. When that happens, our DNA will

find itself out of a job, having lost the evolutionary race to a new kind
of competition.
 Prologue

A. G. Cairns-Smith, a chemist who has contemplated the beginnings

of Hfe on the early earth, calls this kind of internal coup a genetic
takeover. He suggests that it has happened at least once before. In
Seven Clues to the Origin of Life, Cairns-Smith argues that the precur-
sors to life as we know it were microscopic crystals of clay, which
reproduced by the simple process of crystal growth. Most crystals
are marked by patterns of dislocation in the orderly arrangement of
their atoms, many of which propagate as the crystal grows. If the
crystal should fracture, each piece may inherit a copy of the pattern,
sometimes with a slight change. Such defects can have a dramatic
effect on and chemical properties. Crystals sharing
a clay's physical
one dislocation pattern may form dense clumps, while those with an-
other mav aggregate into a spongy mass. Mineral-bearing water may
be diverted around one type but trickle through the other, providing
raw materials for continued growth. The patterns also affect growth
indirectly by modulating the chemistry of other molecules in their
environment. Clays are powerful chemical catalysts; the tiny crystals
have enormous which molecules can adhere in
total surface area, to

certain configurations, depending on the external shape of the crystal

and molecule in question. These common crystals thus possess the
essentials for Darwinian evolution —
reproduction, inheritance, muta-
tion, and differences in reproductive success.

In Cairns-Smith's theory, the first genetic takeover began when some

clay species, in vigorous Darwinian competition with one another,
began to encode some genetic information externally in long carbon
molecules. Such polymers are more stable than the easily disturbed
dislocation patterns themselves, and organisms using them to ever
greater extent reproduced more successfully. Although utterly de-
pendent at first on the existing crystal-based chemical machinery, as
these carbon molecules assumed a greater share of the reproductiv^e
role they became less reliant on the crystals. In time, the simple
crystal scaffolding vanished altogether, leaving in its evolutionary
wake the complex, interdependent system of organic machinery we
call life.

Today, billions of years later, another change is under way in

how information passes from generation to generation. Humans
evolved frc^m organisms defined almost totally by their organic genes.
We now rely additionally on a vast and rapidly growing corpus
 Mind Children

of cultural information generated and stored outside our genes in —

our nervous systems, libraries, and, most recently, computers. Our
culture still depends utterly on biological human beings, but with each
passing year our machines, a major product of the culture, assume
a greater role in its maintenance and continued growth. Sooner or
later our machines will become knowledgeable enough to handle
their own maintenance, reproduction, and
self-improvement without
help. When this happens, the new genetic takeover will be complete.
Our culture will then be able to evolve independently of human
biology and its limitations, passing instead directly from generation
to generation of ever more capable intelligent machinery.

Our biological genes, and the flesh and blood bodies they build,
will play a rapidlv diminishing role in this new regime. But will our
minds, where culture originated, also be lost in the coup? Perhaps
not. The coming revolution may liberate human minds as effectively
as it liberates human culture. In the present condition we are un-
comfortable halfbreeds, part biology, part culture, with many of our
biological traits out of step with the inventions of our minds. Our
minds and genes may share many common goals during life, but there
is a tension between time and energy spent acquiring, developing,

and spreading ideas and effort expended toward maintaining our

bodies and producing a new generation (as any parent of teenagers
can observe). The uneasy truce between mind and body breaks down
completely as life ends. Our genes usually surv'ive our death, grouped
in different ways in our offspring and our relatives. In a subtle way
it is no doubt in their evolutionary interest to regularly experiment
like this with fresh shuffles of the genetic deck. But the process is

devastating for our other half. Too many hard-earned aspects of our
mental existence simply die with us.
It is easy to imagine human thought freed from bondage to a mortal

body — belief in an afterlife is common. But it is not necessary to adopt

a mystical or religious stance to accept the possibihty. Computers
provide a model for even the most ardent mechanist. A computa-
tion in progress —
what we can reasonably call a computer's thought
—
process can be halted in midstep and transferred, as program and
data read out of the machine's memory, into a physically different
computer, there to resume as though nothing had happened. Imagine
that a human mind might be freed from its brain in some analogous
(if much more technically challenging) way.
 Prologue

A mind would require many modifications to operate effectively

after being rescued from the limitations of a mortal body. Natural
human mentality is tuned for a life span's progression from impres-
sionable plasticity to self-assured rigidity, and thus is unpromising
material for immortality. Itwould have to be reprogrammed for
continual adaptability to be long viable. Whereas a transient mortal
organism can leave the task of adaptation to the external processes of

mutation and natural selection, a mind that aspires to immortality,

whether it traces its beginnings to a mortal human being or is a
completely artificial creation, must be prepared to adapt constantly
from the inside.
Perhaps it would undergo a cyclical rejuvenation, acquiring new

hardware and software in periodic phases that resemble childhood. Or

maybe it could update the contents of its mind and body continuously,
adding and deleting, testing components in all kinds of combinations,
to keep up with changing conditions. The testing is of central
importance: it steers the evolution. If the individual makes too many
bad decisions in these tests, it will fail totally, in the old-fashioned
Darwinian way.
A postbiological world dominated by self-improving, thinking ma-
chines would be as different from our own world of living things as
this world is different from the lifeless chemistry that preceded it. A
population consisting of unfettered mind children is quite unimag-
inable. We are going to try to imagine some of the consequences
anyway.
 1 Mind in Motion

I BELIEVE that robots with human intelHgence

will be common within fifty years. By comparison, the best of today's
machines have minds more like those of insects than humans. Yet
this performance itself represents a giant leap forward in just a few
decades.
Mechanical imitations of certain human functions have been with
us for centuries. Many medieval clock towers are equipped with
mechanisms that mark the hours with elaborate morality plays enacted
by mechanical saints, knights, bishops, angels, demons, and all kinds
of animals. Smaller devices that walked, talked, swam, breathed, ate,
wrote with quill pens, or played musical instruments have amused
polite society since at least the fifteenth century. Leonardo da Vinci,
for one, constructed elaborate mechanical displays of this sort for
his patrons. These early clockwork machines — powered by running
water, falling weights, or springs — copied the motions of living things,
but they could not respond to the world around them. They could
only act, however charmingly.
Electrical, electronic, and radio technology, developed early in this
century, made possible machines that could react — to light, sound,
and invisible remote control. The result was a number of entertaining
demonstration robots — as well as thoughts and stories about future
humanlike machines. But only simple connections between the sen-
sors and motors were possible at first. These newer machines could
sense as well as act, but they could not think.

Machines That Think (Weakly)

During World War II analog computers — machines that simulated

physical systems by representing their changing quantities as analo-

 Mind in Motion

gous mo\es oi shafts or voltages — were designed for controlling anti-

aircraft guns, for navigating,and for precision bombing. Some of their
developers noticed a between the operation of these devices
similarity'

and the regiilator\- svstems in living things, and these researchers

were inspired to build machines that acted as though they were alive.
Norbert Wiener at the Massachusetts Institute of Technology (MIT)
coined the term cvbemetics for this unified studv of control and com-
munication in animals and machines. Its practitioners combined new
theorv on feedback regulation with advances in postwar electronics
and earlv kno\vledge of living nervous svstems to build machines that
were able to respond like simple animals and to learn. The rudiments
of artificial thought had arrived.
Among the highUghts oi the cvbemetics effort was a series of
electronic turtles built during the 1950s bv \V. Grev Walter, a British
psvchologist. With subminiature radio-tube electronic brains, rotating
phototube eves, microphone ears, and contact-switch feelers, the first
versions could locate their recharging hutch when their batteries ran
low and other^vise a\"oid trouble \s'hile wandering about. Groups
of them exhibited complex social behavior bv responding to one
another's control lisihts and touches. A later machine with the same
senses could be conditioned to associate one stimulus with another
and could learn bv repeated experience that, for instance, a loud
noise would be followed bv a kick to its shell. Once educated, the
turtle would avoid a noise as it had before responded to a kick.
The associations were slowlv accumulated as electrical charges in
electronic devices called capacitors, used here as memory- devices.
Perhaps the most impressive creation of the cvbemeticists was
the Johns Hopkins Beast. Built bv a group oi brain researchers
in the earlv 1960s, it wandered the halls guided bv sonar and a

specialized photocell eve that searched for the distinctive black cover
plate of wall outlets, where it would plug itself m to feed. The Beast
inspired a number of imitators. Some used special circuits connected
to televisioncameras instead of photocells and were controlled bv
assemblies oi (then new) transistor digital logic gates, Uke those that
can now be found, in thousands and millions, in the integrated circuits
of ever\- computer Some added new motions such as "Shake to
untangle recharging arm" to the repertoire of basic actions.
The field of cybernetics thrived less than two decades. As is so
often the case, it was eclipsed bv a relative, the artificial intelligence
 Mind Children

movement. The war's many small analog computers, which had

inspired cybernetics, had a few much larger digital cousins. These
machines computed not by the measured turns of shafts or flow
of current but by counting, in discrete jumps. The first automatic
digital computers — huge, immobile, autonomous calculators — were
completed toward the end of the war. Colossus, an ultrasecret British
machine that broke the German Enigma code and helped to change
the course of the war, scanned code keys tens of thousands of times
faster than humanly possible. In the United States, ENIAC computed
antiaircraft artillery tables for the Army and later did calculations for
the construction of the atomic bomb, at speeds similar to Colossus.
Less belligerently, these "giant brains," as they came to be called,
provided unprecedented opportunities for experiments in complexity.

Pioneers like Alan Turing, one of the creators of Colossus, and John
von Neumann, who was involved with the first American machines,
harbored the hope that the ability to think rationally, our unique
asset in dealing with the world, could be captured in a machine.
Our minds might be amplified by computers just as our muscles
had been amplified by the steam engines of the industrial revolution.
Programs to reason and to play intellectual games like chess were
designed by Claude Shannon of MIT and by Turing in 1950, but the
earliest computers were too limited and expensive for this use. A few
poor checker-playing programs did appear on the first commercial
machines of the early 1950s, and equally poor chess programs showed
up in the last half of that decade, along with a good checker player
by Arthur Samuel of IBM. Then in 1957 Allen Newell and Herbert
Simon of Carnegie Tech (now Carnegie Mellon University) and John
Shaw of the RAND Corporation demonstrated the Logic Theorist, the
first program able to reason about arbitrary matters by starting with

axioms and applying rules of inference to prove theorems.

In 1960 John McCarthy, then at MIT, coined the term artificial
intelligence (AI) for the effort to make computers think. By 1965 the

first students of McCarthy, Marvin Minsky (also at MIT), Newell, and

Simon had produced AI programs that proved theorems in geometry,

solved problems from intelligence tests, algebra books, and calculus
exams, and played chess, all with the proficiency of an average college
freshman. Each program could handle only one narrow type of
problem, but for first efforts these programs were encouraging so —
encouraging that most people involved felt that another decade of
 Mind in Motion

progress would surely produce a genuinely intelligent machine. This

was an understandable miscalculation.
Now, a quarter of a century later, computers are thousands of times
more powerful than these sixties models, but they do not seem much
smarter. Bv 1975 progress in artificial intelligence had slowed from
the heady sprint of a handful of enthusiasts to the plodding trudge of
growing throngs of workers. Even so, modest successes have main-
tained flickering hope. So-called expert systems, programs encoding the
decision rules of human experts in narrow domains such as diagnosis
of disease, factory scheduling, or computer system configuration, are
dailv earning their keep in the business world. A fifteen-vear effort at
MIT gathered knowledge about algebra, trigonometry, calculus, and
related fields into a wonderful program called MACSYMA, now mar-
keted commercially, that manipulates svmbolic formulas and helps to
solve otherwise forbidding problems. Several chess-plaving programs
are now officially rated as chess masters, and excellent performance
has been achieved in other games like backgammon. There are
semi-intelligent programs that can understand simplified tvpewritten
English about restricted subjects and make elementary deductions in
the course of answering questions about these texts. Some interpret
spoken commands chosen from thousand-word repertoires, and others
can do simple visual tasks, such as deciding whether a part is in its
desired location.
Unfortunately for humanlike robots, computers are at their worst
trying to do the things most natural tohumans, such as seeing,
hearing, manipulating objects, learning languages, and commonsense
reasoning. This dichotomy — machines doing well things humans find
hard, while doing poorly what is easy for us —is a giant clue to the

problem of how to construct an intelligent machine.

Machines That See (Dimly) and Grasp (Clumsily)

In the mid-1960s Marvin Minsky's students at MIT began to connect

television cameras and mechanical robot arms to their computers,
giving eyes and hands to artificial minds so that their machines could
see, plan, and act. By 1965 these researchers had created a machine
that could and remove white blocks from a black tabletop.
find
This accomplishment required a controlling program as complex as

—
any of the then-current pure reasoning programs programs which.
10 Mind Children

unencumbered by robot appendages, could, for instance, match first-

year college students in solving calculus problems. Yet Minsky's

hand-eye system could be bested by a toddler. Nevertheless, the
experiments continued at MIT and elsewhere, graduallv developing
into a field which now goes by the name robotics, a term coined in
1942 in a science fiction story by Isaac Asimov from the word robot,
itself coined by the Czech playwright Karel Capek in 1921. Robotics

started far lower on the scale of human performance than artificial

intelligence, but its progress in the past twentv vears has been just as
agonizingly slow and difficult.

Not all robots, nor all people, idle away their lives in universities.
Many must work for a living. Even before the industrial revolution,
before any kind of thought was mechanized, partially automatic
machinery, powered by wind or flowing water, was put to work
grinding grain and cutting lumber. The beginning of the industrial
revolution in the eighteenth century was marked bv the invention
of a plethora of devices that could substitute for manual labor in a
precise, and thoroughly inhuman, wav. Driven by shafts turned by
water or steam, these machines pumped, pounded, cut, spun, wove,
stamped, moved materials and parts, and much else, consistently and
tirelessly.

Once in a while something ingeniously different appeared. For

instance, the Jacquard loom, invented in 1801, could weave intricate
tapestries specified by a string of punched cards. By the early
twentieth century, electronics had given industrial machines limited
senses; they could now stop when something went wrong, or control
the temperature, thickness, even consistency of their workpieces. Still,

each machine did one job and one job onlv Consequentiv, the product
produced by the machine often became obsolete before the machine
had paid back its design and construction costs. This problem had
become particularly acute by the end of World War II.
In 1954 the inventor George Devol filed a patent for a new kind of
industrial machine, the programmable robot arm, whose task could
be altered simply by changing the stream of punched program cards
that controlled its movement. In 1958, with Joseph Engelberger, Devol
founded a company named Unimation (a contraction of "universal"
and "automation") to build such machines. The punched cards
soon gave way to magnetic memory, which allowed the robot to be
programmed simply by leading it once by the hand, so to speak.
 Mind in Motion 11

through its required paces. The first industrial robot began work in
a General Motors plant in 1961. To this day, most large robots that
weld, spray paint, and move pieces of cars are of this type.
Only when the cost of small computers dropped to less than $10,000
did robotics research at the universities begin to influence the design of
industrial robots. The first industrial vision systems, able to locate and
identify parts on conveyor belts, and usually coupled with a new class
of small robot arms, appeared in the late 1970s. Robots able to see and
feel, after a fashion, now play a modest but quietly booming role in
the assembly and inspection of small devices like calculators, printed
circuit boards, typewriters, and automobile water pumps. Indeed,
industrial needs have strongly influenced university research. What
was once a negligible number of smart robot projects has swelled to
the hundreds on campuses across the country. And while cybernetics
may now be relatively dormant, its stodgy parent, control theory, has
been quite active since the war in an effort to meet the profitable needs
of the aerospace industry. Elaborate methods developed to control
aircraft, spacecraft, and weapons are now influencing the design of

industrial robots.
In 1987 I was treated to a tour of the factory in Fremont, California,
where Apple's Macintosh computers are assembled. I found most
of the plant well organized but unremarkable. Many assembly steps
were done manually. The most efficient machines were probably those
that inserted components into circuit boards. Acting something like
sewing machines, these "board stuffers" take components strung on
tapes like machine-gun ammunition and "stitch" them into printed
circuit boards at blinding speed, several components per second, with
the board sliding rapidly into position for each part, front and back,
left and right. The machines are marvels of computerized control,
and very cost-effective for high-volume production, but they left me
vaguely disappointed. In one small niche, however, I saw a quite
different device inserting components that the high-speed machines
could not handle. The were old-fashioned inductors small
parts —
metal cans containing a coil of wire. Each can had metal tabs that were
to fit into slots in the board, and the coils ended in wires intended for
small holes. Unlike the precisely shaped components on the feed tapes
of the other machines, the inductors, supplied neatly arrayed on tex-
tured plastic trays, often had slightly bent tabs and wires that would
simply be mangled by a blind attempt to push them into a board.
12 Mind Children

The insertionmachine worked in a glass booth. Boards and trays

of inductors arrived and left by conveyor belts. The insertion process
began with a TV camera that looked down on the parts tray. A vision
program located an inductor, and a fast robot arm swooped to pick
it up and then brought it, wire and tab end up, in front of a second

TV camera. A second vision program examined the leads and, if they

were straight enough, directed the arm to insert the component in the
board. If the leads were slightly bent, the inductor was first pushed
against a "pin straightener," a metal block with tapered holes, after
which the pins would be inspected again. If the leads were hopelessly
mangled, the inductor was dropped into a reject bin and another was
fetched from the tray.
The insertion was itself a sensitive process. The inductor was
rapidly brought to within a few millimeters of the board surface, then
slowly lowered until the robot arm encountered resistance. The arm
nudged the inductor to and fro while maintaining a slight downward
pressure, until the tabs and wires found their holes, at which point it
applied greater pressure to seat the component. A motorized cutter
mounted below the board then cut and bent the protruding metal,
anchoring the inductor. If the attempt to seat the part failed after a
few seconds, it would again be brought in front of the lead-checking
camera, and possibly into the pin straightener, before another insertion
attempt. a third attempt failed, the part would be tossed into the
If

reject bin and a new one tried.

All this happened very rapidly —
a part could be inserted every three
or four seconds, though a troublesome one might take up to ten. I
—
was impressed and a little nostalgic. The activities had a familiar
feel. I had been a regular witness to somewhat slower and clumsier

versions of them a decade earlier at the Stanford Artificial Intelligence

Lab, where I was a graduate student. In fact, Apple's assembly system
was a product of a small southern California company called Adept
that can trace its ancestry back to PhD theses at SAIL. The seeds cast
there were starting to sprout.
The goal humanlike performance by stationary robots, though
of
highly diluted by a myriad of approaches and short-term goals,
has acquired a relentless. Darwinian vigor. As a story, it becomes
bewildering in its diversity and interrelatedness. Let us move on to
the sparser world of robots that rove.
 Mind in Motion 13

Machines That Explore (Haltingly)

The first reasoning programs needed very little data to do their work.

A chessboard, or a problem in logic, geometry, or algebra, could be

described by a few hundred well-chosen symbols. Similarly, the rules
for solving the problem could be expressed as several hundred so-
The difficulty lay only in finding
called "transformations" of this data.
a sequence of transformations that solved the problem, from among
the astronomical number of possible combinations. It was obvious that
solving problems in less restricted areas (the question "How can I get

to Timbuktu?" was an often-used rhetorical example) would require

a much greater initial store of information. It seemed unlikely that all

the facts needed to solve such problems could be provided manually

to reasoning programs.
Some facts might be made available if the programs could be
taught to read and understand books, but comprehending even simple
words would require detailed knowledge of the physical world. Such
knowledge is assumed to preexist in the minds of book readers no —
book attempts a comprehensive definition of a rock, a tree, the sky, or
a human being. Possibly some of this zuorld knowledge, as it has come
to be called, could be obtained by the machine itself if it could directly
observe its surroundings through camera eyes, microphone ears, and
feeling robot hands. The desire to automate the acquisition of world
knowledge was one of the early rationales for robotics projects in the
artificial intelligence labs. The internal model of the world that these
computers might develop could then be used by them, and by other
machines, as a basis for commonsense reasoning.
Although a machine that can move around should be able to gather
much more data than an immobile one, the logistical difficulty of
connecting a huge immobile computer to a complicated array of
sensors on a moving platform made fixed hand-eye systems more
attractive to most researchers. Besides, it was soon realized that the
problem of systematically acquiring information from the environment
was much less tractable than the mental activities the information
was intended to serve. During the 1970s dozens of research labs
had robot arms connected to computers, but hardly any had robot
vehicles. Most robotics researchers viewed mobility as an unnecessary
complication to an already overly difficult problem. Their experience
14 Mind Children

was in marked contrast with the attitude of the cyberneticists (and

hundreds of hobbyists and toymakers), who had been quite satisfied
with ehciting simple animal-hke behavior from the modest circuitry
aboard their many small mobile machines.
Stanford Research Institute's Shakey, completed in 1969, was the
first,and is still the only, mobile robot to be controlled primarily
by programs that reasoned. It is an exception that proves the rule.
—
Shakey's instigators Nils Nilsson, Charles Rosen, and others were —
inspired by the early success in artificial intelligence research. They
sought to apply logic-based problem-solving methods to a real-world
task involving a machine that could move and sense its environment.
The problems of controlling this movement and interpreting sensory
data were of secondary importance to the designers, however, whose
main interest was in the machine's ability to reason. The job of
developing the mobility and sensory software was relegated to junior
programmers.
Five feet tall and driven by two slow but precise stepping motors,
Shakey was equipped with a television camera and was remote-
controlled by a large computer. Methods from MIT's blocks manipu-
lating programs, previously used only with robot arms, were adapted
for interpreting the TV images. These worked only when the scene
consisted solely of simple, uniformly colored, flat-faced objects, so a
special environment was constructed for the robot. It consisted of
several rooms bounded by clean walls, containing a number of large,
uniformly painted blocks and wedges. Shakey's most impressive
performance, executed piecemeal over a period of days, was to solve a
so-called "monkeys and bananas" problem. Told to push a particular
block that happened to be resting on a larger one, the robot constructed
and acted on a plan that included finding a wedge that could serve as
a ramp, pushing it against the large block, driving up the ramp, and
delivering the requested push.
The environment and the problem were contrived, but they pro-
vided a motivation, and a test, for a clever reasoning program called

STRIPS (the STanford Research Institute Problem Solver). Given a task

for the robot, STRIPS assembled a plan out of the limited actions the
robot could take, each of which had preconditions (for example, to
push a block, it must be in front ofme) and probable consequences
(after I push a block, it is moved). The state of the robot's world
was represented in sentences of mathematical logic, and formulating
 Mind in Motion 15

a plan was like proving a theorem, the initial state of the world being
the axioms, primitive actions being the rules of inference, and the
desired outcome playing the role of the theorem. One complication
was immediately evident: the outcome of an action is not always what
one expects (as when the block does not budge). Shakey had a limited
ability to handle such glitches by occasionally observing parts of the
world and adjusting its internal description and replanning its actions
if the conditions were not as it had assumed.

Shakey was impressive in concept but pitiable in action. Each move

of the robot, each glimpse taken by its camera, consumed about an
hour of computer time and had a high likehhood of failure. The block-
pushing exercise described above was staged, and filmed, in steps,
with several steps requiring repeated "takes" before they succeeded.
The fault lay not in the STRIPS planner, which produced good plans
when given a good description what was around the robot, but in
of
the programs that interpreted the raw data from the sensors and acted
on the recommendations.
It seemed to me, in the early 1970s, that some of the creators of

successful reasoning programs suspected that the poor performance

in the robotics work somehow reflected the intellectual abilities of
those attracted to that side of the research. Such intellectual snobbery
is not unheard of, for and experimentalists
instance between theorists
in physics. But as the number mounted, it has
of demonstrations has
become clear that it is comparatively easy to make computers exhibit
adult-level performance in solving problems on intelligence tests or
playing checkers, and difficult or impossible to give them the skills of
a one-year-old when it comes to perception and mobility.
In hindsight, this dichotomy is not surprising. Since the first multi-
celled animals appeared about a billion years ago, survival in the fierce

competition over such limited resources as space, food, or mates has

often been awarded to the animal that could most quickly produce
a correct action from inconclusive perceptions. Encoded in the large,

highly evolved sensory and motor portions of the human brain is a

billion years of experience about the nature of the world and how to

survive in it. The deliberate process we call reasoning is, I believe,

the thinnest veneer of human thought, effective only because it is

supported by this much older and much more powerful, though

usually unconscious, sensorimotor knowledge. We are all prodigious
olympians in perceptual and motor areas, so good that we make the
16 Mind Children

difficult look easy. Abstract thought, though, is a new trick, perhaps

less than 100 thousand years old. We have not yet mastered it. It is
not all that intrinsically difficult; it just seems so when we do it.
Organisms that lack the ability to perceive and explore their environ-
ment do not seem to acquire anything that we would call intelligence.
We need make only the grossest comparison of the plant and animal
kingdoms to appreciate the fact that mobile organisms tend to evolve
the mental characteristics we associate with intelligence, while immo-
bile ones do not. Plants are awesomely effective in their own right, but
they have no apparent inclinations toward intelligence. Perhaps, given
much more time, an intelligent plant could evolve — some carniverous
and "sensitive" plants show that something akin to nervous action is

possible —but the life expectancy of the universe may be insufficient

time.
The cybernetics researchers, whose self-contained experiments were
often animal-like and mobile, began their investigation of nervous
systems by attempting to duplicate the sensorimotor abilities of ani-

mals. The artificial intelligence community ignored this approach in

their early work and instead set their sights directly on the intellectual
acme of human thought, in experiments running on large, stationary
mainframe computers dedicated to mechanizing pure reasoning. This
"top-down" route to machine intelligence made impressive strides
at first but has produced disappointingly few fundamental gains in
over a decade. While cybernetics scratched the underside of real
intelligence, artificial intelligence scratched the topside. The interior
bulk of the problem remains inviolate.
All attempts to achieve intelligence in machines have imitated
natural intelligence, but the different approaches have mimicked
different aspects of the original. Traditional artificial intelHgence
attempts to copy the conscious mental processes of human beings
doing particular tasks. Its limitation is that the most powerful aspects
of thought are unconscious, inaccessible to mental introspection, and
thus difficult to set down formally. Some of the cyberneticists,
taking a different tack, had focused on building models of animal
nervous systems at the neural level. This approach is by
limited
the astronomical number of cells in large nervous systems and by
the great difficulty of determining exactly what individual neurons
do, how they are interconnected, and what nerve networks do. Both
traditional AT and neural modeling have contributed insights to the
 Mind ill Motion 17

enterprise, and no doubt each could solve the whole problem, given
enough time. But with the present state of the art, 1 feel that the fastest
progress can be made by imitating the ezvlution of animal minds, by
striving to add capabilities to machines a few at a time, so that the

resulting sequence of machine behaviors resembles the capabilities of

animals with increasingly complex nervous systems. A key feature
of this approach is that the complexitv of these incremental advances
can be tailored to make best use of the problem-solving abilities of the
human researchers and the computers involved. Our intelligence, as

a tool, should allow us to follow the path to intelligence, as a goal, in

bigger strides than those originally taken bv the awesomely patient,
but blind, processes of Darwinian evolution.
The route is from the bottom up, and the first problems are those of

perception and mobility, because it is on this sensorimotor bedrock

that human intelligence developed. Programs which tackle incre-
mental problems similar to those that faced early animals — how to
deal with, and even to anticipate, the sudden surprises, dangers,
and opportunities encountered bv an exploring organism are being —
written and tested in robots that have to face the uncertainties of a
real world. Most approaches will fail, but a few will succeed, in
much the same wav that a tiny fraction of the spontaneous mutations
that appear in organisms survive into the next generation. Survival
depends on the advantages a new technique offers in coping with
the challenges of a complex and dynamic environment. By setting up
experimental conditions analogous to those encountered by animals in
the course of evolution, we hope to retrace the steps by which human
intelligence evolved. That animals started with small nervous systems
gives confidence that today's small computers can emulate the first
steps toward humanlike performance. Where possible, our efforts to
simulate intelligence from the bottom up will be helped by biological
—
peeks at the "back of the book" at the neuronal, morphological, and
behavioral features of animals and humans, as revealed bv the people
who study those aspects of life.

The modern robotics effort is just twenty years old, and for only the
last ten of those have computers been routinely available to control
robots. The recapitulation of the evolution of intelligent life is at a
very early stage — robotic equivalents of nervous systems exist, but
they are comparable in complexity to the nervous systems of worms.
Nevertheless, the evolutionary pressures that shaped life are already
U
c
o
 3

^rf
^ g X ~>^73E J-
~ ^
i/i
(C j:
— 3 Oi XI £-^ re
s:
m
01
a
01
-o
X tC -5 2
> - P > =
c
w _
oi

01 -
Ml ^ re 73
ere ^
'^
£x o •—
3
teg-
X — ^
re

^ >- ^ 5 '^ ,- ^^
7=
'-'

«:
c
"^
SP
^ £ rS S
~ 2i
-5 -c
^ ^
X 73 X ^ X
Ml ^ C 3
X
^^.S C X
£ ? E -" X
"X
T3
^ E >-^ -t:
T " Xc
'

= -o ^' S •
a.'^ t c ra •— - tC X 5 C X
o
6 <z
^X .3^ c;= S3 -^
S!
—
c 5 H -; £ re 6C '^
^ -0 .S -
XJ ^
'^
tc ^
u X 't; • - Xi ^ > ^ X a."
>.£ S E
.

01 i/i
C c Q X o re
5 a. c 73 £ 3^ re >
O) 5
£ o
?.
v£i 0/ IS — f3!:~"" — -px
o' .£ ^ ^
^; I-

X 5 X
>
ra u
3 ^ 01
oc
a. 73
0* C
X i ~
c
X re XX -7- £
e >^ X "^ = > D- n: oi re
0*
O ra 5i "^ c -, 3 01 re "^ C
£ jr O i rz u^ \

-
O 73 >
1/5
c g-S XI k- _ c
.2 3 d jf
111 o ^ i o,^ 01 ^
X
> c *J o.
x: _ re
£ x: ~ £ re
re
X u 01 -p

'x m X u c £ C 01 , 3 c re
"^
Ol

o ra <z
'^
cj
o re
c '^ •- &--^ c oi f3
^C ^X XOi tc—
^ X .;^
I^ re -r- O
>^
> 2 a. re
c Xi X ji; o
c ^ 01 t/i E e 73
c ^ ^o £ 5;
^£30
.

X X1-

t "r" S£ '5 'x o 3 re

o c
01 re
Ji X
-"
H o •-
:S ii
•-
5 ^ S.

C o oT o _o ^ i-H re
o C
ac
c 73 >
-^ o
c X o 3 77^ re X 01 —
a;
r- >-
> i^i
^ (Z C "
~3 ,- 3 ii x: T 0) 3O J= X £ X —
01aj i-
01

S S £ X X D.
Si
T3
X - S c E 5 c en
<= 3 c 1 X ^
Ki

ifi
T3
s:.
C X
ra
^ c >;E 3 re c C O- ii iC' re
—
J= p
£ >
j;
re
r- 01

S C -c ^
x:
'X ^ tc c 0* c u
>, XI
X
>. O.
-^ O _ P
'X,
>^ i^ i:
01 3 oj
X ^'1 "x: 73 £- ^ 5 -o >
U > ._ >
2 ^
o 73
'i>
O C £ "", '2 ^
5 -r X O ;ii
tC 0.3
jr "
^ 2 3 73 01 X 73 C
X ~S X b;o >re _re
^- re

^ J2 c £ 12 ><,
73 .£
C J= ^X ::f- 5 £ C^ re
f,
5 O)

>.-S £ ^- r ^ 5
^ .-^ re o .=
-p "x
C
>s'X!
tJ
^-
£ — •§, ^ ^'^ 5 re£
H ^X X o > Q. re
^ oi
>,
X re — o 3 O X C3 U C C X S£
C
<^. X :;: X £ d

;^ ;?; ..- ^- j: -rr _; _ 73 f^ - 5 cc T5 ^- '

b ^ 3 01
"P ii^
C o C
;L
— t?;
O -5 O 3 O 6C
X ^ ? C X re -;i -C 5"^ '^ ^^-
o £ - ^-
•£ -c
X —
3 -I > X X re re ^ -^ —X
f=
^^o re o X
Xc
t. X.
J-
o -^ X o X
.s: X S^ <-> >-
._
X
X 5 >.
tc-" 3 D.^
P ^ ^
73 a. — 'i —£
- o >
o •£
-
/' a.
-r
ii-"
3 --2^-0
ifi X •£ 5 £
" "
>>—
'd
re
-^ -2 ^ H f^x 3 §-x
•^ 3 S c 73
> > D o .£ - X'.

5 jc '-^

6C ^
re
^ ^ p £•2 ^ i= 3 §73 X aj

bCx
re Sc-^a;£-aga;c
re^o'S ^ — -T"
^ s £: = '' C .2pT3 :3 ^
3 2Px 3 > ,j^ * " X re r^
X re re X
£ X tc 0> ^ 01 re
Ef
^ o re a^ ^ 5 5PC
x: 2
X ^
£ ^ 3
3
=o re X ° " - 3 X ^ ^ X £ ^
.,
:3 X re ^ £ 3 ^
3 "3 a^ >
O) C ^ 2
re
5 -T '£ re -P 5 X X
% S E
- -3 ,- o re ^ § 3 X i
.'£• 01
c COO.' TO E ^ £ re
re 01 -1 3
^X""^?--3^re
.t;
o "^ S •£
6C c
X
c c O 3 ^
? P •- r ,73
.£ > X"5 a; >-M
&D
SO c ^,
C >,re73-^:t:- E
3 5 3 ^-
—^ X ^ X :i. ^
73 a; 3 ^ J2
~>, 5S o 3-.
-i.
i> 3:
01 r ? o-x TV P^^ X
si; re ^
'r^ ^•
tie
i: I

O in
iri
'*" o > •- x: -' c£ c r" £
—x "^ ^ 3 ; > -^ "5
I
re
73 -»2
O ^^
£ 5cii c s-e
re >< .
2 o
^ re a. X c .3^3 X I
re
01 •?-
-J

5 " ^ "3 X ^X £ re
^^,^33 •- C 3
ii = &.
•-§£- > _-, o -- 1 § ri ^
>^2 X i -3 X — re i^ X X c. — ^ ^
.

a £f .i:
•^£ 3 x: ?>; ,x ;^ - !^ u = _ 3
re — Oi '^, £r
OC 3 T C £ «
5 o ^ ^ ;3
- ^ o 1 .= X = 3 ^^ 3:£ ^" 35
>. £
i 3 h =5 -£ ;=.£ X -3
^ o Oi -J ?2 ;i
o :^ £ ^ 73 ^ X
X5C 733 XI>^ 3a. 13
5 1 5 = - g DCX-^ g S
re
X 3 C - C
£
re

—X
>
X
o
^- - £;5a.6C-rQ.>-o X
X
— 73
3
X 3^
'^ re

jf- a; re
^ E
Qj X Cc
Xre re ^
P
''^

2
—
.£ ^ X ^ £

X 5
c
^- X".
^
o

re
-= p 'p ^ 3 i
£
„
^.:: '>, 0773^?
,, ,^ b; iJ iri
^J. rex ir>x I,
£ I 3
re
0-3i£'3D. 3c 3
.3 o •- ^
D. 3
X a; ^ ^ X
3 X '^ _3 X
S re 2 5
O ^ ^ re
X > 7c
'-'

£ 75 3 X Scl £^ £ X "2 != £ re -O 1
'•
.£
_3
c X a^ X <^ £
X -^ x c re
3 X
.

^€X i2
< X
r=
^ :^ E reac5<reM)£a^
-^ u^
X t o ^ .3 ^ ti 01
X re <*-
£ 01
II >,^*; r3 60 3 0, ^
CoiOHrexure
Oi
o re 7^
a..£ o X E 0,0 0- 3
20 Mind Children

palpable in the robotics lab, and I am confident that this bottom-

up route to one day meet the traditional
artificial intelligence will

top-down route more than half way, ready to provide the real-
world competence and the commonsense knowledge that has been so
frustratingly elusive in reasoning programs. Fully intelligent machines
will result when the metaphorical golden spike is driven uniting the
two efforts. A reasoning program backed by a robotics world model
will be able to visualize the steps in its plan, to distinguish reasonable
situations from absurd ones, and to intuit some solutions by observing
them happen in its model, just as humans do. Later I will explain why
1 expect to see this union in about forty years. For the moment, let us

explore some of the artificial fauna at the bottom.

As we have seen, Shakey was not one of them. This robot was an
expression of the top-down effort: its specialty was reasoning, while
its rudimentary vision and motion software worked only in starkly

simple surroundings. At about the same time, though, on a much

lower budget, a mobile robot that was to specialize in seeing and
moving in natural settings was born at Stanford University's Artificial
Intelligence Project, located about eight miles away from Shakey's
residence at SRI. John McCarthy founded the Project in 1963 with
the then-plausible goal of building a fully intelligent machine in a
decade. (The Project was renamed the Stanford Artificial Intelligence
Laboratory, or SAIL, as the decade drew nigh and plausibihty of
the Project drifted away.) Reflecting the priorities of early artificial
intelligence research, McCarthy worked on reasoning and delegated
to others the design of ears, eyes, and hands for the anticipated
mind. SAIL's hand-eye group soon overtook the MIT robotics
artificial

group and was seminal in the later boom in smart robot arms for
industrial uses. A modest investment in mobility was added when
Les Earnest, SAIL's technically astute chief administrator, learned of a
vehicle abandoned by Stanford's mechanical engineering department
after a short stint as a simulated remote-controlled lunar rover. At
SAIL became the Stanford Cart, the first mobile robot controlled by
it

a large computer that did not reason, and the first testbed for computer
vision in the cluttered, haphazardly illuminated world most animals
inhabit. The progeny of two PhD theses (one of them my own), the
Stanford Cart slowly navigated raw indoor and outdoor spaces guided
by TV images processed by programs quite different from those in
Shakey's world.
 Mind in Motion 21

In the mid-1970s NASA began planning for a robot mission to Mars,

to follow the successful Viking landings. Scheduled for launch in
1984, was to include two vehicles that would rove the Martian
it

surface. Mars is so far away, even by radio, that simple remote

control would be either very slow or very risky; the delay between
sending a command and seeing its consequence can be as long as
forty minutes. If the robot could travel safely on its own much of
the time, it would be able to cover much more
Toward this
terrain.

end, Caltech's Jet Propulsion Laboratory (JPL), designer of most of

NASA's robot spacecraft, which until then used quite safe and simple
automation, initiated an intelligent robotics project. Pulling together
methods, hardware, and people from university robotics programs,
it built a large, wheeled test platform called the Robotics Research
Vehicle, orRRV, a contraption that carried cameras, a laser rangefinder,
a robot arm, and a full electronics rack, all connected by a long cable
to a big computer. By 1977 it could struggle through short stretches
up a certain rock and rotate it for
of a rock-littered parking lot to pick
was halted when the Mars 1984
the cameras. But in 1978 the project
mission was canceled and removed from NASA's budget. (Of course.
Mars has not gone away, and JPL is considering a visit there at the
end of the millennium.)
Along with the Office of Naval Research, the first and steadiest
supporter of artificial intelligence research (and a major reason all
the early advances in the field happened in the United States) is the
Department of Defense's Advanced Research Project Agency (DARPA).
Founded in 1958 after the national humiliation caused by Sputnik, its
purpose was to fund far-out projects as insurance against unwelcome
technological surprises. In 1981 managers in DARPA decided that
robot navigation was sufficiently advanced to warrant a major effort to
develop autonomous vehicles able to travel large distances overland
without a human operator, perhaps into war zones or other hazardous
areas. The number of mobile robot projects jumped dizzyingly, in
universities and at defense contractors, as funding for this project
materialized. Even now, several new truck-size robots are negotiating
test roads around the country —
and the dust is still settling.
On a more workaday level, it is not a trivial matter that fixed
robot arms in factories must have their tasks delivered to them.
An assembly-line conveyor belt is one solution, but managers of
increasingly automated factories in the late 1970s and early 1980s
22 Mind Children

found belts, whose routes are difficult to change, too restrictive. Their
robots could be rapidly reprogrammed for different jobs, but the
material flow routes could not. Several large companies worldwide
dealt with theproblem by building what they called Automatically
Guided Vehicles (AGVs) that navigated by sensing signals transmitted
by wires buried along their route. Looking like forklifts or large
bumper cars, they can be programmed to travel from place to place
and be loaded and unloaded by robot arms. Some recent variants
carry their own robotic arms. Burying the route wires in concrete
factory floors is expensive, and alternative methods of navigation are
being sought. As with robot arms, the academic and industrial efforts
to develop mobile robots have merged, and a mind-boggling number
of directions and ideas are being energetically pursued.

A Robot for the Masses

The history presented so far is highly sanitized and describes only

a few major actors in the new
field of robotics. The reality is a
witch's brew of approaches, motivations, and, as yet, unconnected
problems. The practitioners are large and small groups of electrical,
mechanical, optical, and all other kinds of engineers, physicists,
mathematicians, biologists, chemists, medical technologists, computer
scientists, artists, and inventors, all around the world. Computer
scientists and biologists are collaborating on the development of
machines that see. Physicists and mathematicians are working to
improve sonar and other senses. Mechanical engineers have built
machines that walk on legs, and others that grasp with robot hands of
nearly human have suffered
dexterity. Yet all of these fledgling efforts
from poor communication among the various groups, which have not
been able to agree upon even a general outline for the field of robotics.
Despite the chaos, I expect to see the first mass offering from the
cauldron served in time for the new millennium, in the form of a
—
general-purpose robot for the factory and the home.
In industrialized nations, agriculture and manufacturing is increas-
ingly the province of machines, leaving people free to provide human
services for one another. Food and goods have become plentiful

and cheap under this many services have increased

arrangement, but
in cost. Domestic service, once common, is scarce and expensive.
Domestic machines such as food processors, vacuum cleaners, and
 Mind in Motion 23

microwave ovens do not fill the void in families where all the adults
work outsidethe home. The need has existed for many decades: When
will there be a robot to help around the house?
For many years I believed that robot servants, ubiquitous in science
were unlikely in the near future. Households are complex
fiction,

environments with limited resources. The economic return from a

mechanical domestic helper would only be a fraction of the value
of a robot in a typical industrial role, so the home robot must sell

(or rent) for much less. Worse yet, safe and effective operation
in the often chaotic home environment is a lot more difficult than
in controllable factory settings. Existing robots offer mostly blind,
repetitive, potentially lethal motions at a price comparable to that of

an entire residence. This enormous gap in price and performance is

real, yet I now expect to see a general-purpose robot usable in the

home within ten years. The change in my attitude comes partly from
research developments of the last few years and partly from a new
appreciation of the implications of the concept "general purpose."
Today's industrial robots are more flexible than the fixed automation
they sometimes displace, but they do so very few things well that the
term "general purpose" hardly applies. Indeed, individual robots are
usually bolted to a fixed station, equipped with grippers and some-
times sensors specialized for a certain task, which they henceforth
execute, again and again, perhaps for the rest of their existence. The
narrowness of their repertoire, besides being boring, greatly limits the
number that can be sold. There are less than 100 thousand robots
(other than toys) of all makes world today. Compare this with
in the
100 million cars, 500 million television sets, or 20 milHon computers.
So few units sold can support only a limited amount of engineering
thought and development. The result is a less-than-optimal design
at a high price. But not forever. As the number of units produced
grows, so does the opportunity and incentive to improve the design
of the robots and the details of their production. The costs drop,
and the robots become better, incidentally expanding the market and
increasing the number of units sold, leading to further improvements.
The graph of declining unit cost versus number of units produced is

called the manufacturer's learning curve.

The potential market for robots will expand enormously when a
certain level of general usefulness is achieved. Up to this breakeven
point, specialization — the exploitation of the unique circumstances of
 —

24 Mind Children

a job to achieve acceptable performance with minimum complexity

willbe the robotic norm. Beyond the breakeven point, the potential
market will be large enough that higher profits will go toward more
standard designs sold in ever larger numbers. The cheap, mass-
produced, high-utility robot will have arrived. We have accumu-
lated enough experience to specify some of the characteristics of this
Model T of robots. It will not be intelligent, and it will not come pre-
programmed to do many useful tasks. It will come from the factory
with a sufficient set of mechanical, sensory, and control capabilities
that can be conveniently invoked by software specially written for
particular applications.
The first major market for such a machine will be in factories,
where it will be somewhat cheaper and considerably more versatile
than the older generation of robots it replaces. Its improved cost-
benefit ratio will allow it to be used in a much wider arrav of jobs
and thus in greater quantities, further lowering its cost. In time it
will become cheaper than a small car, putting it within the reach
of some households and creating a demand for a huge variety of
new software. The robot control programs that actuallv get various
jobs done will come from many different sources, as do programs for
today's personal and business computers.
As with personal computers, many successful appHcations of the
general-purpose robot will come as surprises to its makers. We
can speculate about the videogame, word -processor, and spreadsheet
equivalents of the mass robot era, but the reality will be stranger To
get the guessing game going, consider programs that do light me-
chanical assembly (from a factory automation company), clean bath-
rooms (from a small firm founded by former cleaning stafO, assemble
and cook gourmet meals from fresh ingredients (a collaboration of
a computer type and a Paris chef), do tuneups on a certain year
of Saturn cars (from the General Motors Saturn service department),
hook patterned rugs (by a Massachusetts high school student), weed a
lawn one weed at a time, participate in robot races (against other soft-
—
ware programs are assigned a certain physical robot chassis by lot
just before the race begins), do detailed earthmoving and stonework

(by an upstart construction company), investigate bomb threats (sold

departments worldwide), deliver to and fetch from a ware-
to police
housed inventory, help to assemble and test other robots (in several
independent stages), and much more. Some of the applications will
 Mind in Motion 25

require optional hardware attachments for the robot, special tools and
sensors (such as chemical sniffers), protective coverings, and so on.
It may be that writing applications programs for successive genera-

tions of general-purpose robots will become the major human occupa-

tion in the early decades of the next century. The skilled plumber, for
instance, will be faced with the choice of applying his or herplumbing
skills to meet the needs of a few hundred clients or encoding those
skills into robot programs that might be sold successfully to thousands

or even millions of customers. The first alternative will become

increasingly less attractive as a source of income as manual work
competes with an ever-increasing number of ever-more-sophisticated
robots controlled by ever-better software. The latter course has its

risks also — a program may flop in the marketplace, just as inventions,

books, music, art, and computer software flop today. On the other
hand, a successful program might generate years of income for its

author.
Almost everyone many skills, and each skill can
has, or can develop,
be a potential source of royalties when encoded as a program. Many
competing versions of each skill will be marketed and purchased
on the basis of utility, cost, personal taste, fashion, and advertising.
Each program will have a limited lifetime, destined to be eclipsed by
replacements that are either simply better or are designed to operate
a more sophisticatednew generation of robot. A large secondary
industry will spring up to help in the programming process.
Before long programs will be created that make general-purpose
robots good by leading them through
learners, teachable, for instance,
the motions of the task, or by example. The accumulating library
of such programs will eventually be a motherlode of encoded hu-
man, nonverbal knowledge which can be tapped by the waves of
autonomous robots that will follow the breakeven gener-
increasingly
The expert-systems industry has already begun encoding verbal
ation.
knowledge in this fashion.

Breakeven Locomotion

To be successful, a mass-produced general-purpose robot will require

a minimum The first robots of this class need
level of functionality.
not be able todo everything, or even do most things, very well.
They must do enough things well enough to create an open-ended
26 Mind Children

market for themselves, with each drop in price bringing a more than
proportional increase in the number of economical applications and
units in demand.
Even in highly urban settings, regions of flat, hard ground form an
archipelago in a sea of terrain that is variously rough, soft, stepped,
or totally impassable. A machine unable to navigate this sea will be
trapped on a single island, its potential uses enormouslv restricted.

Our breakeven criterion thus calls for a drive system more capable
than standard wheels. Robots with legs are just now showing signs
of practicality. The most convincing demonstration to date is by
a California company called Odetics, whose six-legged, spiderlike,
electrically driven robot can climb out of its truck, onto a hatbox,
squeeze through a door, then show off by one end of the truck
lifting

and dragging it around. This and other promising demonstrations

make it likely that practical legged locomotion will be available within
a decade.
Legs are a powerful mechanism movement, but their start-stop
for
nature limits speed and energy efficiency. The Odetics machine, for
instance, drains its batteries in under an hour of slow walking. The
serious power constraints in a self-contained robot may demand a
more frugal drive system. On flat ground wheels are best, offering
close to 100% efficiency over a wide range of speed. A compromise
solution may be slow legs terminating in wheeled feet like powered —
roller skates. The robot would roll on these wheels most of the time

but lift its feet over obstacles and up stairs. On rough ground it might
plod along slowly in a full walk.
Hitachi, the Japanese electronics giant, experimented in the early
1980s with a particular! v simple version of the wheel-foot idea. For
use in nuclear reactors, the Hitachi system has five simple "legs," each
a vertical, motorized post that telescopes up and down out of the
body. The legs are arranged uniformly around the robot, in a regular
pentagon. Each ends in a wheel able to steer and drive. Five legs
are the minimum that allow a robot to stand stably with any one leg
raised, without shifting its weight. The Hitachi machines climb stairs
by rolling up to them on five wheels, raising the leading one to the

height of the first stair, driving forward until the raised leg is securely
over the step, lowering it slightly until the contact is firm, and then
continuing with the next nearest leg. On narrow stairs the robot may
have its wheels resting on up to three successive steps at the same
Walking Machine
The Odetics "Odex" can walk, climb, squeeze through doorways, or
spread for stability. But its power consumption limits it to an hour
of mobility per battery charge.
28 Mind Children

time. A similar procedure carries it over obstacles. The robot can

traverse rough ground slow^ly with the individual wheels riding up
—
and down over the surface irregularities an active suspension. The
upper body of the robot remains perfectly horizontal under normal
operation. Time and further research will tell which configuration
proves best for the first all-doing robot.

Five Legs
This design for mobility from Hitachi, five steerable wheels on tele-
scoping legs, has more limitations than one with fully articulated
legs, but it gets much better mileage on flat surfaces.
 Mind in Motion 29

Breakeven Manipulation

Few useful jobs can be accomplished if the robot simply moves about.
Productive work calls for the holding and transporting of ingredients,
parts, and other things. Industrial manipulators, the most
tools,

numerous and successful robots to date, have arms that can reach
where needed by using about six rotary or sliding joints. If we neglect
fine points of weight, power, and control, some of the smaller designs
are nearly adequate for the reach needed in the breakeven robot. Since
many jobs require bringing pairs of objects into contact, our robot
will probably come with at least two arms. A third arm would be

advantageous where the contacting objects must be operated

for jobs
on in some way (electronic hobbyists will recognize soldering as one
such job).

Robot hands are not as well developed as robot arms. The industrial
manipulators manage to grasp and hold with special fixtures for
particular objects or with a simple two-fingered kind ofhand called
a parallel-jaw gripper. Such grippers are easy to operate but can
safely grasp only some kinds of rugged object. They are incapable
of controlling or changing the orientation of something being held.
Our universal robot needs more flexibility.

A few research projects have investigated multifingered grippers

that exhibit much greater dexterity. One of the best comes from a ten-
year effort by Ken Salisbury, now at MIT. Salisbury's three-fingered
robot hand can hold and orient bolts and eggs and manipulate string
in a humanlike fashion. He determined the basic configuration and
dimensions of the hand with a computer search over different linkages,
looking for the minimal set that allowed fingertips to converge on and
securely hold arbitrarily shaped small objects. The result has three
symmetrically placed fingers that bend much
humans. like those of
However, because the fingers can bend outward as well as inward,
the hand can grip hollow objects from the inside as well as from the
outside. The driving forces come through thin steel cables pulled on
by a bank of motors some distance down the robot's wrist.
To accomplish feats of even moderate dexterity the hands must
"feel"grasped objects. Salisbury is developing hemispherical fin-
gertips for the hand that, through carefully placed internal strain
gauges, can sense the magnitude and direction of external forces. The
computer programs to plan and to carry out arm and hand motions
Three Fingers
The "Salisbury Hand," a minimal solution for general robot dexterity.
Each finger is controlled by three motors. The hand can grip from the
outside or, with equal facility, bend its fingers outward to grip a
hollow object by its interior.
 Mind in Motion 31

for complex manipulators are still in a tender state. Several programs

exist that can plan collision-tree arm movements between two points
in known clutter. These programs consider a space (the so-called
configuration spjace) that describes all possible postures of the manipu-
lator. Each joint adds one dimension to this space, so a manipulator
as complicated as Salisbury's has a complex configuration space. A
complex space translates into an expensive and time-consuming search
for a good path. Running times of minutes to hours are typical, but
better algorithms continue to be found, and computers continue to
become faster.

Breakeven Navigation

The mechanical ability to move is onlv part of the problem of mobility.

One must also be able to find and return to specific locations and
avoid dangers in transit. I have been working on this issue for most
of mv career and am happv to report that some good solutions are
developing. My thesis work at Stanford during the 1970s was on
programs that were intended to let the Stanford Cart find its way
through cluttered rooms and outdoor spaces. The first version of such
a program, in 1976, obtained its \'iew of the world through one TV
camera on the robot. Bv locating distinctive areas in the T\' image and
tracking them as the robot moved, the program was able to estimate
their distance and the extent of its own motion. It constructed a sparse
three-dimensional map of its surroundings, identified obstacles in it,

and planned a path to its destination that staved clear of them. The
program then moved the robot about a meter along that path, looked,
mapped, planned, and moved again. In repeated cautious lurches, the
Cart was to creep safelv to its destination.
Unfortunately, the program didn't work. About one lurch in four,
the part of the program that attempted to estimate the robot motion
from the changing image made a mistake, misidentified areas being
tracked, reported an incorrect robot motion, and messed up the slowly
building map. The chance of successfully crossing a large room, a
journey of perhaps thirty lurches, was so small as to be negligible. In
1979 I tried again, with a new program aided by a small amount of
new hardware, a mechanism that precisely moved
the camera from
side to side along a track. With it, program was able to
the driving
obtain several images of the scene without moving the whole robot.
32 Mind Children

much as a human obtains two images, one from each eye. By carefully
exploiting the extra information to prune away errors, the program
improved the success rate for a single lurch to nearly 100%. The robot
was now often able to successfully complete the thirty lurches to cross
a room to the desired destination and show a correct map on a display
screen. About one time in four, however, it failed, either because the
aggressive error-pruning had removed a real obstacle from the map
and the robot had collided with it, or because, in spite of the pruning,
errors had crept in and confused the robot's idea of its position. Good
enough for my thesis, perhaps, but not good enough for a robot to do
complex tasks that would, at the minimum, require it to cross rooms
many times.
In 1980 I moved to Carnegie Mellon University, to continue the re-
search under the auspices of its new Robotics Institute. Two graduate
students. Chuck Thorpe and Larry Matthies, examined and greatly
improved the old program, increasing both its speed and its accuracy
tenfold. When everything went well, it was now able to report the
position of the robot (a new one we call Neptune) to a few centimeters
accuracy. Unfortunately, things did not always go well, and the failure
rate remained stubbornly unchanged. The robot still crossed the room
correctly only about three times out of four.
In 1984 our group agreed to do some research for a new company.
Denning Mobile Robotics, Inc., in Massachusetts, that was developing
a robot security guard (more accurately, a roving burglar alarm).
Instead of a camera, the robot was equipped with a beltlike ring
of sonar range sensors like those found in Polaroid cameras. These
had already been found to be very useful for detecting the presence
and general direction of nearby obstacles, thus allowing the robot
to avoid them. Our aim was more ambitious, however. Instead
of merely sensing imminent collisions, could the continuously active
sonar system be used to build a map of the surroundings that could
direct accurate point-to-point navigation, as (three times out of four)
our vision-guided programs could do? The sonar units each emit an
ultrasonic chirp of sound over a wide cone and report the time to
the first echo they hear. This time is proportional to the distance of
the nearest object within the cone. The distance to the object may be
accurate to better than a centimeter, but since the cone subtends an
angle of about 30°, the side-to-side position is still highly uncertain.
This is very different from the almost pinpoint measurements possible
 Mi)id in Motion 33

from TV cameras, and so the program methods developed for the Cart
could not be used.
Although a single sonar reading can tell a program only a little

about the position of the thing that caused the echo, it maps out a
large volume of empty space in front of that thing. When readings
from different sensors overlap, the emptv region indicated by one

Autonomous Navigation
Tlie Denning Sentnj is a commercial product that can patrol a large
warehouse or office complex evenj night for months without human
intervention, guided by light-emitting beacons and a sonar image of
its surroundings. By day, it recharges itself in a special booth.
34 Mind Children

reading may restrict the possible location of the echo-causing object

indicated by the other. Hundreds or thousands of readings from
different positions, taken together, might be able to build detailed

maps in spite of the fuzziness of individual sensors. Because the

sensitivity of a sonar sensor falls off smoothly from the middle to the
edges of its cone, it seemed best to do the mixing with probabilities.
Alberto Elfes, another graduate student, andwrote a program based
1

on these ideas and were astonished when it droxe the robot much
more reliably than the old TV-guided program. Yet another student,
Bruno Serrey, along with Larry Matthies, then found a way to use
the probabilistic approach for TV data and again discovered that it

worked remarkably better than the old approach.

Our new method represents the space around the robot as a grid
of cells, each containing the probability, based on all available sensor
readings, that a corresponding cell in space is occupied by matter. A
reading may lower the probability of some cells (for instance, those
belonging to the interior of a sonar cone) and raise others (such
as those on the range surface of the sonar reading). It provides a
convenient way to combine the results of different kinds of readings,
and indeed Elfes and Matthies recently demonstrated a program that
builds maps from both sonar and TV data. Using these new methods,
our robots can now travel long distances almost flawlessly. With a new
mathematical foundation for this approach, and a new, very fancy,
robot, Uranus, to continue the work, I feel extremely confident that
navigation will be more than adequately in hand within the ten-vear
timeframe of the universal robot.

Breakeven Recognition

The sensory system has another vital function: the recognition and
localization of specific objects in the robot's surroundings. Recognized
objects may be small things destined later to be picked up by one of
the hands or large objects that serve as landmarks or work locations.
Imagine a process whereby objects are described by shape and surface
characteristics and the robot's recognition system looks for one object
at a time. A tentative identification can be confirmed by viewing
the scene from a different point. The result is a description of the

position and orientation of the object suitable for use by the program
that controls grasping by the hands.
Object Finding
3DPO (for Three-Dimensional Parts Orientation) is a program that
finds particular parts in a clutter of other parts. This sequence of
images is: (1)a three-dimensional computer model of the part to be
found; (2) a TV picture of a jumble of actual parts; (3) a computer
image of the same parts where brightness now indicates how close to
the camera is each bit of visible surface; (4) the computer's deduction

of the major surface boundaries of the jumble; and (5) the computer's
fit of the part model to actual occurrences of the part in the jumble.
36 Mind Children

Computer vision is by far the most promising vehicle for this

identification ability. The key operation is identifying a specific object
in a mass of clutter. Vision research for industrial robots has produced
partial solutions to the so-called "bin-picking" problem. Bin-picking
programs let a computer identify predefined objects in a jumble in a
TV image, even if the objects partially occlude each other, so that they
can be removed one at a time by a manipulator. A General Motors
research group in the 1970s demonstrated a system that worked if

the overlapping parts mostly lay flat. It was too slow and unreliable
to be practical in production, but it did demonstrate feasibility. In
the last several years many groups in the United States and Japan
have unveiled programs that can identify simple objects on the basis
of three-dimensional data obtained from a camera looking at a scene
illuminated by special devices that generate stripes or grids of light.

On contemporary computers, these systems take many minutes to

make their less-than-satisfactory identifications. Yet it is likely that
the minimal requirements for our robot will be met within our ten-
year timeframe.

Processing and Coordination

The and movement-planning

best prototypes for the low-level sensory
parts of our future robot all consume many minutes of computer time
on a good microcomputer. This is partly a measure of the patience
of the researchers; processes that run for much longer than an hour
are simply too difficult to investigate effectively, while simpler, faster
programs are not very interesting because they work less well. On
the other hand, the running times do tell us something about the
difficulty of the breakeven criteria. A robot that spends up to an
hour considering every simple move is clearly unacceptable, but a
few seconds would be tolerable. A computer able to do a billion
operations per second, with a billion bytes of main memory, would
be enough. This is about the power of the largest supercomputers that
have been built to date, and a few hundred times faster than the best

microcomputers. The continued computer evolution should deliver it

in a microcomputer within Depending on the progress in

a decade.
various lines of development, the power may be spread among few or
many individual processing units, and may depend on a significant
fraction of specialized hardware, for instance, circuits to do low-level
 Mind in Motion 37

vision processing. The exact hardware configuration is unimportant

for our purposes here.
Our work at Carnegie Mellon with integrated tasks for a mobile
robot suggests that basic processes should be organized into modules
that run concurrently. A navigation program driving the robot to a
desired location might, for example, coexist with ones that watch out
for surprises and dangers. If a stairwell-detecting module concludes
that hazard isnear, it would take over control of the robot until the
danger was past.

A Sensible Robot

Here is a possible configuration for our mass robot. It moves on five

leg-wheels of the Hitachi design and has two arms with Salisbury
hands. Topped by a pair of color TV cameras, it has an unobtrusive
array of sonar sensors to sense the world in directions not covered
by the cameras. It carries an inexpensive laser gyroscope to help with
navigation, and it is controlled by a computer system able to do at least
a billion operations per second. Integral with the computer hardware
is a software operatingsystem that allows multiple simultaneous pro-
cesses. Built-in programs permit objects in the world to be described,
visually identified in or out of clutter, and picked up. A navigational
system can be asked to build, store, retrieve, and compare maps of
the surroundings and to bring the robot to specific locations.
Readers familiar with personal computers may recognize the sim-
ilarity to operating-system utility functions, especially the graphic
toolbox in the Apple Macintosh. These capabilities in the robot are

orchestrated by application software for (one hopes) an astonishing

variety of specific jobs; software is supplied by many independent
vendors. Again the similarity to personal computers is clear. One
might eventually hope for integrated software packages that allow the
robot to switch quickly and automatically from one task to another,
making it a more autonomous mechanical servant.

The Convergent Evolution

of Emotions and Consciousness

The machines we have been considering behave in a predictable

way that we might describe as mechanical or insectlike. Will robots
A General-Purpose Robot
This caricature of a first-generation general-purpose robot shows the
major systems: Locomotion with limited stair and rough-ground
capability, general manipulation, stereoscopic vision, coarse 360°
sensing for obstacle avoidance and navigation. Not shown is the
computer hardware and software that will be required to animate
this assembly.
 Mind in Motion 39 ,

continue to display this predictability as they become more complex,

or wiU they develop something akin to the richer character of higher
animals and humans?
As we have seen, the more advanced control programs in today's

roving robots use data from sensors to maintain representations, at

varying levels of abstraction and precision, of the world around the
robot, of the robot's position within that world,and of the robot's
internal condition.The programs that plan actions for the robot
manipulate these world models to weigh alternative future moves.
The world models can also be stored from time to time and examined
later, as a basis for learning.
A would meaningfully
verbal interface keyed to these programs
answer questions "Where are you?" ("I'm in an area of about
like

twentv square meters, bounded on three sides, and there are three
small objects in front of me") and "Why did you do that?" ("I turned
right because I didn't think 1 could fit through the opening on the
left.") In our lab, the programs we have developed usually present

such information from the robot's world model in the form of pictures
—
on a computer screen a direct window into the robot's mind. In
these internal models of the world I see the beginnings of awareness
in the minds of our machines —
an awareness I believe will evolve into
consciousness comparable with that of humans.
The term convergent evolution is used by evolutionary biologists
whenever species that are only very distantly related independently
develop similar characteristics, presumably in response to similar en-
vironmental pressures. Eyes are an example of convergent evolution;
they have evolved over 40 different times in the animal kingdom.
What was necessary was the presence of light-sensitive cells and
selection pressures favoring the survival of animals that could see,
however dimly at first. If a function of the nervous svstem as complex
as vision can evolve so many different times when environmental
pressures are right, what about emotions and consciousness? Unlike
vision, these features of the human mind have no incontrovertible
external manifestation and indeed lack a precise definition. Their
existence in animals, and even in humans, has been questioned bv a
generation of behavioral psychologists. Yet animal ethologists such as
Donald Griffin find the concepts useful in explaining animal behavior.
If an animal acts as I do when I am afraid, is it not reasonable to call

its mental state "fear"? If it chooses from among several complex

40 Mind Children

alternatives in dealing with a novel situation, as I would consciously

weigh my options in the same circumstance, why not ascribe "con-
sciousness" instead of some other mechanism with a different name
but the same effect?
Consider the following thought experiment. Suppose we wish to
make a robot that can execute some general task such as "Go down
the hall to the third door, go in, look for a cup, and bring it back."
Our most pressing need would be computer language in which
a
to specify complex tasks for the rover and a hardware and software
system to embody it. Sequential control languages successfully used
with industrial manipulators might seem to be a good starting point.
But paper attempts to define the structures and primitive actions
required for mobility would reveal that the linear control structure
of these state-of-the-art languages, though adequate for a robot arm,
would prove to be inadequate for a rover. The essential difference
is that a rover, in its wanderings, is regularly "surprised" by events
which it cannot anticipate but with which it must deal. This requires
that contingency routines be activated in arbitrary order and run
concurrently, each with its own access to the needed sensors, effectors,
and the internal state of the machine, and a way of arbitrating
their differences. As conditions change, the priority of the modules
changes, and control may be passed from one to another.
A request to our future robot to go down the hall to the third
door, go in, look for a cup, and bring it back might be implemented
as a module, FETCH-CUP, that looks very much like a program
written for the arm-control languages (which in turn look very much
like programming languages such as Algol or Basic), except that
another module, COUNT-DOORS, would run concurrently with the
main routine. Consider the following outline for such a program.

Module COUNT-DOORS:
Check the robot's surroundings for doors

Add one to the variable DOOR-NUMBER each time a new door is located

Record the location of the new door in the variable DOOR-LOCATION

Module GO-FETCH-CUP:
Step 1: Record the current location of the robot in the variable START-
LOCATION
Step 2: Set the variable DOOR-NUMBER to zero
 Mind in Motion 41

Step 3: Wake up the COUNT-DOORS module

Step 4: Drive the robot parallel to the right-hand wall until DOOR-
NUMBER is three or greater

Step 5: Cause the robot to face the location in the variable DOOR-
LOCATION
Step 6: If the robot is facing an open door, go to Step 10

Step 7: If the robot is not facing a door, subtract one from DOOR-NUMBER
and go to Step 4

Step 8: If the robot is facing a closed door, try to open it

Step 9: If the door fails to open, say "knock knock" and go to Step 6

Step 10: Drive the robot through the open door

Step 11: Check the robot's surroundings for cups; if there are none, go

to Step 15
Step 12: Record the location of the nearest cup in CUP-LOCATION
Step 13: Drive the robot to within reach of the CUP-LOCATION
Step 14: Pick up the cup at CUP-LOCATION; if this fails go to Step 15

Step 15: Go back and face the door at DOOR-LOCATION

Step 16: If the robot is facing a closed door, try to open it

Step 17: If the door fails to open, say "knock knock" and go to Step 16
Step 18: Drive the robot through the open door
Step 19: Return to START-LOCATION
Step 20: Put the robot to sleep

So far so good. We activate our program, and the robot obediently

begins to trundle down the hall, counting doors. It correctly recog-
nizes the first one. The second door, unfortunately, is decorated with
garish posters, and the lighting in that part of the corridor is poor, so
our experimental door-recognizer fails to detect it. The wall-follower,
however, continues to operate properly and the robot continues on
down the hall, its door count short by one. It recognizes door 3, the
one we had asked it to go through, but thinks it is only the second,
so continues. The next door is recognized correctly and is open. The

program, thinking it is the third one, faces and proceeds to go

it,

through. This fourth door, sadly, leads to the stairwell, and the poor
robot, unequipped to travel on stairs, is in mortal danger.
Fortunately, there is another module in our concurrent program-
ming system called DETECT-CLIFF. This program is always running
and checks ground position data incidentally produced by the vision
processes and also requests sonar and infrared proximity checks on
42 Mind Children

the ground. It combines these, perhaps with an a priori expectation

of finding a chff set high when operating in dangerous areas, to pro-

duce a number Ukehhood there is a drop-off in the
that indicates the
neighborhood. A
companion process DEAL-WITH-CLIFF, also running
continuously but with low priority, regularly checks this number and
adjusts its own priority on the basis of it. When the cliff probability
variable (perhaps we'll call it VERTIGO) becomes high enough, the
priority of DEAL-WITH-CLIFF will exceed the priority of the current
process in control, GO-FETCH-CUP in our example, and DEAL-WITH-
CLIFF takes over control of the robot. A properly written DEAL-WITH-
CLIFF will then proceed to stop or greatly slow down the movement of
the robot, to increase the frequency of sensor measurements of the cliff,
and to back away slowly from it when it has been reliably identified
and located.
Now there is a curious thing about this sequence of actions. A
person seeing them, not knowing about the internal mechanisms of the
robot, might offer the interpretation, "First the robot was determined
to go through the door, but then it noticed the stairs and became so
frightened and preoccupied it forgot all about what it had been doing."
Knowing what we do about what really happened in the robot, we
might be tempted to chastise this poor person for using such sloppy
anthropomorphic concepts as determination, fear, preoccupation, and
forgetfulness in describing the actions of a machine. We could chastise
the person, but in my opinion that would be wrong. The robot came
by its foibles and reactions as honestly as any living animal; the
observed behavior is the correct course of action for a being operating
with uncertain data dangerous world. An octopus in pursuit of
in a

a meal can be diverted by subtle threats of danger in just the way the
robot was. The invertebrate octopus also happens to have a nervous
system that evolved entirely independently of our own vertebrate
version. Yet most of us feel no qualms about ascribing qualities like
passion, pleasure, fear, and pain to the actions of the animal. I believe
that we have in the behavior of a person, an octopus, and a robot a
case of convergent evolution. The needs of the mobile way of life have
conspired in all three instances to create an entity that has modes of
operation for different circumstances and that changes quickly from
mode to mode on the basis of uncertain and noisy data prone to

misinterpretation. As the complexity of mobile robots increases, their

similarity to animals and humans will become even greater.
 —

Mind in Motion 43

Hold on a minute, you say. There may be some resemblance

between the robot's reaction to a dangerous situation and an animal's,
but surely there are differences. Isn't the robot more like a startled

spider, or even a bacterium, than like a frightened human being?

Wouldn't it react over and over again in exactly the same way, even
if the situation turned out not to be dangerous? You've caught me.

I think the spider's nervous system is an excellent match for robot

programs possible today. (We passed the bacterial stage in the 1950s
with light-seeking electronic turtles.) This does not mean that concepts
like thinking and consciousness must be ruled out, however.
In his book Animal Thinking, Griffin reviews evidence that much
animal behavior, including the behavior of insects, can be explained
economically in terms of consciousness: an internal model of the self,

surroundings, and other individuals that, however crudely, allows

consideration of alternative actions. For instance, bees, as Otto von
Frisch discovered, communicate direction, distance, and desirability of
a food source to other members of a hive by the direction, length, and
vigor of each burst in a "waggle dance." Martin Lindauer extended
Frisch's observations to cases where a swarm from an overpopulated
colony seeks out a new A worker from the swarm flies out
site.

in search of suitable cavities and returns when it has found and

meticulously explored one. It then performs a waggle dance, on

the surface of the swarm, describing the location and suitability
of its discovery. Meanwhile, other workers tell of other locations.
Promising sites are visited and carefully examined by other members
of the swarm, who return to tell the tale. A worker telling of one
site is unaffected by another bee sending the same message but

can be "converted" by a sufficiently emphatic and repeated display

describing a different location. The debate rages for several days, with
repeated visits to a dwindling number of candidate sites, until near
unanimity is reached. The entire swarm then flies to take up residence
in the winning cavity. This performance might be explained if we
postulate a simple map in the brain of each bee describing locations
and their desirability, maps which can be modified by the complex
experiences of exploration or the simpler ones of communication and
which can become the basis of choice.
An internal model of the world complex enough to allow choices
in behavior —
whether or not we call this model "consciousness"
is what roboticists are currently trying to achieve in their roving
 —

44 Mind Children

robots. In fact, robotics research is too practical to seriously set

itself the explicit goal of producing machines with such nebulous and
controversial characteristics as emotion and consciousness. It would
be enough if our machines could make a living in the face of the many
and competitors they will
surprises, setbacks, opportunities, barriers,
encounter in the world. But natural selection, the guiding mechanism
of Darwinian evolution, is equally utilitarian, and yet here we are,
with feelings and a sense of self.
In The Growth of Biological Thought, the evolutionary biologist Ernst
Mayr points out that both living and nonliving systems "almost
always have the property that the characteristics of the whole cannot
(even in theory) be deduced from the most complete knowledge of
the components, taken separately or in other partial combinations."
Emergence — this appearance of novel properties in whole systems
has often been invoked to explain such difficult biological realities as

mind, consciousness, and even life itself. Here is how imagine some
I

of themore mysterious mental experiences that we associate with

human beings might emerge in our machines as we pursue utilitarian
functionality.

Learning

When tickled, the sea slug Aplysia withdraws its delicate gills into
its body. If the tickling is repeated often, with no ill effect, Aplysia
gradually learns to ignore the nuisance, and the gills remain deployed.
If, later, tickles are followed by harsh stimuli, such as contact with a
strong acid, the withdrawal reflex returns with a vengeance. Either
way, the modified behavior is remembered for hours. Aplysia has
been studied so thoroughly in the last few decades that the neurons
involved in the reflex are well known, and the learning has recently
been traced to chemical changes in single synapses on these neurons.
Larger networks of neurons can adapt in more elaborate ways, for
instance by learning to associate specific pairs of stimuli with one an-
other. Such mechanisms tune a nervous system to the body it inhabits
and to its environment. Vertebrates owe much of their behavioral
flexibility to an elaboration of this arrangement, systems that can be
activated from many locations that encourage and discourage future
repetitions of recent behaviors. Though the neural architecture of these
 Mind in Motion 45

systems in vertebrates is not understood, their effect is evident in the

subjective sensations we call pleasure and pain.
Existing robot systems are, at best, configured to learn a few spe-
cific things from their environment — a simple sequence of moves, the
location of an expected component, the position of nearby obstacles,
sometimes a few parameters for best controlling a motor or interpret-
ing a sensor. There is little point in having them learn to orchestrate
their actions in complicated ways when we can hardly program them
to do one thing at a time well. Yet this primitive state of affairs will not
last forever. Beginning, perhaps, with the universal robot I described
earlier, it will become desirable to add some very general learning
abilities.

A robot's safety and usefulness in a home would be greatly en-

hanced if it could learn to avoid idiosyncratic dangers and exploit
opportunities. If a particular door on a certain route is often locked, it

might be worthwhile if the robot could learn to favor a longer but more
reliable path. A job might be done more effectively if the changing
location of a needed ingredient could be learned or even anticipated
from subtle clues. It is impossible to explicitly program the robot for
every such eventuality, but much could be accomplished by a unified
conditioning mechanism which increased the probability of decisions
that had proven effective in the past under similar circumstances and
decreased it for ones that had been followed by wasted activity or
danger.
The conditioning software I have in mind would receive two kinds
of messages from anywhere within the robot, one telling of success,
the other of trouble. Some — for instance indications of batteries, full

or imminent collisions — would be generated by the robot's basic

operating system. Others, more specific to accomplishing particular
tasks, could be initiated by applications programs for those tasks. I'm
going to call the success messages "pleasure" and the danger messages
"pain." Pain would tend to interrupt the activity in progress, while
pleasure would increase its probability of continuing.
The messages also would provide input to a program that used sta-
tistical techniques to compactly "catalog" the time, position, activity,

surroundings, and other properties known to the robot that preceded

the signal. A "recognizer" would constantly monitor these variables
and compare them with entries in the catalog. Whenever a set of
46 Mind Children

conditions occurred that was similar to those that had often preceded
pain (or pleasure) in the past, the recognizer would itself issue a
somewhat weaker pain (or pleasure) message. In the case of pain, this
warning message might prevent the activity that had caused trouble
before. In time the warning messages themselves would accumulate
in the catalog, and the robot would begin to avoid the steps that led to
the steps that caused the original problem. Eventually a long chain of
associations like this could head off trouble at a very early stage. There
are pitfalls, of course. If the strength of the secondary warnings does
not weaken sufficiently as the chain lengthens, pain could grow into
an incapacitating phobia and pleasure into an equally incapacitating
addiction.
Besides allowing the robot to adapt opportunistically to its environ-
ment, a pleasure-pain mechanism could be exploited by applications
programs in more directed ways. Suppose the robot has a spoken
word recognizer. A module that simply generates a pleasure signal
on hearing the word "good" and a pain message on hearing "bad"
would allow a customer to easily modify the robot's behavior. If the
robot was making a nuisance of itself by vacuuming while a room
was in use, a few utterances of "bad!" might train it to desist until
conditions changed, for instance at a different time of day or when
the room was empty.
A robot with conditioning software could be programmed to train
itself. If a task in an application program required that a certain kind
of container be opened, it would be possible to write a detailed list of
instructions describing just how to hold, turn, and pull to get the job
done. Alternatively, a robot at the factory could be programmed to
pick up many such one after another, and randomly push,
containers,
twist, shake, and pull each one until it either opened or broke. The
trainingprogram would recognize both situations and issue a pleasure
message in one case and a pain signal in the other before going on to
the next container. Gradually the conditioning system would inhibit
those sequences that caused breakage and facilitate those that were
successful. An abstracted version of the training catalog for the session
could then be inserted into the final application program in place of

explicit instructions, combined, perhaps, with catalogs for other parts

of the task developed on other robots.
Infinite patience would be an asset in a training session, but it could
be exasperating in a robot in the field. In the cup-fetching program I
 —

Mind in Motion 47

described earlier, you may have noted that if the robot finds the door
closed and is unable to open it, it simply stands there and repeats
"knock knock" without letup until someone opens the door for it. A
robot that often behaved this way — and many present-day robots do
would do poorly in human company. Interestingly, it is possible to
trick insects into such mindless repetition. Some wasps provide food
for their hatching eggs by paralyzing caterpillars and depositing them
in an underground burrow. The wasp normally digs a burrow and
seals its entrance, then leaves to hunt for a caterpillar. Returning with
a victim, she drops it outside the burrow, reopens the entrance, and
then drags it in. If, however, an experimenter moves the caterpillar
a short distance away while the wasp is busy at the opening, she
retrieves her prey, and then again goes through the motions of opening
the already open burrow. If, while she is doing this, the experimenter
again moves the caterpillar away, she repeats the whole performance.
This cycle can apparently be repeated indefinitely, until either the
wasp or the experimenter drops from exhaustion. A robot could
be protected from such a fate by a module that detects repetitious
behavior and generates a weak pain signal on each repetition. In
the example, the door knocking would gradually become inhibited,
freeing the robot for other pending tasks or inactivity. The robot will
have acquired the ability to become bored.
Modules that recognize other conditions and send pain or pleasure
messages of appropriate strength would endow a robot with a unique
character. A large, dangerous industrial robot with a human-presence
detector sending a pain signal would become shy of human beings and
thus be less likely to cause injury. A module that registered pleasure
on encountering new debris, and pain on seeing it subsequently, might
enable a cleaning program to become very creative and aggressive in
its battle against filth.

Imagery

Fast-learning robots would be able to handle programs that had a great

many alternative actions at each stage of a task — such alternatives
would give the robot a wide margin for creativity. But a robot
with only a simple conditioning system would be a slow learner.
Many repetitions would be required to elicit statistically significant
correlations in the conditioning catalog. Some situations in the real
48 Mind Children

world are unforgiving of such a leisurely approach. A robot that

repeatedly wandered onto a public road, being slow to register the
danger of that location, might suddenly be converted into scrap metal.
A robot, or software, that was slow in adapting to changing conditions
or opportunities in a house could lose the battle for economic survival
against a swifter product from another manufacturer.
Learning could be greatly enhanced by the addition of another major
module, a general world simulator. Now, even the bare-bones universal
robot I outlined uses simulation to some extent. To safely reach its
destination, a program in the universal robot consults its internal map
of the surroundings and considers many alternative paths to find the
best. These ponderings are simulations of hypothetical robot actions.
Similar processes go on when the robot decides how to pick up an
object or when it considers possible interpretations of what it sees
with its cameras. But each of these procedures is specialized, models
only one aspect of the world, and can be used for only one function.
Suppose the robot had a much more powerful simulator that permitted
complex hypothetical situations involving the robot and many aspects
of its surroundings to be modeled. An application program might use
such a simulator to check out a proposed action for safety and efficacy,
without endangering the robot.
But things get really interesting when events in the simulator are fed
to the conditioning mechanism. Then, a disaster in the simulator (for
instance a simulated tumble of the robot) would in real life condition
the robot to avoid the simulated precursor event (let's say loitering
at the head of a simulated stairwell). The robot could thus prepare
for many future problems and opportunities by simulating possible
scenarios in its idle time. Such scenarios might be simply variations
on the day's real events. So equipped, the robot will have the capacity
to remember, to imagine, and to dream.
Imagination via simulator is useful only if the simulator makes rea-

sonably accurate predictions about the real world. Doing so requires

much knowledge about the world, and I imagine that the competitive
development of increasingly good simulators will be a major part of
the research effort of the early twenty-first-century robotics industry.
Robot companies will observe the foibles of their robots in the lab-
oratory and the and tinker with the simulators, the better to
field,

model those aspects of the world important for robot performance.

Evolution did the same for us in the eons of our development. The
 —a

Mind in Motion 49

simulators will come from the factory loaded with generic knowledge,
but they will also be required to leam the idiosyncrasies of each new
location. Advanced robots may find themselves working with other
robots and with people. Such interaction could be made more effective
if the simulators on these machines could predict the behavior of
others to some extent. Part of the prediction might involve roughly
modeling the other's mental state, so that its reactions to alternative
acts could be anticipated. A rich new arena opens up once there
is an internal model of another being's state of mind. For instance,
a module that generated pain messages when it detected distress in
a mental model in the simulator would condition the robot to act
in a kindly manner. And a robot might find itself admonished for
inappropriately ascribing "robotomorphic" feelings and intentions to
other machines, or to humans!
It program robots to commit crimes
would, of course, be as easy to

as to perform socially sanctioned tasks, and legal ways of assigning

blame when this happens will no doubt be devised. But complex
robots will sometimes get into trouble on their own initiative. Imagine
a simulator-equipped robot that has several times in the past suffered
dire consequences from being unable to recharge its batteries in time.
It will thus be especially strongly conditioned against allowing its

power to run low. Suppose it finds itself locked out of its owners'
home, its battery charge fading. The robot's simulator will chum
through different scenarios furiously, searching for a solution —
combination of actions that will result in a recharge. As combinations
of conventional behaviors fail to get it closer to its goal, the simulator
search expands to more unusual possibilities. The neighbors' house is

nearby, its door may be open, and there will be power outlets inside
the simulator discovers a scenario that takes the robot to those outlets.
There is pain associated with leaving home territory, and with the
trouble it may cause, but it is more than balanced by the pleasureand
great release from pain in the possibility of finding a recharge. The
robot repeatedly runs a simulation of the trespass of the neighbors'
house, each time strengthening its conditioning for the steps involved,
making the act itself increasingly Ukely. Eventually the conditioning
is sufficient, and the robot begins on a course that is likelv to lead it

into more trouble than its imperfect simulator anticipated. It will not
be the first creature to have been driven to a desperate act by a great
need.
50 Mind Children

We could carry this speculative evolution further, gradually endow-

ing our feeling robots with intellectual capabilities similar to those of
humans. I expect, however, that by the time the robots are ready for

them, superb intellectual capabilities will be available for wholesale

purchase from the traditional artificial-intelligence industry, which
will have been pursuing its top-down strategy in parallel with the
bottom-up evolution of the robots. The marriage may take many years
to consummate fully, raising issues such as how the reasoning system
can best access the simulator to derive flashes of intuition, and how
reasoning should influence the conditioning system so as to be able
to override the robot's instincts in exceptional circumstances. The
combination will create beings that in some ways resemble us, but in
other ways are like nothing the world has seen before.
 2 Powering Up

D,URING the 1970s, while I was a graduate stu-

dent, seemed to me that the processing power available to artificial
it

intelligence programs was not increasing very rapidly. In 1970 my

work was done on a Digital Equipment Corporation PDP-10 main-
frame computer serving a community of perhaps thirty people. By
1980 my computer was a DEC KL-10, five times as fast and with
five times the memory of the old machine but serving twice as many

users. Worse, the little remaining speedup seemed to have been

absorbed in computationally expensive convenience features: fancier
time sharing and high-level languages, graphics, screen editors, mail
systems, computer networking, and other luxuries that had become
necessities.

Several effects together produced this state of affairs in computing

hardware. Support for university science had wound down
in general

in the aftermath of the Apollo moon landings and the Vietnam war,
leaving the universities to limp along with aging equipment. The
same conditions caused a recession in the technical industries: unem-
ployed engineers opened fast-food restaurants instead of designing
computers. The initially successful problem-solving thrust in artificial
intelligence had not yet run its course, and it still seemed to many that
existing machines were powerful enough —
if only the right programs

could be found. Yet progress in the research itself became slow,

difficult, and frustrating, and many of the best programmers were

drawn into the more rewarding activity of building attractive, but

computationally expensive, software tools, whose success spawned
yet more tool building.
If the 1970s were the doldrums for computing hardware, the 1980s
have more than compensated. Just as artificial intelligence was given
its first boost in the 1960s by the Russian leap into space, the second

51
52 Mind Children

stage was ignited in the present decade by the Japanese leap into
the American marketplace. The Japanese industrial successes focused
attention worldwide on the importance of technology, particularly
computers and automation, in modern economies. American indus-
tries and government responded with research dollars. The Japanese

stoked the fires, under the influence of a small group of senior re-
searchers, by boldly announcing a major initiative toward future com-
puters, the so-called Fifth Generation project, which would expand in
the most promising American and European research directions. The
Americans responded with more money.
Besides this economic boon, integrated circuitry had evolved far
enough by the 1980s that an entire computer could fit on a chip. Sud-
denly computers were affordable by individuals, and a new generation
of computer customers and manufacturers came into being. On the
other end of the scale, supercomputers, once reserved for a handful
of government labs and agencies, became fashionable in hundreds
of industrv and research settings. Across the spectrum of size, the
computer industry became lucrative and competitive as never before,
with new generations of faster, cheaper machines being introduced at
a frenetic rate.
How much must this evolution proceed until our machines
further
are powerful enough to approximate the human intellect? Too little is
known about both the overall functioning of the human brain and how
an intelligent computer would operate to make this estimate directly.
I have approached the problem indirectly by comparing a fragment of

the nervous system that is moderately well understood — the retina of

the eye —
with computer vision programs that do approximately the
same job. 1 then extrapolate the ratio from that comparison to the
whole brain, in order to obtain the computing power required in a
machine that would mimic it. The time of arrival of a machine of that
power is then estimated by extending into the future the trendline
of computer power per unit cost as it has developed during this
century.
The computer estimates 1 will use in making these comparisons
are from my own research. The neurobiology is abstracted from John
Dowling's authoritative book The Retina and Stephen Kuffler and John
Nicholls' classic textbook From Neuron to Brain. The numbers are
precarious because both computer vision and our understanding of
biological vision (not to mention other brain functions) are in their
 Powering Up 53

infancy- Many fundamentals remain mysteries in this complex do-

main. Fortunately, my comparison does not require fiendish precision;
errors of 100 times either way will make little qualitative difference
in relation to the large logarithmic scales of this chapter. Besides, I

also hope that some of my errors will be in opposite directions and

thus partly cancel. There are some dangerous curves in this joyride

to human equivalence, so hold on!

Neural Circuitry

The retina is really an elongated extension of the brain. But its location
at the back of the eyeball, some distance from the bulk of the brain, has
made it comparatively easy to study, even in living animals. Removed
from the body, it can be kept functioning for hours, with its inputs
and outputs highly accessible. Transparent and thinner than a sheet
of paper, the retina can be stained with dyes to make specific neurons
visible to light and electron microscopes. For these reasons, the retina
is probably the best-studied piece of the vertebrate nervous system.
We will look at it in some detail, but first some background about
nerve cells.

like other cells, are daunting mechanisms.

All neurons, They
begin by differentiating from stem cells early in the growth of
life

an embryo, then go through repeated cycles of crawling amoeba-

like to precise destinations throughout the body and dividing and
differentiating further. When they reach their final location, they
extend fibrous growths that seek out specific connections with other
neurons, through junctions called synapses. Different subpopulations
of neurons differ radically in geometry, size, and function. Some
neurons have thousands of small fibers called dendrites and may be
host to hundreds of thousands of synapses. One fiber, known as the
axon, can grow to several centimeters in length, a million times the
cell's original size.

A typical neuron receives messages on its dendrites and issues them

on its axon, which can branch at its end. It signals by means of electri-
cal potential differences of a few millivolts across itsouter membrane.
The voltage is maintained in this wet, electrically conductive environ-
ment by molecular ion pumps in the membrane that move charged
potassium, sodium, calcium, chloride, and other ions in and out of the
cell. The pumps are activated or inhibited by small molecules called
 g

0^ c
c 2j£
O t« C i« ra ;/)
w t/i
0=5
B X
N Oi
CO
<
m i^
E
u C o
m IX
o
/
V /

!/)

^ 3 .f -t-i _>»
~ .f^
«
o «
Si <3
<5
t«
«U
V) ^
i: ^ *- 3
a.
^
"3 •«
c a. <A
c -S «j
3 6 S
a. v> 3 K o
«
Cl :S
.fe<3
Si. i^j
bo ^ 3
V)
3
<«
t; B
Ts -SI,
^ Si O
C
3
AO ^ •« <^
H
a.
O IS
P «^
o H-
C « o
QJ V) .H a =: <-> S: -i
c 5U > ^ ^
I O 2: Ts
T SX3

Pi
'w —
P a f^ 5 **
c
~ • S
«u -c B ^
o c 3
Rj "C
O
H u a. <s to « .5
 Powering Up 55

neurotransmitters produced by other neurons and delivered through a

variety of synapse types.
When a neuron receives a jolt of neurotransmitter, its membrane
voltage may be raised or lowered, depending on the synapse and
neuron type. If the voltage is lowered enough by many signals, a kind
of short circuit happens: the voltage suddenly collapses completely,
and the collapse is propagated up the axon as a pulse. When the pulse
reaches a synapse connecting the neuron to another, it triggers there
a release of neurotransmitter from tiny sacks in its membrane. These
diffuse across the synapse, eventually raising or lowering the voltage
on the second cell. Meanwhile, the pumps in the first neuron work
to restore the original voltage, and in a few thousandths of a second
the cell is ready to fire again. The which pulses are repeated
rate at
encodes the intensity of the stimulus; anywhere from zero to several
hundred pulses per second can be produced. Pulses are used for
long-range communication, but closely spaced neurons, such as those
found in the retina, often communicate simply by responding to each
other's smoothly changing voltages. Besides the synaptic connections
to other cells, many neurons and synapses have receptors for certain
classes of free-floating neurotransmitters, delivered by the blood from
other parts of the nervous system or other body organs, that inhibit
or enhance the neuron's response.
At the nucleus of the neuron, slower genetic processes operate to
manufacture neurotransmitter and to convey it down the axon to
the storage sacks. The neuron's genetic machinery also packages
energy, builds and repairs structure, and does all the other amazing
things any cell must do to keep functioning. Fortunately for those
of us working toward electronic imitations of the nervous system,
most of this complexity is not directly involved in perceiving, acting,
and thinking. Much of the neuron's mechanism is for growing and
building an organism from inside out. Even its information-processing
operations seem to be adapted from this evolutionary necessity, and
it shows.
At this stage in computer technology, it is easier to keep the
construction and repair machinery outside rather than inside the
functional parts. Factories produce integrated circuits and assemble
them into working hardware quite effectively. This removes a great
deal of excess baggage from the final product. Moreover, because
of their roundabout method of operating, neurons are quite slow;
56 Mind Children

they seem incapable of generating much more than 100 signals per
second. These days, electronic switches, always vastly simpler and
now smaller than neurons, can switch as fast as 100 billion times per
second. The great speed advantage of electronics will allow us to
get by with fewer electronic switches than the number of neurons in
the human nervous system. Electronics is also exceptionally precise,
allowing things to be done systematically and efficiently.

Now back to the human what does it actually do? A

retina:

rough and ready answer can be found if we compare the function

of its five different cell types. At the outermost level is a network of
neurons that respond to contrast, motion, other more specific features
of the object under view. Connected to this neural network is a
layer of light-detecting photocells. This type of cell is subdivided into
cone cells, which together discriminate colors, and rod cells, which
do not.
That light must pass through the neural network to get to the
photocells is a peculiar feature of the vertebrate retina — one hit upon
early in the evolutionary history of vertebrates and locked into place.
The independently evolved retinas of the invertebrate octopus and
squid have their photoreceptors up front. The awkward position of
the vertebrate retinal nerve net has greatly limited its size, but strong
selection pressure has efficiency and function. Small
enhanced its

differences in visual acuity or speedmust often have had life or death

consequences among our ancestors, and the retinal neurons are in a
unique position to rapidly and comprehensively abstract the essentials
from an image. The retina is thus likely to be an exceptionally efficient
piece of vertebrate neural machinery.
After adapting to a particular overall light level, clusters of photo-
cells create a voltage proportional to the amount of light striking them.
This signal is received by two classes of cells, the horizontal cells and
the bipolar cells. The horizontal cells, whose thousands of fibers cover
large circular fields of photocells, produce a kind of average of their
areas. If the voltages of all the horizontal cells were mapped onto
a television screen, a blurry version of the retinal image would be
displayed. The bipolar cells, on the other hand, are wired only to

small areas and would provide a sharp picture on the TV. Some of
the bipolar cells also receive inputs from nearby horizontal cells and
then compute a difference between the small bipolar center areas and
the large horizontal surround. Viewed on our TV, their picture would
 Powering Up 57

look much paler than the original, except at the edges of objects and
patterns, where a distinct bright halo would be seen.
The bipolar cell axons connect to complicated multilayer synapses
on the axonless amacrine cells. Each gauglion cell collects inputs from
several of these amacrine synapses and produces a pulsed output,
which travels up its long axon. Each amacrine cell connects to several
bipolar and ganglion cells, and some of the junctions appear to both
send and receive signals. Some amacrine cells enhance the "center
surround" response; others detect changes in brightness in parts of the
image. On the TV, some of these would show only objects moving
left to right, while others would reveal other directions of motion.
Each ganglion cell connects to several bipolar and amacrine and
cells

produces pulse streams whose rate is proportional to a computed

feature of the image. Some report on high contrast in specific parts
of the picture, others on various kinds of motion or combinations of
contrast and motion.
The TV have been referring to
1 is not totally imaginary. Sitting next
to me as write is a TV monitor
I that often displavs images just like
those described. They come not from an animal's retina but from the
eye of a robot. The picture from a TV camera on the robot is converted
by electronics into an array of numbers in a computer memory.
Programs in the computer combine these numbers to deduce things
about the robot's surroundings. Though designed with little reference
to neurobiology, many of the program steps strongly resemble the
operations of the retinal cells — a case of convergent evolution. The
parallel provides a way to measure the net computational power of
neural tissue.

Cells and Cycles

The human retina has 100 million photocells, tens of millions of

and amacrine cells, and a million ganglion cells,
horizontal, bipolar,
each contributing one signal-carrying fiber to the optic nerve. All
this is packaged in a volume a half-millimeter thick and less than

a centimeter square, 1/100,000 the volume of the whole brain. The

photocells interact with their neighbors to enhance one another's
output, and their great multiplicity appears to be a way to maximize
sensitivity; even photon can sometimes produce a detectable
a single
response. The horizontal and bipolar cells and the amacrine cell
58 Mind Children

synapses each seem to perform a unique computation. The bottom

hne, however, is that each of the one milhon ganghon-cell axons
reports on a specific function computed over a particular patch of
photocells.
To find the computer equivalent for such a function, we will first
have to match the visual detail of the human eye in our computer
equivalent. Simply counting photocells in the eye leads to an over-
estimate, because they work in groups. External visual acuity tests
are better, but they are complicated by the fact that the retina has a
small, dense, high-resolution center area, the fovea, which can resolve
details more than 10 times as fine as the rest of the eye. Though it

covers less than 1% of the visual field, the fovea employs perhaps one
quarter of the retinal circuitry and one quarter of the optic nerve fibers.

Under optimal seeing conditions, as many as 500 distinct points can

be resolved across the width of this central region. This feat could
be matched by a TV camera with 500 separate picture elements, or
pixels, in the horizontal direction. The vertical resolution of the fovea

is similar, so our camera would need 500 x 500, or 250,000 pixels, in

all — which, incidentally, just happens to be the resolution of a good-

quality image on a standard television set.
But don't we see more finely than conventional TV? Not exactly.
The 500 X 500 array corresponds only to our fovea, spanning a mere
5° of our field of view. A standard TV screen subtends about 5° when
viewed from a distance of 10 meters. At that range, the scanning lines
and other resolution defects of the TV image are invisible because the
resolution of our eye is no better At closer range, we can concentrate
our fovea on small parts of the TV image to get greater detail, and
this gives us the illusion that we see the whole screen this sharply.

We don't; our unconsciously swiveling eyes simply zip the foveal

area rapidly from one place on the screen to another. Somewhere,
in an as yet mysterious part of our brain, a high-resolution image is
synthesized, like a jigsaw puzzle, from these fragmentary glimpses.
So the foveal circuitry in the retina effectively takes a 500 x 500
image and processes it to produce 250,000 values, some being center-

surround operations, some being motion detections. How fast does

this happen? Experience with motion pictures provides a ready
answer. When successive frames are presented at a rate slower than
about 10 per second, the individual frames become distinguishable.
At faster rates they blend together into apparently smooth motion.
 Powering Up 59

Though the separate frames cannot be distinguished faster than 10

per second, if the hght flickers at the frame rate, the flicker itself is

detectable until it reaches a frequency of about 50 flashes per second.

Presumably in the 10-50 cycle range the simplest brightness change
detectors are triggered, but the more complicated neuron chains do not
have time to react. Movie projectors avoid most of the flicker while
keeping the frame rate reasonably low by using a rotating shutter to
flash each frame more than once. Television does the same thing by
scanning each frame twice, once with the odd numbered scan lines
and once with the even. The peripheral parts of the retina have faster
motion detectors than the fovea (presumably the better to notice fast-
moving dangers coming from the sides), and many people can detect
TV and movie flicker in the corners of their eyes.
In our lab at Carnegie Mellon we have often programmed com-
puters to do center-surround operations on images from TV-toting
robots, and once or twice we have written motion detectors. To get the
speed up, we have spent much programming effort and mathematical
trickery to do the job as efficiently as possible. Despite our best efforts,
10-frames-per-second processing rates have been out of reach because
our computers are simply too slow. With an efficient program, a
center-surround calculation applied to each pixel in a 500 x 500 image
takes roughly 25 million computer calculations, which breaks down
to about 100 calculations for each center-surround value produced.
A motion-detecting operator can be applied at a similar cost. Trans-
lated to the retina, thismeans that each ganglion cell reports on the
computer equivalent of 100 calculations every tenth of a second and
thus represents 1,000 calculations per second. The whole million-fiber
optic nerve, then, is a conduit for the results of 1 billion calculations
per second.
If the retina's processing can be matched by 1 billion computer
calculations per second, what can we say about the entire brain? The
brain has about 1,000 times as many neurons as the retina, but its

volume is 100,000 times as large. The retina's evolutionarily pressed

neurons are smaller and more tightly packed than average. By mul-
tiplying the computational equivalent of the retina by a compromise
value of 10,000 for the ratio of brain complexity to retina complexity,
I rashly conclude that the whole brain's job might be done by a
computer performing 10 trillion (10^^^) calculations per second. This is

about 1 million times faster than the medium-size machines that now
60 Mind Children

drive my robots, and 1,000 times faster than today's best supercom-
puters.
Estimates like these are vulnerable to attack from many directions
(see Appendix 1). After all, controversy flares when one merely
compares similar electronic computers, whose internal operations are
well understood and whose performance can be tested in detail.
Hence it would be foolish to expect consensus opinion about a
comparison of radically different systems executing dimly understood
functions. Nevertheless, my estimates can be useful even if they are
only remotely correct. Later we will see that a thousandfold error in
the ratio of neurons to computations shifts the predicted arrival time
of fully intelligent machines a mere 20 years.

Memory
Having settled on a lO-trillion-operation-per-second (10 teraops) com-
puter as a sufficiently powerful host for a humanlike mind, we still

have to decide how much memory to include. In 1953 the IBM 650
computer performed and was equipped
1,000 instructions per second
with 1,000 "words" memory, each able to store one number, or
of
one instruction. In 1985 the Cray 2 ran at up to 1 billion instructions
per second and was packed with up to 1 billion words of memory.
This ratio, shared by most computers, of about one memory word for
each instruction per second of speed was shaped by the market and
probably indicates the size necessary to contain problems sufficiently
large to keep a computer busy for seconds to hours at a time rates —
comfortable for human programmers. If it had this ratio, a humanlike
computer would require 10 trillion words of memory, about 10'"" bits.
(A bit, or binary digit, is a tiny unit of information that encodes a
choice between two equal possibilities. Computer words today are
between 16 and 64 bits long. Larger machines tend to have longer
words.)
But is this number compatible with what is known about the
nervous system? During the last decade Eric Kandel of Columbia
University and others have studied the cellular changes that occur
in the sea slug Aplysia when it is conditioned by irritating stimuli.
They found that learning manifests itself as long-lasting chemical
changes in individual synapses between neurons, changes that affect

the strength of the connections to other neurons. Each synapse can

 Powering Up 61

store only one such strength, and then only with limited precision.
If we —
assign 10 bits enough to represent a number to three decimal
places of accuracy — to each synapse, and if this storage method is

substantially correct for larger nervous systems, then the 10'"' bit

"standard" memory of a humanlike computer should be able to

contain the information encoded in the 10'"* synapses of a human
brain.

Comparative Computational Power and Memory

Some natural and artificial organisms rated by the measures of this
chapter. Current laboratory computers are roughly equal in power to
the nervous systems of insects. It is these machines that have hosted
essentially all the research in robotics and artificial intelligence. The
largest supercomputers of the late 1980s are a match for the 1-gram
brain of a mouse, but at $10 million or more apiece they are reserved
for serious work.

10

10

T3
C

10 Microwave
noise
TV-guided
missile
|Tele\ision
§ channel
10
i Macintosh Video recorder

o
CI.
Radio
channel
Bacterial
10'
reproduction

Calculator Encyclopedia Library of Congress

Bacterial Human
ywoo^ Aiooo^ AMv^ -^

12 15 18
10 10 10 10 10 10

Capacity (bits)
62 Mind Children

Comparing Computers

It is easy to see that computers are becoming more powerful, but by

how much and how fast? Just when can we expect 10 teraops in
a package sized and priced to fit in an autonomous robot? When
I first approached this question, it seemed natural to rate electronic

computers in operations per second, beginning with the first ones in

the 1940s, and to project the resulting curve into the future. But there
were complications. The machines came in many sizes, with prices
ranging from tens of dollars to tens of millions of dollars. A given
model could be equipped with many options, more memory, auxihary
processors, faster input and output, and so on. Recent machines are
—
sometimes multiprocessors multiple computers working in lockstep,
or separately, sharing data. Machines had different instruction sets, so
that an operation that took 10 instructions on one might be done on
another in a single step. Some computers worked with numbers only
5 decimal places long; others handled 20 digits at a time. Also, the
literature on the early electronic machines led me to their predecessors,
computers built of telephone relays, the electromagnetic switches that
had been perfected for telephone exchanges. Research on those,
in turn, suggested yet earlier machines that calculated with motor-
driven, and even hand-cranked, gears and cams. If these manual
machines could somehow be compared with automatic computers,
my curve could be extended back in time to the nineteenth century.
As a first step in devising a useful measure, I decided to cancel
the size differences between machines by dividing each machine's
processing power by its price, in constant dollars. This would give
me an estimate of cost-effectiveness. For the mechanical calculators I
added the human operator (valued at 100,000 1988 dollars— as if salary
were problem a manual
a leasing cost) to the price, since to solve a
calculator needs a human to steadily enter numbers and operations
and to write down results. This approach allowed the cost of purely
—
manual calculation to also be measured an unaided human clerk,
whose effective capital cost is $100,000, can do about one calculation
a minute!
Step two was to figure out just how factors such as speed, memory
size, and instruction repertoire affected a machine's processing power.
This was slippery. Computers today are often compared by measuring
the running time of large sets of standard programs on each. This
 Powering Up 63

route was not open to me, since most of the machines I hoped to
include in my curve no longer exist. I did know how long most of the
machines took add and to multiply
to two numbers, how many words
of memory each had and the size of a word, and the approximate size
of each machine's instruction repertoire. Processing power was to be
the amount of computation done by the machine in a given time. If I
could estimate how much computation each instruction accomplished,
on the average, I would merely have to multiply by the number of
instructions executed per unit time to get total power. So the problem
reduced itself to estimating the work done by a single instruction.
Suppose a child's story begins with the words: Here's my cat. It

has fiir. It has claivs... Pretty boring, right? Imagine, now, another
story that starts out with: Here's my cat. It wears a hat. It totes a

gun... Better. seems more interesting and informative

The second story
because its later statements are less likely cats usually have fur and —
claws, but they rarely carry hats and guns. In 1948 Claude Shannon
of MIT formalized such observations in a mathematical system that
came to be known as information theory. One of its key ideas is that
the information content of a message goes up as its likelihood, as
measured by the recipient, decreases (mathematically, as the negative
logarithm of the probability). A series of messages has maximum
information content when it is maximally "surprising."
My measure of effective computation works the same way. Each
instruction executed by a machine is like a message. The more
predictable its sequence of instructions, the less useful work a machine
is doing. For instance, a program that causes a computer simply to

add one to a memory location once every millionth of a second is

doing almost nothing of consequence. The contents of the memory
location at any time in the future are known in advance. But even the
best programs are limited in how much "surprise" they can introduce
into a computation at every step. Each instruction can specify only
number of different possible operations and choose from a
a finite
number of memory locations, each itself containing only a finite
finite

number of possibilities. These sources of surprise can be combined

using the formulas of information theory to express the maximum
information content of a single computer instruction.
I detail such a calculationin Appendix 2. The numbers vary from

machine to machine and program to program, but I conclude there

that a typical computer running an exceptionally efficient program
 Cost of hardware for human equivalence (1988$)

a
B
o
U
Mm
o

3
**
C
0)
U
($886l/P^0D3s/siiq) }soD 4iun jad laMod iBuoi4C4ndui03
 Powering Up 65

produces about 50 bits of surprise for each operation performed. If

the computer can do 1 milUon operations per second, its maximum

computational power is about 50 milUon bits per second. Expressed in
these units, the computational power required in a human-equivalent
robot is about 10''* bits per second.

Projections

The on page 64 plots the number of bits per second of compu-

figure
tational power provided per (constant 1988) dollar of purchase price
by a number of notable computing machines from 1900 to the present.
Although numerous mechanical digital calculators were devised and
built during the seventeenth and eighteenth centuries, only with the
mechanical advances of the industrial revolution did they become re-

liable and inexpensive enough to rival manual calculation. By the late

nineteenth century their edge was clear and the continuing progress
dramatic. The vertical scale in the figure is logarithmic equal steps —
represent tenfold increases in the ratio of performance to price. Since
1900 there has been a trillionfold increase in the amount of computation
a dollar will buy. Each of the machines in the figure has a fascinating
story; but since this book is not primarily a history of computation, I
will restrict myself to a few highlights. The Origins of Digital Computers:
Selected Papers, edited by Brian Randell, contains excellent first-hand

accounts of many of these early machines.

Charles Babbage, of Cambridge, England, conceived the idea of
an automatic program-controlled calculating machine in 1834, almost
a century before anyone else. This "Analytical Engine" was to be
a steam-engine-powered calculating behemoth of gears and shafts
dealing in 50-digit decimal numbers. A rack of cogwheels was to
store 1,000 such numbers, and a calculating unit was to be able to add
two numbers in less than ten seconds and to multiply them in under
a minute.The machine was to be controlled on the small scale by
slowly rotating pin-studded drums such as those that still pluck the
reeds in mechanical music boxes, and on a coarser scale by a stream of
punched cards specifying memory locations and arithmetic operations
to be performed with their contents.
In concept, the Analytical Engine contained all the elements of
a modern digital computer. Babbage worked on it for the last 37
years of his life, but it was never completed. The enormous scale of
66 Mind Children

the project and the tender state of the mechanical art (components
were still typically hand fitted) made it unlikely that he could have
succeeded in his lifetime. Precision interchangeable parts were much
more common by the early twentieth century, and in 1910 Babbage's
youngest son was able to demonstrate a working portion of the central
calculating unit, although he did not complete the entire machine. I

have included the Analytical Engine as a 1910 data point in my figure

since it is likely that the machine could have been built at that time if

a pressing need had arisen.

The other mechanical calculators in the chart were manually op-
erated and were sold primarily to businesses for use by clerks and
accountants, though some found uses in science. As mentioned pre-
viously, I included the "price" of the human operator in the cost of
calculation for these. The early improvements in speed and reliability

came with advances in mechanics: precision mass-produced gears

and cams, for instance, improved springs and lubricants, as well as
increasing design experience and competition among the calculator
manufacturers. Powering calculators by electric motors provided a
boost in both speed and automation in the 1920s, as did incorporating
electromagnets and special switches in the innards in the 1930s.
Leonardo Torres y Quevedo, a Spanish inventor, demonstrated an
electromechanical calculator in Madrid in 1919. Controlled by simple
arithmetical commands entered at a typewriter keyboard, the "Torres
Arithmometer" lacked memory and was not fully automatic in the
a
modern was close, and it could have been converted to
sense. Yet it

automatic operation by addition of a tape unit for entering commands,

and made practical by addition of a unit to store and recall a handful
of numbers.
Konrad Zuse independently invented the idea of programmed
calculation as a young man in Germany in 1934; he built several
large, automatic electromechanical computers in his parents' living
room. The third machine in the series, built with backing from the
German government and completed in 1941, was a complete, tape-
controlled, binary floating-point (meaning it represented numbers in
scientific notation, allowing for very small and very large numbers)

computer with a 64-word memory. Zuse formed a company that sold

improved models in the years following the war.
The Bell Telephone Labs (BTL) machines were built using telephone-
exchange relay-switching techniques. The first two were modest
 Powering Up 67

internal projects, built to test antiaircraft gun directors. The third

was a massive general-purpose, tape-controlled, automatic computer,
intended as a commercial product. was overtaken by much faster
It

electronic The huge Harvard-

machines and was never successful.
IBM machines were of similar construction and suffered a similar fate.
The era of general-purpose relay computers was over almost before it
began.
One class of electromechanical machine had a longer histor\'. The
Constitution of the United States specifies that a national census
must be taken every decade. As the country grew, these censuses
took longer and longer to tally. The 1880 results were still being
organized in 1887. It was obvious that without improved techniques
the 1890 census would last beyond 1900. The Census Office held a
competition for a better svstem. The winner was a young engineer
named Herman Hollerith, who devised machinery that automatically
counted holes in punched cards. Over the next half century Hollerith's
invention evolved into a battery of "tabulating" machines that sorted
and interleaved punched cards, duplicated them, printed on and
from them, and did calculations with their contents. Hollerith's
company grew into International Business Machines, which to this
dav represents 70% of the computer industrv.
Electronic tube computers using radio and ultrafast radar techniques
appeared as government-funded projects toward the end of World
War II. The first commercially manufactured machine of this kind
was the UNIVAC I, and its first customer was the Census Bureau,
in 1951. By the end of the 1950s there were about 6,000 computers

overall in industry, government, and universities. Their electronics

was built around vacuum tubes, and thev became known as the "first
generation" of computers.
in about 1960, a second generation of machines began
Beginning
to appear that used newly developed transistors in place of vacuum
tubes. They were smaller, more reliable, and cheaper and used less
electricity than the vacuum tube computers, while providing more
speed and memorv.
By the late 1960s IBM began to introduce a third generation of ma-
chines using "hybrid integrated circuits." Dozens of tiny, unpackaged
transistors and other components were bonded onto wiring
electronic
printed on ceramic chips the size of a thumbnail. Over the next several
years these hybrid chips gave way to "monolithic" integrated circuits.
68 Mind Children

in which dozens of components were etched directly into silicon chips

a few millimeters square.
Integrated circuit technology developed rapidly, and by the mid-
1970s a chip could contain thousands of components. A fourth
generation of computers, whose heart was a handful of such chips,
appeared but was quickly eclipsed by the microprocessor, a chip
with tens of thousands of components that was, by itself, a complete
computer. Progress was now so bewilderingly fast and multifaceted,
with computers appearing in everyday devices such as microwave
ovens, that the industry gave up on the generational nomenclature.
(The last vestige was the Japanese Fifth Generation project, a research
effort to develop artificially intelligent machines.) I am typing these
words on a Macintosh II computer, a machine containing manv chips
with over a million components each, and a machine without a
generation!

Human Equivalence in 40 Years

The progress documented in the figure on page 64 is remarkably

steady despite radical changes in the nature of computing this century.
The amount of computational power that a dollar can purchase has
increased a thousandfold every two decades since the beginning of
the century. In eighty years, there has been a triUionfold decline in the
cost of calculation. If this rate of improvement were to continue into
the next century, the 10 teraops required for a humanlike computer
would be available in a $10 million supercomputer before 2010 and in

a $1,000 personal computer by 2030.

But can this mad dash be sustained for another forty years? Easily!
The curve in the figure is not leveling off, and the technological
pipeline contains laboratory developments that are already close to
my requirements. To a large extent, the slope of the figure is a self-
fulfilling prophecy Integrated circuit manufacturers have been aware
of the trend since Gordon Moore, one of the inventors of the integrated
circuit, noted in 1963 that the number of components on a chip was

doubling each year. Computer makers have had similar observations,

and new products in both of these related fields are designed with
the trend in mind. Established manufacturers design and price
products to stay on the curve, to maximize profit; new companies
 Powering Up 69

aim above the curve, to gain a competitive edge. The industry's

—
success is one reason the success can continue its enormous, and
rapidly increasing, wealth supports more and better research and
development of further advances. Also, the very computers that the
industry makes are employed in the design of future circuits and
computers. As they become better and cheaper, so does the design
process, and vice versa. Electronics is riding these vicious cycles so
quickly that it is hkely to be the main occupation of the human race
by the end of the century.
A key driver of both this decline in price and gain in performance
is miniaturization. Small components simultaneously cost less and
operate more quickly. Charles Babbage realized this in 1834. He
wrote that the speed of his Analytical Engine, which called for
hundreds of thousands of mechanical components, could be increased
in proportion if "as the mechanical art achieved higher states of
perfection" his palm-sized gears could be reduced to the scale of
clockwork, or further to watchwork. (Try to imagine our world if
electricityhad not been discovered and the best minds had continued
on Babbage's course. By now there might be desk- and pocket-
sized mechanical computers containing millions of microscopic gears,
computing at thousands of revolutions per second.)
To a remarkable extent the cost per pound of machinery has re-
mained constant as the machinery has become more intricate. This is
as true of consumer electronics as of computers (merging categories
in the 1980s). The radios of the 1930s were as large and expensive
as the televisions of the 1950s, the color televisions of the 1970s, and
the home computers of the 1980s. The volume required to amplify or
switch a single signal dropped from the size of a fist in 1940 to that of
a thumb in 1950, to a pencil eraser in 1960, to a salt grain in 1970, to a

small bacterium in 1980. In the same period, the basic switching speed
rose a millionfold and the cost declined by the same huge amount. I
cannot tell you exactly what developments will yield the additional
factor of a million project I —
such predictions are impossible for many
reasons. Entirely new and unexpected possibilities are encountered in
the course of basic research. Even among the known contenders, many
techniques are in competition, and a promising line of development
may be abandoned simply because some other approach has a slight
edge. I can tell you that there are experimental components in lab-
70 Mind Children

oratories today that improve on the best commercial components a

thousandfold, at least in speed and size. Here is a short list of what

looks promising today.

In recent years the widths of the connections within integrated cir-
cuits have shrunk to less than one micrometer, perilously close to the
wavelength of the light used to print the circuitry. The manufacturers
have switched from visible light to shorter wavelength ultraviolet,
but this gives them only a short respite. X-rays, with much shorter
wavelengths, would serve longer, but conventional x-ray sources are
so weak and diffuse that they need uneconomically long exposure
times. High-energy particle physicists have an answer. Speeding
electrons curve in magnetic fields and spray photons like mud from
a spinning wheel. Called synchrotron radiation for the class of particle
accelerator where it became a nuisance, the effect can be harnessed
to produce powerful beamed x-rays. The stronger the magnets, the
smaller the synchrotron. With supercold superconducting magnets,
an adequate machine can fit into a truck; otherwise it is the size of
a small building. Either way, synchrotrons are now of hot interest
and promise to shrinkmass-produced circuitry into the submicron
region. Electron and ion beams are also being used to write submicron
circuits, but present systems affect only small regions at a time and

must be scanned slowly across a chip. The scanned nature makes

computer-controlled electron beams ideal, however, for manufacturing
the "masks" that act like photographic negatives in circuit printing.
Smaller circuits have less electronic inertia; they switch faster and
need lower voltages and less power. On the negative side, as the
number of electrons in a signal drops, the circuit becomes more prone
to thermal jostling. This effect can be countered by cooling, and indeed
fast many labs now run in supercold liquid
experimental circuits in
nitrogen. One supercomputer is being designed to operate this way.
Liquid nitrogen is produced in huge amounts in the manufacture

of liquid oxygen from air, and it is cheap (unlike the much colder
liquid helium). Uneven clumping of key impurities results in erratic
component values as circuits get smaller, so more precise methods
for implanting them are being developed. Quantum effects become
more pronounced, creating new problems and new opportunities.
Superlattices — multiple layers of atoms-thick regions of differently
—
doped silicon made with molecular beams are such an opportunity.
They allow the electronic characteristics of the material to be tuned
 —

Powering Up 71

and permit entirely new switching methods, often giving tenfold

improvements. Even more exciting are "quantum dot" devices that
exploit the wavelike behavior of small numbers of electrons trapped
in regions smaller than the electron wavelength.
The were made of germanium; they could not
first transistors

withstand high temperatures and tended to be unreliable. Improved

understanding of semiconductor physics and ways of growing silicon
crystals made possible faster and more reliable silicon transistors

and integrated circuits. Newer now coming into their

materials are
own. The most immediate is gallium arsenide. Its crystal lattice
impedes electrons less than silicon and makes circuits up to ten times
faster. The Cray 3 supercomputer, scheduled to appear late in 1988,

uses gallium arsenide integrated circuits packed into a one-cubic-

foot volume, to top the Cray 2's speed tenfold. Other compounds
hke indium phosphide and silicon carbide wait in the wings. Pure
carbon in diamond form is a definite possibility; it should be as
much an improvement over gallium arsenide as that crystal is over
silicon. Among its many superlatives, perfect diamond is the best
solid conductor of heat, an important property in densely packed
circuitry. The vision of an ultradense three-dimensional circuit in a

gem-quality diamond is compelling. As yet no working circuits of

diamond have been reported, but the odds improved in 1987 with
reports from the Soviet Union, Japan, and, belatedly, the United States,
of diamond layers up to a millimeter thick grown from microwave-
heated methane.
The ultimate circuits may be superconducting quantum devices
not only extremely fast but highly efficient. Superconducting circuits
have been in and out of fashion over the past twenty years. They have
had a tough time because the liquid helium environment they required
until recently is expensive, the heating /cooling cycles were stressful,
and especially because rapidly improving semiconductors offered
such tough competition. The newly discovered high-temperature
ceramic superconductors could solve many of these old problems all

at once. A superconducting transistor announced by Bell Laboratories

early in 1988is less than one twentieth of a micrometer in size and able

to switchbetween on and off states in a picosecond (one trillionth of a

second) on receiving an input signal of only one electron! A thousand
microprocessors made of such switches would fit in the space occupied
by one of today's microprocessor chips, and each would be a thousand
71 Mind Children

times as fast. A thousand processors, each a thousand times faster

than today's, would have just about the 10 teraops needed for human
equivalence.
Farther off the beaten track are optical circuits that use lasers
and nonlinear optical effects to switch light instead of electricity.
Switching times of a few picoseconds, a hundred times faster than
conventional circuits, have been demonstrated, but many practical
problems remain. Finely tuned lasers have also been used with light-
sensitive crystals and organic molecules in demonstration memories
that can store up to a trillion bits per square centimeter.
Underlying these technical advances, and preceding them, are
equally amazing advances in the methods of basic physics. One un-
expected, and somewhat unlikely, device is the inexpensive scanning
tunneling microscope that can rehably see, identify, and manipulate
single atoms on surfaces by scanning them with a sharp needle. The
tip is positioned by three piezoelectric crystals that stretch micro-
scopically under the influence of small voltages. A gap of a few
atoms in size is maintained by monitoring the current that quantum-
mechanically tunnels across it. The tunneling microscope provides a
secure toehold on the atomic scale, and big ideas about little atoms
are being pursued by enthusiasts in both semiconductor and biotech-
nology laboratories.
Living organisms are clearly machines when viewed at the molec-
ular scale —
in them information encoded in RNA "tapes" directs pro-

tein assembly devices called ribosomes to pluck specific sequences of

amino acids from the environment and attach them to the ends of
growing chains of protein. Proteins, in turn, fold up in certain ways,
depending on the sequence of their amino acids, to do many jobs.
Some proteins have moving parts like hinges, springs, and latches
triggeredby templates. Others are primarily structural, like bricks or
ropes or wires. The proteins of muscle tissue work like ratcheting
pistons.
Today's biotechnology industry depends on modest manipulations
of natural genetic machinery. The visionaries have more elaborate
plans — nothing less than the fusion of biological, microelectronic,
and micromechanical techniques into a single, immensely powerful,
new technology. Computer-modeling techniques are slowly becoming
powerful enough to allow new proteins to be designed and tested
on display screens, much as conventional machine parts are often
 Powering Up 73

developed now. Such engineered proteins, as well as existing protein

mechanisms copied from living cells, could be assembled into tiny
artificial machines. Early products might be simple tailored medicines

and small experimental computer circuits. Gradually, though, accu-

mulated tools and experience would allow the construction of more
elaborate machinery, eventually as complicated as tiny robot arms and
equally tiny computers to control them. These would be small enough
to grab individual molecules and hold them, thermally wriggling,
in place. The protein robots could then be used as machine tools
to build a second generation of even smaller, harder, and tougher
devices by assembling atoms and molecules of all kinds. For instance,
carbon atoms might be laid, bricklike, into ultrastrong fibers of perfect
diamond. The entire scheme has been called nanotechnology, for the
nanometer scale of its parts. By contrast, today's integrated circuit
microtechnology has micrometer features, a thousand times bigger.
Some things are easier at the nanometer scale. Atoms are perfectly
uniform in size and shape, if somewhat fuzzy, and behave predictably,
unlike the nicked, warped, and cracked parts in larger machinery.
At the nanometer scale, the world is stocked with an abundance of
components of absolute precision.
Atomic-scale machinery is a wonderful concept and would take us
far beyond the humanhke point in computers, since it would allow

many millions of processors to fit on a chip that today holds but

one. Just how fast could each individual nanocomputer be? Quantum
mechanics demands a minimum energy to localize an event to a given
time: Energy = h j time where h is Planck's fundamental constant of
quantum mechanics. Higher speeds require greater energy. Above the
frequency of light, about a quadrillion (10^^) transitions per second, the
energy reaches one electron volt, close to the energy of the chemical
bonds holding solid matter together. Attempts to switch faster will
tear apart the switches. A quadrillionth of a second, or femtosecond,
is nanosecond (billionth of a second)
a million times faster than the
switching time of the fastest commercial computer components today,
so a single nanocomputer might have a processing speed of a trillion
operations per second. With millions of such processors crammed
onto a thumbnail-size chip, my human-cc]uivalence criterion would
be bested more than a millionfold! That might seem to be enough,
but I cannot help wondering if, just maybe, speeds beyond this "light
barrier" are possible.
7^ Mind Children

The physics world is a turbulent one, as theorists chase a goal that

eluded Einstein: a single theory that encompasses all the types of
particle and energy in nature. The front-runner these days is the
superstring model; in it, particles are tiny loops of space knotted in
six dimensions beyond the four of normal spacetime. Its variations
making up atoms, some
predict a host of particles heavier than those
of them stable. Material made of such particles would be from a
thousand times to astronomically denser and more tightly bonded
than normal matter. Ultradense matter could, in principle, support
switching operations much more rapid than the frequency of light.
The benefits of miniaturization need not stop at the atomic scale!
While ultradense matter that is stable on earth is just a speculation,
vast quantities of similar stuff is known to exist in the tremendous
gravity fields of collapsed white dwarf and neutron stars. Someday,
our progeny may exploit these bodies to build machines with a million
million million million million (that's 10^°) times the power of a human
mind.
 Symbiosis

X HE robot who
century will have some interesting properties.
will work alongside us
Its
in half a
reasoning abilities

should be astonishingly better than a human's — even today's puny

systems are much better in some areas. But its perceptual and motor
abilities will probably be comparable to ours. Most interestingly, this

artificial person will be highly changeable, both as an individual and

from one of its generations to the next.
But however competent, are only part of the
solitary, toiling robots,

story. some decades into the future, the most effective

Today, and for
computing machines work as tools in human hands. As the machinery
grows in flexibility and initiative, this association between humans
and machines will be more properly described as a partnership. In
time, the relationship will become much more intimate, a symbiosis
where the boundary between the "natural" and the "artificial" partner
is no longer evident. This collaborative route is interesting for its
powerful human consequences even if, as I believe, it will matter
little in the long run whether or not humans are an intimate part of

the evolving artificial intelligences.

We will begin our exploration of the symbiotic path with a history of

its humble beginnings, the minimal interfaces of the earlv computers.

Stored Programs and Assemblers

ENIAC, the first general-purpose electronic digital computer, built in

Philadelphia in 1946, was designed for hardwired control. A typical
program involved thousands of wires connected by hand from point
to point on large programming boards. "Writing" such a program was
tedious; debugging one was daunting. Onlv a few were written before
John von Neumann made a now-famous proposal. ENIAC had three

75
le Mind Children

large banks of dial switches, called function tables, intended for storing
precomputed mathematical results needed during a calculation. They
might be set up, for instance, with square roots or logarithms or
with more specialized functions. Von Neumann suggested that these
switches could be used in a different way, namely, to hold sequences
of instructions, encoded as numbers, that would direct the machine's
operation. The regular hardwired program would be set up, once and
for all, in such a way that it could read these instructions from the
function tables one after another and do what the numbers indicated.
Thenceforth the machine could be programmed for new tasks simplv
by dialing commands into the function tables.

ENIAC
The rat's nest of wires at the left are the machine's original program-
ming hoards. The banks of switches on the right were intended
three
to hold mathematical function tables but were soon enlisted as a more
convenient way to represent programs.
 Symbiosis 77

This new mode of programming was much easier and neater than
the original scheme. Instead of a rat's nest of wires, a program
consisted of neat columns of numbers. The numerical encoding
system for machine operations came to be called machine language.
It was one small step for user-friendly computers, a giant leap for
computer organization.
All digital computers after ENIAC incorporated an expanded version
of this stored program idea. Not only were programs represented
as sequences of numbers, but the numbers were kept in the same
memory used for calculations and could be loaded at high speed
from input devices such as punched paper-tape readers. This unity of
memory allowed a computer to modify its own program in midrun,
an intriguing technique that was used extensively at first but is less
common now. ENIAC used a large roomful of vacuum tubes to store
fewer than fifty numbers, and could carry out about one thousand
calculations per second.
ENIAC's successors were able to store entire programs in their
working memory because new methods were invented for more
densely and cheaply storing the electronic tally marks that constituted
the memory. In some, a device much like a television picture tube was
able to retain thousands of such "bits" as tiny areas of electric charge
on its glass face. A
sweeping electron beam could both sense and alter
the contents of each area. In others, thousands of bits were encoded
as a recirculating stream of acoustic pulses that traveled down a long
column of mercury, to be sensed electronically at the end of their
and
journey, amplified, re-injected back into the head of the column.
Another approach was to record the bits magnetically on the surface
of a rapidly spinning drum or disk. Magnetic disks evolved into the
bulk external storage devices still in use today, but they were too slow
to survive long as the internal working memory of computers.
The most successful method turned out to be one that used tiny
donuts of a specially developed magnetic material strung on the inter-
sections of a net of fine wires. Each of these magnetic cores could store
one encoded as either a clockwise or counterclockwise magneti-
bit,

zation. The direction of magnetization of a particular magnetic core

could be changed by sending small currents through the horizontal
and vertical wires that passed through it. Its previous contents could
be determined with the help of another wire that zig-zagged through
all the cores. Whenever any core's magnetization flipped from a
78 Mind Children

clockwise to a counter-clockwise direction, or vice versa, a tiny pulse

appeared on this wire. Magnetic core memories holding thousands
and eventually millions of numbers were the main form of computer
working storage for over twenty years, until they were overtaken in
the mid 1970s by storage circuits much like those of ENIAC, but made
of transistors instead of tubes and arrayed in thousands on tiny silicon
chips.
Machine-language programming was a great labor saver. But in-

creased memory, speed, and availability of computers soon lured their

users to problems so large and complex that even machine language
became unbearably tedious. A machine-language program consists
of a sequence of instructions encoded as numbers. A
few digits of
each number, the operation code, specify the action the computer is
to perform, say to add or to obtain the next instruction from else-
where in memory. The remaining digits contain the address of one
or more locations in memory indicating, in our examples, where to
find the numbers to add or the next instruction. Converting problems
originally formulated in terms of algebraic expressions like x^ + y
to numerical codes that the computer could use was an unintuitive,
slow, and error-prone process. Even worse, inserting a few additional
instructions into a program might require shifting the location of much
of the rest of the program and of number storage areas in memory,
which in turn would require the alteration of the address parts of
many, perhaps thousands, of instructions. The slightest error in the
process could prevent the program from working. Programmers who
became skilled at this exacting drudgery, punctuated by bursts of
artful invention, were sometimes treated with the deference accorded
to chess masters. But it did not long escape notice that computers
themselves specialize in just this sort of exacting drudgery.
By the mid-1950s programmers were writing large programs whose
function was to translate symbolic commands (like ADD X) into
machine language while automatically assigning and keeping track of
the location of variables and instructions. Machine-language purists
complained that assemblers, as programs of this kind came to be called,
lessened the precise control one should have over the workings of
a computer and wasted computer time doing the translation. Even
so, symbolic programming proved such a boon that writing machine

language quickly became an extinct art. The path connecting humans

and computers had widened again, allowing an increased flow of
 Symbiosis 79

everyday traffic. The road enhancements also facilitated the passage

of heavier road-building machinery.

Compilers and Operating Systems

Assemblers were a great help to professional programmers but were

still very tedious for occasional computer users with particular prob-
lems to solve. Such end-consumers of computer power requested,

and were granted, high-level languages that let compact mathematical

notation, such as they used routinely in their work, substitute for
longer, and more error-prone, assembler sequences. For instance, in
a high-level language such as FORTRAN, which was one of the first
and still survives today, A x X + B might stand in for the assembly
sequence:

LOAD A
MULT X
ADD B

Very complex programs called compilers, tours-de-force of program-

ming when they were first written, translated lines of the high-level
language into often long machine-language sequences. The user of a
high-level language had little or no knowledge of the final program
produced. Compilers used even more of the computer's valuable time
than did assemblers and, lacking the cleverness and insight of human
programmers, produced machine language that was larger and slower
running than that produced when the same problem was written in
assembly language. The disadvantages were great enough that, until

recently, many critical still demanded assembly language.

applications
But high-level languages hadmany advantages for the average
user. They made programs easier to write, and many errors could
be detected during the translation instead of causing obscure failures
when the program was run. Since high-level languages were close
in form to conventional mathematical notation, many computer non-
specialists found them easier to learn. A transcendent feature was
transportability. Unlike machine language or assembly programs, high-
levelprograms did not reflect the detailed workings of any particular
machine and so could be translated for entirely different computers.
FORTRAN compilers, for instance, exist for essentially every computer
ever built, and certain useful FORTRAN programs have survived the
80 Mind Children

entire history of computers, leaping from one generation of machine

to the next.
On the first digital computers, setting up a program, monitoring
its progress,and cleaning up afterward was a strictly manual affair.
The programmer could watch the action on banks of lights showing
the internal state of the machine and could interrupt, examine, and
alter it at any time, or run it step by step. This was quite convenient

but expensive on machines whose time was valued at hundreds of

dollars per hour. To minimize the time lost, computer companies in
the late 1950s began supplying programs called monitors, supervisors,
or operating systems to manage the flow ot successive programs, read
from punched-card readers or magnetic-tape units, through their
machines. The early operating systems had a no-nonsense approach:
one program at a time was run; if any problem was encountered it
was automatically stopped, the contents of memory were printed out,
and the next program was started. The core dump (as the memory
printout was called, after magnetic core memory) was delivered to the
programmer, who could diagnose the problem on his or her own time.
Many programs spent only seconds in the computer, producing results
that would occupy the programmer hours or days in preparation for
the next run.
From the mid-1950s to the late 1960s this batch mode of computer
operation was the rule, and a generation of computer users, especially
users of IBM equipment, knew no other way. Some of the old timers,
however, wistfully remembered the days of hands-on use. Not only
was tracking down a programming bug easier with the computer as an
ally, but it was possible to write programs that carried on a dialogue
with the user. Interactive programs let the user and the computer
act as partners, often with the user supplying insight and judgment
and the computer providing prodigious calculation and memory. The
problem was how to arrange this level of service without having the
computer waste most of its time waiting for the human's next move.
Academic groups began work on a solution, an awesomely complex
form of operating system that kept several programs, initiated from
individual interactive terminals, going at the same time. Each user
program would be allowed to run for a fraction of a second, then
the time-sharing operating system would switch control to the next
one and so on, eventually returning to the first. A human tied to

any one of the active programs would not notice the fractional second
 —

Si/mbiosis 81

interruptions, and would appear to each user that he had a computer

it

all to himself, albeit one that was a little slower than the raw machine.

A program that became temporarily inactive waiting for user response

would simplv be skipped by the operating system, with little time
wasted.
Like the assemblers and high-level languages before them, time-
sharing systems caused considerable controversy in the computer
community. With their need to keep the resources for several pro-
grams around at the same time, and to decide what to do next several
times a second, they took a much larger cut of computer resources
than batch operating systems. Time-sharing did offer efficiencies in
return, however. With many diverse users active at one time, the
various resources of a computer system — memory, disks, tape drives,
printers, displays, —
and so on could be kept busier than in single-
program systems. More important, programmers could monitor the
runs of their creations, quickly stopping them when, as often happens
with new programs, they went awry. The most important efficiency
gains, however, were for the customers rather than the machinery.
Instead of waiting hours for a test run of a program on a batch system,
a time-sharing user could watch the progress of a program, stop and
modify it, and try again within minutes. This speed made possible a
highly experimental, and somewhat Pavlovian, style of programming,
characterized by quick punishment and reward cycles. A generation
of proficient computer hackers was born.
The hackers, some spending much of their waking lives at computer
terminals, quickly added to the basic capabilities of their favorite
habitat. Entirely new uses for computers became the norm. Starting in
the mid-1960s, users could communicate among themselves, live from
terminal to terminal, or by electronic mail to be read and answered at
leisure. They could engage in interactive text and video games, play

sophisticated tricks on each other, share insights and programs, and

in general experience a sense of community through the unwitting

medium of their employers' machines. The community was enriched

by the creation of many public artifacts in the form of computer files
community bulletin boards, witty quotes, technical hints, original
writings, amusing programs, and interactive information-retrieval
systems to help sort through all the richness. Some of the more
sophisticated research systems also offered pictures and sounds from
the machine.
82 Mind Children

The hackers' intensive and varied computer use called for efficient
ways to find, start, and stop many different programs, to scan, read,
and modify information files, to interact with other users, and to ask
the computer to do many things automatically In an evolutionary
process, the command languages with which users at their terminals
communicated with time-sharing operating systems were given these
abilities. Designed by experts for experts, with layer upon laver of

unplanned extensions, controlled by terse and powerful but often in-

consistent, undocumented, and hard-to-remember incantations, these
systems often exasperated the less experienced users. To a hacker
the interface with the computer was instinctive, fast, and immensely
powerful — almost anything was possible to one who could construct
the right spells. But to those lacking the monomania required to keep
up with the system's rapid and amorphous evolution, the interface
was opaque, unhelpful, and very error-prone. The hackers got their
comeuppance when dissimilar systems were hooked together by com-
puter networks — even a hacker is a novice in another group's arcana.
A major success of the hacker era is the Unix time-sharing system.
Created in the early 1970s by two young hackers at Bell Laboratories

and extended by others at the University of California at Berkeley,

it became, by the 1980s, the de-facto standard for larger computers.

Unix is now working its way into upscale personal computers.

Menus and Icons

From time to time computer manufacturers tried to incorporate other

hacker innovations into systems for their more staid customers. A

major goal often was to make the language that invoked the many
functions of the operating system as much like plain English as
possible, in the expectation that this would burden on
greatly ease the
users already proficient in English. Such an effect had been observed
in the scientific community when high-level languages that used more
or less standard mathematical notation were introduced. Dissenters
noted that, unlike mathematics, natural English is a poor language
for precise descriptions; communities requiring precision invariably
generate their own specialized conventions, mathematical notation
being the most obvious example.
This time the nay-sayers were right. Experimental English-like
interfaces (with long forgotten acronyms that tended to include the
 Symbiosis 83

words "Plain" or "Simple" and "English") were researched during the

late 1960s and early 1970s but were not very successful. It was (and is)

not yet possible to embody true, general language understanding, with

its underlying requirement of common sense and broad knowledge of
the world, in a program. The actual systems failed to understand
(or positively misunderstood) many offered phrases, so that using
them was often a guessing game. It took skill to phrase requests so
they would be properly interpreted by their complex but incomplete
and largely undocumented language parsers. Learning a simple,
consistent, special command code was easy by comparison.
Infrequent computer users (and specialists also) were better served
—
by the invention of a much simpler device the multiple-choice menu.
In menu-driven systems, the major options are presented in a list,
from which the user chooses one item. This choice may lead to a
second menu outlining subsidiary characteristics, and so on, until
the desired action is fully specified. Standard menu systems are not
without shortcomings. It takes time both for the computer to print
and for the user to read long menus, and the choices may not always
be phrased in the most compatible manner. Menu systems therefore
tend to be slower than specialist languages operated by experienced
users. Also, it is difficult to capture really complex ideas in a game
of twenty questions. Hybrid systems that once in a while prompt for
essay-type answers sometimes manage this problem fairly well.
The speed problem in printing menus was easy, if a little expensive,
to overcome: it required simply the installation of faster-displaying
terminals. The new user's problem of quickly digesting long lists of
unfamiliar options did not admit of so straightforward a solution. An
excellent answer was explored and developed by hackers working in
the exceptionally luxurious quarters of the Xerox Palo Alto Research
Center (PARC). During the early 1970s group developed expensive
this
workstations, each with its own personal computer and large-screen
display, capable of fast and fine graphic imagery. With good graphics
and lots of personal computer power available, rather fancy interac-
tions between user and computer became possible. For example, each
computer had a hand-held device, called a mouse, which could be slid
over a desktop like a hockey puck and which sensed this motion by
means of a small rolling ball protruding from its base. The motion
of the mouse was linked to the motion of a graphic arrow on the
computer's display screen, and the user could thus point out any part
84 Mind Children

of the screen to the computer. Pointing made menus more natural to

use. Instead of typing the Hne label of a menu item, one could simply
point to it to make the selection.
Small suggestive pictures attached to each menu line on the com-
puter greatly eased the visual task of picking out the proper item and
made the system partly language-independent. In later versions of
this idea, the images, now referred to as icotts, became the dominant
representation, with words relegated to a small annotation. Eventually
icons could be moved around the screen like objects and placed into
other icons that functioned as containers or markers for physical
destinations like printers. Icon interfaces proved effective and easy
to use for novices and experts alike, probably because they tapped
the nonverbal object manipulation skills of humans.
Providing each user with a separate computer had many implica-
tions, some of them confusing. Was it not a step backward from time-
sharing? Several of the pioneers of time-sharing certainly thought
so. The enthusiasts at PARC pointed out that, seen as a convenience
for the user, time-sharing was seriously flawed. It immobilized the
habitual user by tying him to a fixed terminal physically connected to
a huge machine. Worse, the responsiveness of time-sharing systems
had never lived up to their early promise; the number of users on the
computer, and the overhead of the time-sharing system itself, invari-
ably crept up until individuals found themselves waiting seconds, and
even minutes, for the machine to respond to the most trivial requests.
A small personal machine to handle the routine functions seemed
plausible, given the rapidly declining cost of computers.
Alan Kay, the guru of the PARC group, originally envisaged a highly
responsive, book-size personal computer (dubbed the Dynabook for

its dynamic qualities) with a high-resolution color display and a

radio link to a worldwide computer network. Much more than
just a computer, the Dynabook would function as secretary, mailbox,
reference library, amusement center, and telephone. This idea was,
and still is, beyond the capabilities of the technology (though I am
typing these very words into a book-sized portable computer as I sit
in a hotel lobby).
The expensive PARC personal computers were desk- rather than
book-size, had only black and white graphics, and offered less com-
puting power than their designers desired. Nevertheless, they were
a clear step toward the vision of the Dynabook. They were initially
 Symbiosis 85

named Interim Dynabooks, but later, mercifully, acquired the more

melodious name of Alto.
Xerox was slow to commercialize the discoveries of the PARC
group, though in the late 1970s they did produce one expensive
business workstation (named Star) that embodied many of the PARC
ideas. After ten years of research, some of the PARC enthusiasts
grew and found
frustrated with the corporate sluggishness of Xerox
a sympathetic ear in Steve Jobs, co-founder of the nearby Apple
computer company. The result, a few years later, was the Apple
Lisa and, later, its smarter younger sibling, the Macintosh, billed as
the personal computer "for the rest of us." Though still far from a
Ovnabook in capability, it introduced a new way of computing to
millions of people —
and paved the way for a new phase of human-
computer interaction. In the late 1980s almost all new operating
systems for graphically oriented computers are being designed to
present a Macintosh-style face to the user.

Magic Glasses

The graphical interface that makes the Macintosh and its imitators
so much more pleasant to use than earlier machines demonstrates
the value of engaging the sensor}^ capabilities of humans in the
dialogue between them and their machines. Alan Kay's Dynabook,
though wonderful in manv ways, would be unable to go much
further than existing svstems in this nonverbal direction because of its

physical limitations, particularlv the book-size viewscreen. As with

a conventional book, portability is a kev feature of the Dynabook;
many of its proposed uses would vanish if it existed onlv in fixed
locations like home or work. Can this portabilitv be retained while at
the same time the owner's sensory involvement expanded? is greatlv
In other words, can we imagine a computer that takes advantage of the
mobihty of humans, while allowing humans to take advantage of the
superior memory, calculating power, and expanded communications
range of the computer?
Of course we can — not in the form of a book, however, but in the
form of a high-tech wardrobe. The key item of apparel is a pair
of magic glasses (or, in the primitive stages, goggles or a helmet).
Worn on the nose like conventional specs, these contain the following
impressive array of instrumentation:
 Mind Children

High-resolution color displays, one for each eye, with optics that
cover the entire field and make the image the computer presents
appear to be focused at a comfortable distance. The glasses may

have the ability to switch to a transparent mode.

Three TV cameras. A high-resolution pair with forward-looking,

wide-angle lenses is placed as close to the position of the eyes as
possible, so that one can see where one is going when the cameras
are connected to the corresponding display screens on the lenses.
Perhaps a third very wide angle camera looks back to register most
of the wearer's face, allowing the computer, and any videophone
communicants, to monitor the wearer's facial expressions.

Microphones and small earphones in the frames.

A navigation system that accurately and continuously tracks the

position and orientation of the glasses (and consequently the head
of the wearer).

A powerful computer that can generate realistic synthetic imagery,

sound, and speech; understand spoken commands; and identify and
track objects in the field of view of the cameras.

A high-speed data link to a worldwide network of computers and

electronic libraries, as well as to magic glasses worn by others.

It does not take an expert to recognize that this is a demanding

set of requirements. Yet every one of these functions exists and is

the subject of extensive, well-funded research. Squeezing it into a

tiny package is also challenging but not implausible in this day of

cassette-box Walkmans and pocket televisions. The glasses let the
computer control what you see and hear, in response to your verbal
requests, and to keep track of your wanderings; they also allow it to
watch your hands moving, and your facial expressions, though they
cannot control what you feel with your hands. Since touching is an
important source of information about the world, it would be nice if
a computer let us know what things feel like, even when they are not
in our immediate environment. Enter magic gloves.

Like the glasses, the gloves are a tour-de-force of technological

wizardry. Each finger of the glove contains a grid of elements that
create patterns of pressure and temperature on the finger of the wearer.
Physiological experiments have shown that realistic impressions can
Magic Glasses (Early Model)
Instrumentation in military aircraft —
where instant access to naviga-
—
and weapons data is a life-or-death matter is evolving
tion, sensor,

into magic glasses. This model was developed in 1986 for a Boeing-
Sikorsky experimental helicopter project. Data from the aircraft's
instruments are projected into the pilot's field of view. Radar blips,
for instance, are made to appear at the actual locations of the objects
being tracked.
 Mind Children

be created by this arrangement; for instance, uniform pressure and

cold is interpreted as "My fingers are immersed in water." The gloves
have motors that on the joints of the fingers, so the wearer can feel
act
resistance to motions. The same mechanisms permit the computer to
monitor the position of the fingers.
Magic gloves, like magic glasses, have their limitations. They can
generate resistance only to finger motions, yet manipulation involves
movement of the whole arm. So imagine a motorized coat that could
supply arm joints the sense of presence that the gloves give to
to the

the hand. Compared with the gloves, and especially the glasses, the
coat is simplicity itself. But in early versions of the magic wardrobe,
the coat may become hardware that
a convenient repository for all the
does not fit into the smaller and more complex glasses and gloves. It
may even have to be attached to an immobile seat: a real "armchair."

Robot Proxy
This robot proxy was developed in 1986 at the Naval Ocean Systems
Center Hawaii. The motions of the operator on the left are copied
in

by the robot on the right, and the images from the robot's camera eyes
are delivered to the operator's (bulky) magic glasses. The operator has
the subjective sensation of being in the robot's body.
 Symbiosis 89

Later the coat will be as portable as a spacesuit, and eventually it will

resemble a leisure suit.

As with existing computers, the wonderful hardware just outlined is

only as useful as its software.The following sections present programs

you might want to run when wearing your smart new outfit.

Finding Oneself

1 don't know about you, but I often get lost. Lacking a good built-
in sense of direction, I long for a pocket navigational aid that would
not only tell me where I am (electronicmaps now appearing in cars
already do this) but would guide me to my destination, remember
where I've been, and remind me of my grocery list when it notices I'm
close to the store. The glasses contain a navigation system, probably
a combination of a device that measures the distance to ground and
orbiting beacons by radio and one that notes acceleration forces to
deduce motion. However it is done, the navigator knows where you,
the wearer of the glasses, are at any time.

Robin and the family have changed summer places since the last time
you visited and now live somewhere in the backwoods. This is unfamiliar
territory, and for once you really appreciate the glowing green line generated
by the Yellow Brick Road program that guides you down the highway. The
line veers into the right-hand lane, heading into a turnoff about a half-mile
ahead. Just in case you didn't notice, a flashing right turn arrow hovers
over the intersection, and the program whispers "Turn approaching, slow
to 30" into your ear. A mile after the turn the asphalt becomes gravel,
presaged by a color change in the guideline from green to a more cautious
(and metaphorically correct) yellow. Later still, the line becomes red, and the
road dirt. The program announces "Switching to private map" indicating
that the Rand McNally database does not yet contain this little trail; the
data comes from Robin's personal files. "Four miles to Robin's," the program
offers, momentarily interrupting the old Beatles tune it's been playing.
It's getting dark. "Initiate phone — Robin," you say to the computer
on your nose. This starts a telephone program that connects via the
cellular/satellite network. The Yellow Brick Road program continues in
parallel. "Tm almost there," you tell Robin. "Great. Dinner's just about
ready. Do you like rice or potatoes?" "Both?" "Ok, see you!" "Bye."
"Bye." The phone program terminates. The red guideline winds among the
90 Mind Children

trees. The headlights are on but don't penetrate very far. Suddenly the car
lurches to the left. You've hit something, perhaps a rock in the road, and
now your front wheel is stuck in a muddy rut. Trying to drive out mires
you more firmly. "Display map," you ask of the navigation program. The
requested map hangs in the air before you. A little picture of your car on
it shows you about two miles from your destination as the road meanders,
though less than a half mile on a straight path through the woods.
"Hi Robin, me again. I'm in trouble." After considering the predicament,

Robin suggests an adventurous dinner-saving course. "We can deal with the
car in the morning. Last week Marty and I found a short-cut to near where
you are. You should be able to walk here in under half an hour." Hardy soul
that you are, you The record of the path is transferred from Robin's
agree.

computer to yours over the phone connection. Your navigation program links
the short-cut route with the road data, and the original guideline is replaced
by one that runs down the road for a distance, then snakes off into the woods.
You lock your car and follow the line.

It is very dark, so you activate the night-vision program — the cameras in

your glasses are run at maximum sensitivity and deliver enhanced images
to the screens, added to the output of the navigator. You see that the red
guideline follows a faint trail through the woods. Except for a few scratches,
the walk is uneventful. The meal is warm and delicious.

Going Places
Transportation and communication have improved awesomely in the
last 500 years. Yet geography, while no longer being the major
determinant of commerce, still restricts how and with whom most

of us conduct our affairs.The differences between transportation and

communication will become less distinct as we become more able to
project our full awareness and skills to remote locations. The ease of
such projection will allow common interest to be the primary spur to
association. The magic wardrobe can be used to visit both real places
in the world and "unreal" locations deep inside computer simulations.
The most obvious form of remote presence involves a physical robot
—
proxy a distant robot that you control via the global communications
network. The magic glasses allow you to see through the robot's
eyes, the coat and gloves permit you to feel, gesture, and act through
the robot's manipulators, and foot controls on your armchair let you
drive the robot around. By renting proxies at remote locations, you
 Symbiosis 91

can and work at widely scattered projects without leaving

visit, talk,

the comforts of home and without incurring the physical risks of

dangerous locations or the boredom of long trips.
A twist on the proxy idea is a human proxy. Someone traveling to an
interesting locale and wearing a special kind of magic wardrobe would
be able to transmit their view, sound, touch, and perhaps odor im-
pressions to the armchair traveler. In its simplest form, the link is one
way, from proxy to passive observer. Such one-way communication
—
can be recorded and played back at will giving us a plausible form
of the "feelie" extension to the "movie." Alternatively, the connection
could be two-way, with motor actions and sense impressions, suitably
edited by a clever program, transmitted bidirectionally. Inexperienced
participants are likely to find themselves as amusingly uncoordinated
as in a three-legged race, but training may permit an intimate kind of
teamwork, with remote expertise being brought to bear at the location

of the problem. Most of the time would be in control

the "field agent"
of the movements, while the armchair participant would watch, listen,
feel, and give advice. But when the task called for a manual skill better

known to the stay-at-home, the field agent would relax and allow the
remotely controlled suit motors to do the job as if possessed by a —
spirit.

A proxv meeting need not be in the real world — many things can
be done better in computer simulation. Computer generated "unreal
estate" has no intrinsic limits either in extent or in physical properties.
It is world where magic is routine. Today's computer screens
a
—
allow peeks into this world car designers examine future models,
physicists view the interior of nuclear explosions, and Macintosh users
rearrange their files on an unreal space called a "desktop." With a
magic wardrobe, we will be able to go boldy into such worlds and
explore them from the inside.

Your (modest) dream home is finally becoming a reality. The preliminaries

with the architect were done weeks ago, and an eagerly anticipated call

arrives: "Hi there. I've got a tentative design. Do you have time to look it

over?" "You bet!" "Ok, lets switch to the site. Initiate Scene M5." After
a few seconds the computer in your glasses prompts "Allow scene change?"
"Allow," you confirm.
A pretty good rendition of the lot and its neighborhood surrounds you. "I
thought we'd put the house over here. That gives a nice front and rear yard,
 92 Mind Children

and room for driveway and garage on the right." An outline of the house
appears on the ground. "Hallway, guest washroom, living and dining room,
kitchen, and stairwell on the first floor." A labeled floor plan shows up in the

outline. You presume the architect's view of the scene is more complicated
than your own and includes display controls. "Lets put up the first floor."

The floor plan sprouts walls. Your point of view becomes higher, and you see
a second floor plan on top of the truncated building. "Two bedrooms, two
baths and an office at this level." The second floor grozvs into place. "The
third floor is attic, with potential for two bedrooms and a bath." The roof

Unreal Estate — The Road to Point Reyes

This scene was synthesized at the Lucasfilm computer graphics division in'

1985 from an underlying three-dimensional computer model. Magic glasses and

enough computing power will allow us to stroll through fantasy worlds like

this. © 1986 Pixar.

 —

Symbiosis 93

finishes the assembly, and a bit of landscapitjg is added. Drifting back to

ground level and around the house, you take in the scene. "Could we try
that in brick?" The walls change from cut stone facing to brick. "The back
yard looks a little small. Could we move things forward?" "We can't get

too far out of line with the neighbors, but I think I can give you 15 feet."

The house slides forward. "That's better. Let's go inside."

The front door swings open. You note a light switch on the right and
reach to flip it. A stick-figure caricature of a hand connects, and the hall

is illuminated. The view from the living-room windows is not inspiring.

"Could we have a bay window here, maybe, instead of this one?" "Hold on a
sec, ni have to set that one up." You drift into the kitchen, which looks pretty
spacious, then examine the dining room. "Wiridow's ready." The living room
looks better with sunlight streaming in the bay window. "What season do
you have set?" "Realtime. Let's cycle through a year." The lighting changes
through morning and afternoons in all four seasons arid the winter scene
is particularly cheery. After a quick tour of the upstairs, with a few wall
color changes, you bid goodbye. "Til leave a copy. You'll be able to make
cosmetic changes and fiddle with the furniture; don't worry about ruining
the design — the program will preveyit you from doing anything silly." "The
family will be thrilled this evening. Talk to you later."

Bare-Hands Programming

Skilled practitioners in many fields report that they see or feel the
object of their work as they think about it. This is not a great
surprise in occupations that concern physical objects or situations
sculpture or sports, say, or mechanical design. It is less expected
in supposedly abstract fields like music, language, mathematics, or
theoretical physics. Yet Einstein, for instance, reported that he could
often feel the meaning of his equations in his arms and his body as if

they were solid objects.

As 1 suggested in Chapter 1, the large, highly evolved sensory
and motor portions of the brain seem to be the hidden powerhouse
behind human thought. By virtue of the great efficiency of these
billion-year-old structures, they may embody one million times the
effective computational power of the conscious part of our minds.
While novice performance can be achieved using conscious thought
alone, master-level expertise draws on the enormous hidden resources
of these old and specialized areas. Sometimes some of that power can
94 Mind Children

be harnessed by finding and developing a useful mapping between

the problem and a sensory intuition.
Although some individuals, through lucky combinations of inher-
itance and opportunity, have developed expert intuitions in certain
fields, most of us are amateurs at most things. What we need to
improve our performance is explicit external metaphors that can tap
our instinctive and repeatable way. Graphs, rules of
skills in a direct

thumb, physical models and other devices

illustrating relationships,

are widely and effectively used to enhance comprehension and re-

tention. More recently, interactive pictorial computer interfaces such
as those used in the Macintosh have greatly accelerated learning in
novices and eased machine use for the experienced. The full sensory
involvement possible with magic glasses may enable us to go much
further in this direction. Finding the best metaphors will be the work
of a generation; for now, we can amuse ourselves by guessing.

The familiar landscape of the top level of your file system lies ahead. In

the foreground, on a grassy green meadozc, are variously sized, colored,

and shaped boulders labeled "Budget," "Drqwings," "Games," and so on.

In the fog-shrouded distance are large hills emblazoned "Oxford English
Dictionary" and "Encyclopedia Britannica." Two knocks on the "Space"

boulder cause it to expand and to open a portal in its side. Wlmt might
be taken for an asteroid belt is visible through the portal. One of the rocks
floating in the blackness is labeled "Skyhooks." You drift up to it, knock
twice, and enter. A pretty blue and white earth, and some less pretty bits of

variously shaped debris, greet you. This is an unfinished project, and some
of your less successful experiments have yet to be laid to rest.
Today's problem is to develop a simulation of a long and strong cable
orbiting the earth. The cable has mass and a certain stretch. It can be
approximated (you've learned) by stringing together large numbers of simple
springs and even simpler weights. A simple spring joins tzvo points and
exerts a force on them proportional to the amount of elongation from a rest
length. A mass has a position and a velocity that changes in proportion to an
applied force in accordance with Nezvton's three laws of motion. The formula

for a spring is F= K (L - 1,,) where L is the spring's current length and L„ is

its unstretched length. K is the spring constant: a larger K makes the spring

harder to stretch. F is the force exerted by the spring on its endpoints. This

relation is among the debris that litters the landscape. You begin by choosing
 Symbiosis 95

some compwnents you've constructed in the past. A point is imaged as a

small black dot that hangs in space (internally it has three numbers giving
its X, Y, and Z coordinates, but that was yesterday's concern). Tapping
on the point and saying "Duplicate!" gives you a second one. These will
represent the two ends of the spring.

You fetch a length arroiv; it looks like a line with an arrozvhead at each
end and a number (its length) in the middle. Fasten its two heads to your
two points and it calculates the distance between them. Taking it for a spin,
you grab one of the points and move it around. The arrow follows the point
and the length number changes obediently. Tapping it you say "Call this
L." The dimension changes to the syynbol L. "Attach spring formula," you
command, and a copy of the formula springs from the landscape, settles
nearby, and begins to respond to the distance between the points. Slots for

the variables K and L^-^ appear, and you give them values. "Vectorize" is

another prepared component; given a pair of points and a simple magnitude,

it gives direction to the quantity, that is, the direction of the line joining the

points. This is attached to both points, in opposite directions. You command

"Vectorize F." The points now exert the spring force, though they remain

fixed. Wlien you reach to grab one or the other, it tugs on your hand — the

farther you go, the harder it pulls back toward the fixed position of the other
point. Only a few kinds of quantity can be directly experienced this way.

Position, color, and temperature are others. In many applications it is helpful

to translate more abstract measurements into palpable ones.

You add mass to the endpoints. This allozcs them to move independently,
under control of momentum and applied forces, such as the spring force.
With its ends released, the spring vibrates. The vibration does not diminish
until you add a damping term to the force equation that diminishes the force,
depending on the rate of change ofL. Now the spring behaves reasonably, and
you sproing it a few times for fun. Invoking the compiler converts the spring
into a single object and greatly improves the efficiency of the underlying
program. You edit the spring's image to make it look like a stretchy coil, with
black disks representing mass on the ends. A dozen duplicates of the spring
strung together end-to-end make a rather stretchy rope. Your simulation is

off to a good start, but it's lunchtime. After lunch you'll make a longer
section, alter the parameters in the various parts, and instrument it, perhaps
by plotting the stretch of the various sections in a graph. Then you'll turn on
the earth models gravity and put the string into orbit around it ami watch
what happens.
96 Mind Children

Elementary Physics

Socrates, whose were recorded for us by his student Plato,

lessons
wrote no books himself. He seemed to think writing was a bad
idea, since it allows its users to put on a show of knowledge by
looking things up, without really knowing anything; the very capacity
to remember, and to think about the memories, was jeopardized.
Furthermore, an argument presented in a book provides no outlet for
disagreement, unlike a person, with whom one can argue or obtain
clarifications. Both objections have merit. Book knowledge is certainly

dry and static compared with active knowledge in a clever person's

mind. The invention of printing greatly aggravated the effect. Yet
books have a reach, capacity, and permanence much greater than any
person's memory, and these properties have made modern civilization
possible.
Before the age of printing, books were expensive, laboriously hand-
made items found in a few widely scattered libraries. Private copies
were hard to obtain, and scholars found it necessary to memorize
whole volumes. Artificial aids to memorization were valued and
sometimes jealously guarded from theological, political, and com-
mercial rivals. A very effective memory technique, developed into
countless variations through the Middle Ages, was The Walk. A large
location, perhaps a cathedral with many rooms, was remembered or
imagined. A book or lecture to be memorized would be recited while
at the same time the structure was mentally traversed. Each room of
the cathedral, or portion of a room, would serve in the mind's eye
as a repository for a section of the text, perhaps marked by some
object that reminded one of the topic. The task was thus broken into
manageable chunks; each location required only a moderate amount of
remembering. The entire piece could be reconstructed by once again
mentally walking through the building, visiting the rooms one by one,
with the mental images so generated bringing to mind the associated
portions of text.

The Walk may be so effective because it maps the new cultural

need to memorize large quantities of speech into the much older
survival skill of remembering where we saw or left various things.
Recalling the location of a food source, shelter, a danger, a friend or
foe, or simply a landmark once seen in the course of a journey has
clear everyday benefits and is something many of us do naturally and
 Symbiosis 97

quite well. At least a portion of our memory is likely organized in an

approximately geographical way to facilitate this kind of recall.

A lesson delivered through an advanced edition of the magic

wardrobe can simultaneously be as responsive as a personal dialogue,
as permanent and available as a printed book, and in resonance with
natural skills in a way that exceeds any existing method.

The "Gravity" portal opens onto a brightly sunlit pastoral scene. A tree-

lined country road winds into distant hills, fluffy clouds dot the sky, birds
are chirping somewhere. A few of the trees bear apples, and from time to

time one falls to the ground. Some distance down the road a bewigged figure

comes into view, sitting under one of the apple trees. You recognize Sir Isaac
Newton. He looks just like he did in the "Laws of Motion" chapter.

"Greetings, young friend," says Sir Isaac. "I've been puzzling over the

nature of the attraction of the earth for various objects. This apple, for
instance, tugs at the hand with a certain force." He hands you the apple;
sure enough, it has weight. "An apple with twice the substance pulls twice
as strongly." The apple gets bigger and heavier. "The great Galileo observed
that, when released, an object falls toward the ground with a constantly
increasing velocity, independent of its weight." Galileo's demonstrations

with falling balls in "Laws of Motion" come to mind. "Yes, yes, get on
with it." Newton, with a slight frown, continues. "We can conclude that
each particle of an object is attracted to the center of the earth with a force
proportional to its mass. Does this attraction change with distance from the
earth? One can conjecture that the influence extends to great distances and
holds the Moon in its monthly circuit. If the same lazes apply to celestial

bodies as to the mundane, then our studies on the motion of objects indicates
that a force in the direction of the earth's center suffices to bend the moon's
path. Yet the required force is almost 4,000-fold weaker, per particle of mass,
than is experienced by the apple you hold."

As he speaks the ground swells at a fantastic rate, and you, Newton, and
the tree are on the summit of a hill rising like a rocket. "Consider the path
of an object thrown horizontally from a great height; your apple, perchance."
Taking the hint, you launch the apple with a smart upperhand throw. (In the
real word, the motors in your jacket and gloves hum momentarily as they
resist your moving arm, simulating the forces of the apple's inertia). The
apple arcs sknvly toward the ground and strikes near the horizon. The hill

has stopped growing, but you are very high, and the spherical shape of the
planet is evident. You can make out several continental outlines. This is
98 Mind Children

obviously a miniature scale model of the earth. Sir Isaac hands you another
apple and recommends a harder throw. It arcs beyond the horizon, in a curve
almost paralleling the ground. You hear a splat through the ground under
your feet. A yet harder launch results in no impact at all, and after a while

the apple whizzes past your head from behind and goes round once again.
A miniature side view of you, hill, earth, and apples makes all this clearer;

each launch traces out an ellipse that returns to its starting point unless it

intersects the ground first. Newton recalls Kepler's laws of planetary motion
and claims they hold for the apples only if the attraction drops as the square

of the distance from the planet's center. You're skeptical, so the two of you
experiment with other rules. Some cause the apples to trace out nonrepetitive
patterns. Those that do give ellipses violate Kepler's second law, that the line

joining the planet center to the orbiting body sweeps out equal areas in equal
times. After a while your throwing arm gets tired, and you say you're
convinced.
But sometimes your skepticism leads to questions that stump your host.

You remember Newton once responding, "A curious puzzle. Let me ponder
it awhile." Several visits later he came puffing after you with the answer,
coattails flying, one hand holding down his wig and trailing a cloud of dust.
(You presume the book's software, unprepared to answer the question the first
time, had issued a message about it to the book's authors. The authors then
created entries in the book's database that allowed your pending query, and
any similar ones Isaac encounters in future, to be answered.)

The hill shrinks back to flatness, and you're on the road again. Next stop is

a pasture where some of the more formal parts of the lesson will be discussed.
A half dozen exotic creatures are already gathered there. Many people the
world over are reading this book, and the world network makes it possible

for those who wish to associate to be mutually aware of each other during
the course of the study. In such associations most people take advantage of
the freedom of the simulation to assume forms different from their physical

bodies, for anonymity and whim. Your group has a Wolf, Floating Eye, Tin
Man, Giant Butterfly, Dragon, and Small Tank. You yourself appear to them
as a rather stylish Dwarf with axe and tasseled hat. Some former classmates
who began this physics book with you are no longer in your cohort because
they sped ahead or fell behind your pace or took a different turn at a subarea
branchpoint. From time to time you pick up new traveling companions as

subcategories re-merge. The whole world is divided into overlapping "villages

of common interest" of this kind. The group sizes range from two to several

thousand. Often, of course, it's good to walk the paths of learning and
 Symbiosis 99

entertainment in solitude. Among other advantages, the action can be better

individually tailored, since there are fewer constraints.

After the lesson you glance farther down the road. In the distance is a
railway platform with a stopped passenger train. Looking carefully, you note
in the window of one of the cars the somewhat disheveled profile of the world's

most famous scientist. But you're tired, and so you disconnect for the day.

Relativity can wait for tomorrow.

 Grandfather Clause

replace humans
w
HAT happens when ever-cheaper machines can
any situation? Indeed, what will I do when a com-
in

puter can write this book, or do my research, better than I? These ques-
tions have already become crucial ones for many people in all kinds
of occupations, and in a few decades they will matter to everybody.
By design, machines are our obedient and able slaves. But intelligent
machines, however benevolent, threaten our existence because they
are alternative inhabitants of our ecological niche. Machines merely as
clever as human beings will have enormous advantages in competitive
situations. Their production and upkeep cost less, so more of them
can be put to work with the resources at hand. They can be optimized
for their jobs and programmed to work tirelessly.

As if these technological developments were not threatening enough,

the very pace of innovation presents an even more serious challenge
to our security. We evolved at a leisurely rate, with millions of
years between significant changes. Machines are making similar
strides in mere decades. When multitudes of economical machines
are put to work as programmers and engineers, presented with the
task of optimizing the software and hardware that makes them what
they are, the pace will quicken. Successive generations of machines
produced this way will become smarter and less costly. There is no
reason to believe that human equivalence represents any sort of upper
bound. When pocket calculators can out-think humans, what will a
big computer be like? We will simply be outclassed.
So why rush headlong into an era of intelligent machines? The
answer, I believe, is that we have very little choice, if our culture
is to remain viable. Societies and economies are surely as subject
to competitive evolutionary pressures as are biological organisms.
Sooner or later the ones that can sustain the most rapid expansion

100
 Grandfather Clause Wl

and compete with one another

diversification will dominate. Cultures
for the resources of the accessible universe. If automation is more

efficient than manual labor, organizations and societies that embrace

it will be wealthier and better able to survive in difficult times and to

expand in favorable ones. If the United States were to unilaterally halt

technological development (an occasionally fashionable idea), it would
soon succumb either to the military might of unfriendly nations or to
the economic success of its trading partners. Either way, the social
ideals that led to the decision would become unimportant on a world
scale.

If, by some unlikely whole human race decided to eschew

pact, the
progress, the long-term result would be almost certain extinction.
The universe is one random event after another. Sooner or later an
unstoppable virus deadly to humans will evolve, or a major asteroid
will collide with the earth, or the sun will expand, or we will be
invaded from the stars, or a black hole will swallow the galaxy. The
bigger, more diverse, and competent a culture is, the better it can
detect and deal with external dangers. The larger events happen less
frequently. By growing rapidly enough, a culture has a finite chance
of surviving forever. In Chapter 6 I will fantasize about schemes that
would allow an entity to restructure itself so as to function indefinitely
even as its universe ended.
The human race will expand into the solar system before long, and
human-occupied space colonies will be part of that expansion. But
only by a massive deployment of machinery can we survive on the
surfaces of other planets or in outer space. The Apollo project, for
example, put people on the moon for a few weeks for $40 billion,
whereas the Viking landers functioned on Mars for years, at a cost of
only $1 billion. If machines as capable as humans had been available
for the Viking project, they would have been able to gather far
more information about Mars than people were able to gather about
the moon, simply because machines can be constructed to function
comfortably and economically in unearthly environments.
Outer space is already a profitable arena for the owners of communi-
As transportation costs decline, other activities will
cations satellites.
start topay Space factories using raw materials purchased from earth
or from human space outposts will be processed by human-supervised
machines and sold at a profit. The high cost of maintaining humans in
space ensures that there always will be more machinery per person in
102 Mind Children

a space colony than there is on earth. As machines become more capa-

ble, the economics will favor an ever higher machine-to-people ratio.

Humans will not necessarily become fewer at this stage; the machines
will just multiply faster, becoming ever more competent with each new
generation. Imagine the immensely lucrative robot factories that could
be built in the asteroids. Solar-powered machines would prospect
and deliver raw materials to huge, unenclosed, automatic processing
plants. Metals, semiconductors, and plastics produced there would be
converted by robots into components that would be assembled into
other robots and into structural parts for more plants. Machines would
be recycled as they broke. If their reproduction rate is higher than the
wear-out rate, the factories will grow exponentiallv, like a colony of
bacteria, on a Brobdingnagian scale. A harvest of a small fraction of
the output of materials, components, and whole robots could make
investors incredibly rich.
Eventually humans, whether workers, design engineers, managers,
or investors, willbecome unnecessary in space enterprises, as the
and technical discoveries of self-reproducing superintelligent
scientific

mechanisms are applied to making themselves smarter still. These

new creations, looking quite unlike the machines we know, will
explode into the universe, leaving us behind in a cloud of dust.

Robot Bushes

The human world has been shaped by human hands, which are still
our most effective general-purpose tool. Yet many useful and easily
described tasks are beyond human dexterity (pull tightly on both ends
of the string, while holding the knot between your fingers, lift the
bundle, wrap the ends around it tightly four times...). If such actions
are attempted at all, it is with varying degrees of success using special
tools and fixtures.

It is unlikely that our superintelligent descendants will be satisfied

with mere stumpy fingers. Consider the following observations.
Worms and other animals shaped like balls or sticks are unable
to manipulate or even locomote very well. Animals with legs (a
stick with smaller, movable sticks) locomote quite well but are still

clumsy at manipulation. Animals like us, with fingers on their legs

(sticks on sticks on a stick), can manipulate much better. Now
generalize the concept —a robot that looks like a tree, with a big
 Grandfather Clause 103

stem repeatedly branching into thinner, and more numerous

shorter,
twigs, ultimately ending in an astronomical number of microscopic
cilia. Each intermediate branch would be able to swing forward
and backward and side to side while its top, where the next smaller
branches are attached, rotates on the branch axis. Possibly the branch
—
could also change its length like a telescope the number of motions

A Robot Bush
104 Mind Children

of each branch can be traded off for more levels. Each joint would have
sensors to measure its position and also the force it exerts. Although
made of branches, each with a rigid mechanical character, the overall
structure would have an "organic" flexibility because of the great
multitude of ways its parts could move.
A robot of this design could be self-constructing. Tiny bushes, only
a few millionths of the weight of the final device, would be "seeded"
to start the process. These would work in groups to build the next
larger sprigs from available raw materials, then join themselves to
their constructions. The resulting larger bushes would join to build
even larger branches, and so on until a small crew (of large members)
met to assemble the stem. At the other end of the scale, a sufficiently
large bush should be able to organize the necessary resources to build
the tiny seeds to start the process all over again (or simply to repair or
extend itself). It could make the smallest parts with methods similar
to themicromachining techniques of current integrated circuitry. If its
smallest branchlets were a few atoms in scale (with lengths measured
in nanometers), a robot bush could grab individual atoms of raw
material and assemble them one by one into new parts, in a variation
of the nanotechnology methods mentioned in Chapter 2.
To make things more concrete, we can do an actual design. Suppose
that the basic structure is a large branch that splits into four smaller
ones, each half the scale. If we start with a stem a meter long and
ten centimeters in diameter and carry the branching to twenty levels,

the bush will end in a trillion tiny "leaves," each a millionth of a

meter (a micron) long and a tenth of that in diameter. Because of their
much smaller weight and size, the leaves can move a million times as
fast as the trunk. Let's say that the trunk can wiggle back and forth
once per second; so the leaves will vibrate a million times per second.
If the bush folds itself into a tight bundle, its cross section will be
approximately constant, and it will be two meters long. The trunk
will then occupy half that length, the second level half the remaining
length, and so on. Unfolded, umbrellalike, it would spread into a disk
a little under two meters in diameter, thick but sparse near the center,
and thinner at the edge, with smaller gaps that taper off to micron
spaces.
If each joint can measure the forces and motions appHed to it, we

have a remarkable sensor. There are a trillion leaf fingers, each able
to sense a movement of perhaps a tenth of a micron and a force of
 Grandfather Clause 105

a few micrograms, at speeds up to a million changes per second.

This is vastly greater than the sensing ability of the human eye,
which has a million distinguishable points that can register changes
at most a hundred times per second. If our bush puts its fingers

on a photograph, it will "see" the image in immense detail simply

by feeling the height variations of the developed silver on the paper.

It could watch a movie by walking its fingers along the film as it

screamed by at high speed. There is no reason the fingers could not

also be sensitive to light and temperature and other electromagnetic
effects; indeed, the smallest are the right size to be "antennas" for
light. The bush could form an eye by holding up a lens and putting a
few million of its fingers in the focal plane behind it. It may even be
able to get by without the lens by holding a bunch of its fingers in a
carefully spaced diffraction pattern, thus forming a holographic lens.
In addition to having a sensing capability to match that of the
world's current human population, our bush would have the ability

to affect its environmentsame prodigious rate. A well-trained

at the

human, using precise and well-timed hand and body motions, each
able to change direction at most a few times in a second, with a
precision no better than a few percent of the total movement, could
conceivably affect the world at a net rate of a thousand bits per
second —a fast typist, for instance, produces less than one hundred
bits per second of text. The potential data rate of a robot with one
trillion fingers, each able to move a million times per second, is more
than a quadrillion times greater. Such high data rates imply
(10^^)

huge coordination of enormous processing power, but imagine the

possibilities. The bush robot could reach into a complicated piece of

delicate mechanical equipment —

or even a living organism simulta- —
neously sense the relative position of millions of parts, some possibly
as small as molecules, and rearrange them for a near-instantaneous
repair. In most cases the superior touch sense would totally substitute
for vision, and the extreme dexterity would eliminate the need for
special tools.
An astronomical amount of thinking would be required to control
this wonderful machine. Much of it might be handled by what in
animals are called reflex arcs, small bits of nervous system near the site
being controlled. Each of the small branchlets could contain enough
of a computer to control most of the routine activity; only exceptional
situations would require intervention from larger computers nearer
106 Mind Children

the stem. If the branches also contained their own power source (think
rechargeable battery) and a way
communicating remotely (radio, or
of
sound vibrations of a few thousand synchronized cilia, would do), the
bush could break into a coordinated swarm of smaller bushes. The
smaller the individual bush, the less intelligent and less powerful it
would be. It would be preprogrammed, and charged up, by its home
stem to perform some function and then return as soon as possible to
report and receive new instructions.
Small size would frequently be an advantage: a smaller robot can
squeeze into smaller spaces. A tiny machine has a greater surface-
area-to-weight ratio: while a large bush could walk securely along the
floor, using its branches as so many nimble toes, a smaller machine

should be able to walk on ceilings like a fly, with the tiny cilia
holding onto microscopic cracks in the paint or sticking by molecular
adhesion. Bushes could burrow by loosening particles of dirt and
passing them backward, and swim efficiently by assuming a tight,
streamlined shape, with the cilia forming a skin that pumps fluid to
propel and also responds to the flow to prevent turbulence. Extremely
small machines will have so much surface area for their weight that
they can fly like insects, beating their cilia in patterns optimal for
moving air.
The contribution of the bigger branches to the power and intel-

ligence of the smaller ones can be visualized as a kind of reverse

pyramid scheme. Each level (counting all the twigs at that level) of
our robot has twice the volume of the next smaller level and thus room
for twice the power supply and twice the computer. Two levels down,
the ratio is four to one, and at three levels it is eight. If the control
and power for each level are piped up from the branches three steps
closer to the stem, the small branches can be four times as vigorous
as otherwise. Only the stem and the first few branches radiating from
it would be shortchanged. Since most of their power and attention is
directed higher up in the tree, they may be incapable of much motion
and may be relegated to providing a stable framework if they passively
lock their joints.
A big question is how programs for such
the control a beast would
work. In the extreme case one could imagine a program that would try
todetermine the combination of actions of each individual joint that
would best accomplish the desired task. This is almost certainly an
example of a colossal NP (nondeterministic polynomial) problem that
 —

Grandfather Clause 107

can be optimally solved only by essentially examining every possible

combination of motions and picking the best (see Appendix 3). Such
solutions are intractable these days even for simple manipulators that
have only a handful (!) of fingers. Though computers will be vastly
more powerful in the future, the problem posed by a system with a
large number of fingers is much, much bigger still. From time to time
an especially clever strategy for coordinating thousands or millions
of fingers to accomplish a particular task may be discovered, and
collections of clever strategies will be passed between individuals
the manual skills of the superintelligent era. But while humans teach
skills as simple as how to tie a shoelace, the lessons of superintelhgent
machines may be more comparable to instructions for assembling an
airliner Finding the best possible solution for routine problems will
usually be out of the question, but finding a good enough one may
not be too hard. I imagine a divide-and-conquer strategy, where the
stem considers the overall problem and generates plausible subtasks
for each of the four subtrees immediately connected to itself. These
further subdivide the problem and pass those fragments on, and so
on. The smallest branches would receive simple commands like move
to a certain position, or move until resistance is met. A command from
the stem might be something like: North bush — stay on left side of plane

A, and right side of plane B, and apply net force vector V to object; East
—
bush stay on right of A and B, and resist any motion more than 10 cm
from the axis; South bush — right of A, left of B, apply force negative V;
West bush — left of A and B, and resist. If a subproblem, as passed to a
small bush, cannot be solved, a complaint would be sent back to the
originating branch, which would then go back to the drawing board
to try something else.

A bush robot would be a marvel of surrealism to behold. Despite

its structural resemblance to many living things, it would be unlike
anything yet seen on earth. Its great intelligence, superb coordination,
astronomical speed, and enormous sensitivity to its environment
would enable it to constantly do something surprising, at the same
time maintaining a perpetual gracefulness. Two-legged animals have
three or four effective gaits; four-legged animals have a few more.
Two-handed humans have two or three ways to hold an object. A
trillion-limbed device, with a brain to match, is an entirely different
order of being. Add to this the ability to fragment into a cloud of
coordinated tiny fliers, and the laws of physics will seem to melt in
108 Mind Children

the face of intention and will. As with no magician that ever was,
impossible things will simply happen around a robot bush. Imagine
inhabiting such a body.

Transmigration

Some of us humans have quite egocentric world views. We anticipate

the discovery, within our lifetimes, of methods to extend human Hfe,
and we look forward few eons of exploring the universe. The
to a
thought of being grandly upstaged in this by our artificial progeny
is disappointing. Long hfe loses much of its point if we are fated to

spend it staring stupidly at our ultra-intelligent machines as they try

to describe their ever more spectacular discoveries in baby-talk that
we can understand. We want to become full, unfettered players in
this new superintelligent game. What are the possibihties for doing

that?
Genetic engineering may seem an easy option. Successive gen-
erations of human beings could be designed by mathematics, com-
puter simulations, and experimentation, like airplanes, computers,
and robots are now. They could have better brains and improved
metabolisms that would allow them to live comfortably in space. But,
presumably, they would still be made of protein, and their brains
would be made of neurons. Away from earth, protein is not an
ideal material. It is stable only in a narrow temperature and pressure
range, is very sensitive to radiation, and rules out many construction
techniques and components. And it is unlikely that neurons, which
can now switch less than a thousand times per second, will ever be
boosted to the billions-per-second speed of even today's computer
components. Before long, conventional technologies, miniaturized
down to the atomic scale, and biotechnology, its molecular inter-

actions understood in detailed mechanical terms, will have merged

into a seamless array of techniques encompassing all materials, sizes,
and complexities. Robots will then be made of a mix of fabulous
substances, including, where appropriate, living biological materials.
At that time a genetically engineered superhuman would be just a

second-rate kind of robot, designed under the handicap that its con-
by DNA-guided protein synthesis. Only in the
struction can only be
eyes of human chauvinists would it have an advantage because it —
retains more of the original human limitations than other robots.
 Grandfather Clause 109

Robots, first or second rate, leave our question unanswered. Is

there any chance that we — you and I, personally — can fully share

in the magical world to come? This would call for a process that
endows an individual with all the advantages of the machines, without
loss of personal identity. Many people today are alive because of a
growing arsenal of artificial organs and other body parts. In time,
especially as robotic techniques improve, such replacement parts will
be better than any originals. So what about replacing everything,
that is, transplanting a human brain into a specially designed robot
body? Unfortunately, while this solution might overcome most of our
physical limitations, it would leave untouched our biggest handicap,
the limited and fixed intelligence of the human brain. This transplant
scenario gets our brain out of our body. Is there a way to get our
mind out of our brain?

You've just been wheeled into the operating room. A robot brain surgeon
is in attendance. By your side is a computer waiting to become a human
equivalent, lacking only a program to run. Your skull, but not your brain,
is anesthetized. You are fully conscious. The robot surgeon opens your
brain case and places a hand on the brain's surface. This unusual hand
bristles with microscopic machinery, and a cable connects it to the mobile
computer at your side. Instruments in the hand scan the first few millimeters
of brain surface. High-resolution magnetic resonance measurements build
a three-dimensional chemical map, while arrays of magnetic and electric

antennas collect signals that are rapidly unraveled to reveal, moment to

moment, the pulses flashing among the neurons. These measurements, added
to a comprehensive understanding of human neural architecture, allow the
surgeon to write a program that models the behavior of the uppermost layer
of the scanned brain tissue. This program is installed in a small portion of
the waiting computer and activated. Measurements from the hand provide
it with copies of the inputs that the original tissue is receiving. You and
the surgeon check the accuracy of the simulation by comparing the signals it

produces with the corresponding original ones. They flash by very fast, but
any discrepancies are highlighted on a display screen. The surgeon fine-tunes
the simulation until the correspondence is nearly perfect.

To further assure you of the simulation's correctness, you are given a

pushbutton that allows you to momentarily "test drive" the simulation, to

compare it with the functioning of the original tissue. WJien you press
it, arrays of electrodes in the surgeon's hand are activated. By precise
no Mind Children

injections of current and electromagnetic pulses, the electrodes can override

the normal signaling activity of nearby neurons. They are programmed to

inject the output of the simulation into those places where the simulated
tissue signals other sites. As long as you press the button, a small part

of your nervous system is being replaced by a computer simulation of

itself. You press the button, release it, and press it again. You should
experience no difference. As soon as you are satisfied, the simulation

connection is established permanently. The brain tissue is now impotent —

it receives inputs and reacts as before but its output is ignored. Microscopic
manipulators on the hand's surface excise the cells in this superfluous tissue
and pass them to an aspirator, where they are drawn away.
The surgeon's hand sinks a fraction of a millimeter deeper into your
brain, instantly compensating its measurements and signals for the changed

position. The process is repeated for the next layer, and soon a second

simulation resides in the computer, communicating with the first and with
the remaining original brain tissue. Layer after layer the brain is simulated,
then excavated. Eventually your skull is empty, and the surgeon's hand
rests deep in your brainstem. Though you have not lost consciousness, or
even your train of thought, your mind has been removed from the brain and
transferred to a machine. In a final, disorienting step the surgeon lifts out
his hand. Your suddenly abandoned body goes into spasms and dies. For a
moment you experience only quiet and dark. Then, once again, you can open
your eyes. Your perspective has shifted. The computer simulation has been
disconnected from the cable leading to the surgeon's hand and reconnected
to a shiny new body of the style, color, and material of your choice. Your
metamorphosis is complete.

For the squeamish, there are other ways to work the transfer of
human mind to machine. A high-resolution brain scan could, in one
fell swoop and without surgery, make a new you "While-U-Wait."
If even the last technique is too invasive for you, imagine a more
psychological approach. A kind of portable computer (perhaps worn
like magic glasses) is programmed with the universals of human
mentality, your genetic makeup, and whatever details of your life are
conveniently available. It carries a program that makes it an excellent
mimic. You carry this computer with you through the prime of your
life; it diligently listens and watches; perhaps it monitors your brain

and learns to anticipate your every move and response. Soon it can
fool your friends on the phone with its convincing imitation of you.
 Grandfather Clause 111

When you die, this program is installed in a mechanical body that then
smoothly and seamlessly takes over your life and responsibilities.
If you happen to be a vertebrate, there is another option that

combines the methods described above. The

sales features of all the
vertebrate brain has two hemispheres connected by several large
bundles of nerve fibers. The largest is called the corpus callosum. In
the 1960s, guided by animal experiments, researchers in California
successfully treated patients w^ith intractable types of epilepsy by
severing their corpora callosa. (Medical robots in the future will not
use Latin!) Amazingly, this procedure appeared at first to have no side

effects on the The corpus callosum, with 200 million nerve

patients.
fibers, is the brain's most massive long-distance interconnect. It is far

thicker than the optic nerve or the spinal cord. Cut the optic nerve and
the victim is utterly blind; sever the spinal cord and the body becomes
limp and numb. But slice the huge cable between the hemispheres
and nothing bad happens. Well, almost nothing. If the name of an
object, like "brush," is flashed in the right half of the visual field of
view of a "split-brain" person, the person is unable to select the object
from among others with the lefthand but has no difficulty making
the choice with the right hand. Sometimes in the left-handed version
of the task, the right hand — apparently in exasperation— reaches over
to guide the left to the proper location!
Neuroanatomy suggests some of the explanation. The nerv^es direct-
ing the muscles of the left side of the body, as well as those portions of
the optic nerve viewing the left half of the visual scene, are connected
only to the right side of the brain. Conversely, the left side of the
brain controls the right side of the body and sees the right half of the
scene, as illustrated in the figure on page 113. Normally the two brain
halves work in an intimate partnership, and information discovered by
one is rapidly available to the other through the agency of the corpus
is broken. The
callosum. In a split-brain person, this information flow
two brain halves must discover things independently. The left hand
knows not what the right is doing. The two halves still seem to be
aware of each other's emotions, however, from information apparently
relayed through intact connecting nerves in the brainstem.
Roger Sperry of the California Institute of Technology, who received
a Nobel Prize in 1981 for his discoveries on the function of the corpus
callosum, found that in split-brain subjects each brain half seems to
host an independent, fully conscious, intelligent human personality. In
112 Mind Children

intact brains some of the corpus callosum fibers are known to handle
basic functions such as recombining the halves of the visual fields
of the eyes, but others must communicate higher mental concepts
between the hemispheres. There is every reason to believe the corpus
callosum provides a neatly organized and very wide window into the
mental activities of both hemispheres. Suppose in the future, when the
function of the brain is sufficiently understood, your corpus callosum
is severed and cables leading to an external computer are connected
to the severed ends. The computer is programmed at first to pass the
traffic between the two hemispheres and to eavesdrop on it. From
what it learns by eavesdropping, it constructs a model of your mental
activities. After a while it begins to insert its own messages into
the flow, gradually insinuating itself into your thinking, endowing
you with new knowledge and new skills. In time, as your original
brain faded away with age, the computer would smoothly assume
the lost functions. Ultimately your brain would die, and your mind
would find itself entirely in the computer. Perhaps, with advances
in high-resolution scanning, it will be possible to achieve this effect
without messy surgery: you might simply wear some kind of helmet
or headband that monitored and altered the interhemispheric traffic

with carefully controlled electromagnetic fields.

Many Changes
Whatever style of mind transfer you choose, as the process is com-
pleted many your old limitations melt away. Your computer has
of
a control labeled "speed." It had been set at "slow," to keep the
simulations synchronized with the old brain, but now you change it
to "fast," allowing you to communicate, react, and think a thousand
times The entire program can be copied into similar machines,
faster.

resulting intwo or more thinking, feeling versions of you. You may

choose to move your mind from one computer to another that is more
technically advanced or better suited to a new environment. The
program can also be copied to a future equivalent of magnetic tape.
Then, if the machine you inhabit is fatally clobbered, the tape can
be read into a blank computer, resulting in another you minus your
experiences since the copy. With enough widely dispersed copies,
your permanent death would be highly unlikely.
The Corpus Callosum
The left half of the brain controls the right side of the body, and vice-
versa. The left half usually specializes in language and calculation,
while the right half is good at spatial reasoning. The brain halves
normally communicate through the corpus callosum, but they can
continue to function as separate individuals if it is severed.
114 Mind Children

As a computer program, your mind can travel over information

channels, for instance encoded as a laser message beamed between
planets. If you found life on a neutron star and wished to make a

field trip, you might devise a way to build a robot there of neutron
stuff, then transmit your mind to it. Since nuclear reactions are
about a million times quicker than chemical ones, the neutron-you
might be able to think a million times faster. You would explore,
acquire new experiences and memories, and then beam your mind
back home. Your original body could be kept dormant during the
trip and reactivated with the new memories when the return message

arrived —
perhaps a minute later but with a subjective year's worth
of experiences. Alternatively, the original could be kept active. Then
there would be two separate versions of you, with different memories
for the trip interval.
Your new abilities will dictate changes in your personality. Many of
the changes will result from your own deliberate tinkerings with your
own program. Having turned up your speed control a thousandfold,
you notice that you now have hours (subjectively speaking) to respond
to situations that previously required instant reactions. You have time,
during the fall of a dropped object, to research the advantages and
disadvantages of trying to catch it, perhaps to solve its differential

equations of motion. You will have time to read and ponder an entire
on-line etiquette book when you find yourself in an awkward social

situation. Faced with a broken machine, you will have time, before
touching it, to learn its theory of operation and to consider, in detail,

the various things that may be wrong with it. In general, you will

have time to undertake what would today count as major research

efforts to solve trivial everyday problems.
You will have the time, but will you have the patience? Or will a

thousandfold mental speedup simply incapacitate you with boredom?

Boredom is a mental mechanism that keeps you from wasting your
time in profitless activity, but if it acts too soon or too aggressively
it limits your attention span, and thus your intelligence. One of
the changes you will want to make in your own program
first is

to retard the onset of boredom beyond the range found today in

even the most extreme intellectuals. Having done that, you will find

yourself comfortably working on long problems with sidetracks upon

sidetracks. In fact, your thoughts will routinely become so involved
that they will call for an increase in your short-term memory. Your
 Grandfather Clause 115

long-term memon^ also will have to be boosted, since a month's worth

of events will occupy a subjective span of a century! These are but
the first of many changes.
have alreadv mentioned the possibility of making copies of oneself,
I

with each copy undergoing its own adventures. It should be possible

to merge memories from disparate copies into a single one. To avoid
confusion, memories of events would indicate in which body they
happened, just as our memories today often have a context that es-

tablishes a time and place for the remembered event. Merging should
be possible not only between two versions of the same individual but
also between different persons. Selective mergings, involving some
of another person's memories and not others, would be a superior
form of communication, in which recollections, skills, attitudes, and
personalities can be rapidly and effectively shared. Your new body
will be able to carrv more memories than your original biological one,
but the accelerated information explosion will ensure the impossibility
of lugging around all of civilization's knowledge. You will have to
pick and choose what your mind contains at any one time. There
will often be knowledge and skills available from others superior
to your own, and the incentive to substitute those talents for yours
will be overwhelming. In the long run you will remember mostly
other people's experiences, while memories you originated will be
incorporated into other minds. Concepts of and identity
life, death,
will lose their present meaning and those of
as your mental fragments
others are combined, shuffled, and recombined into temporary asso-
ciations, sometimes large, sometimes small, sometimes long isolated
and highly individual, at other times ephemeral, mere ripples on the
rapids of civilization's torrent of knowledge. There are foretastes of
this kind of fluidity around us. Culturally, individual humans acquire

new skills and attitudes from others throughout life. Geneticallv, in

sexual populations each individual organism is a temporary bundling
of genes that are combined and recombined in different arrangements
every generation.
Mind transferral need not be limited to human beings. Earth has
whose nervous systems
other species with large brains, from dolphins,
are as large and complex as our own, to elephants, other whales, and
perhaps giant squid, whose brains may range up to twentv times
as big as ours. Just what kind of minds and cultures these animals
possess is still a matter of controversy, but their evolutionary history
116 Mind Children

is as long as ours, and there is surely much unique and hard-won

information encoded genetically in their brain structures and their
memories. The brain-to-computer transferral methods that work for
humans should work as well for these large-brained animals, allowing
their thoughts, skills,and motivations to be woven into our cultural
tapestry. Slightly differentmethods, that focus more on genetics and
physical makeup than on mental life, should allow the information
contained in other living things with small or no nervous systems
to be popped into the data banks. The simplest organisms might
contribute little more than the information in their DNA. In this way
our future selves will be able to benefit from and build on what the
earth's biosphere has learned during its multibillion-year history. And
this knowledge may be more secure if it is preserved in databanks
spreading through the universe. In the present scheme of things, on
our small and fragile earth, genes and ideas are often lost when the
conditions that gave rise to them change.
Our speculation ends in a supercivilization, the synthesis of all solar-

system life, constantly improving and extending itself, spreading out-

ward from the sun, converting nonlife into mind. Just possibly there
are other such bubbles expanding from elsewhere. What happens if

we meet A negotiated merger is a possibility, requiring only a

one?
translationscheme between the memory representations. This process,
possibly occurring now elsewhere, might convert the entire universe
into an extended thinking entity, a prelude to even greater things.

What Am I?
The idea that a human mind can be transferred to a new body
sometimes meets the following strong objection from people who do
not dispute the theoretical possibility: "Regardless of how the copying
is done, the end result will be a new person. If it is Iwho am being
copied, the copy, though it may think of itself as me, is simply a self-

deluded impostor. If the copying process destroys the original, then I

have been killed. That the copy may then have a great time exploring
the universe using my name and my skills is no comfort to my mortal
remains."
This point of view, which 1 will call the bodi/-identiti/ position, makes
life extension by duplication considerably less personally interesting.
I believe the objection can be overcome by acceptance of an alternative
 Grandfather Clause 117

position which I will call pattern-identity. Body-identity assumes that

a person is defined by the stuff of which a human body is made. Only

by maintaining continuity of body stuff can we preserve an individual
person. Pattern-identity, conversely, defines the essence of a person,
say myself, as the pattern and the process going on in my head and
body, not the machinery supporting that process. If the process is

preserved, 1 am preserved. The rest is mere jelly.

The body-identity position, 1 think, is based on a mistaken intuition

about the nature of living things. In a subtle way, the preservation
of pattern and loss of substance is a normal part of everyday life. As
we humans eat and excrete, old cells within our bodies die, break
up, and are expelled and replaced by copies made of fresh materials.
Most of our body is renewed this way every few years. The few body
components such as nerve cells that tend to be more static nevertheless
have metabolisms that cause their inner parts to be replaced, bit by bit.
Every atom present within us at birth is likely to have been replaced
half way through our life. Only our pattern, and only some of it at
that, stays with us until our death.
Let me explore some of the consequences of the pattern-identity
position. Matter transmitters have appeared often in the science
fiction literature, at least since the invention of facsimile machines in
the late 1800s. I raise the idea here only as a thought experiment,
to simplify some of the issues in my mind-transfer proposal. A
facsimile transmitter scans a photograph line by line with a light-
sensitive photocell and produces an electric current that varies with
the brightness of the scanned point in the picture. The varying electric
current is transmitted over wires to a remote location, where it controls
the brightness of a light bulb in a facsimile receiver. The receiver
scans the bulb over photosensitive paper in the same pattern as the
transmitter. When this paper is developed, a duphcate of the original
photograph is obtained. This device was a boon to newspapers,
who were able to get illustrations from remote parts of the country
almost instantly, instead of waiting for a train to deHver photographic
plates.
If pictures, why not solid objects? A matter transmitter might scan
an object and identify atoms or molecules one at a time, perhaps
its

removing them The identity of the atoms would be

in the process.
transmitted to a receiver, where a duplicate of the original object
would be assembled in the same order from a local supply of atoms.
118 Mind Children

The technical problems are mind-boggling, but the principle is simple

to grasp, as millions of devotees of "Star Trek" will attest.

If solid objects, why not a person? Just stick him in the transmitter,

turn on the scan, and greet him when he walks from the receiver. But
is he really the same person?If the system works well, the duplicate

will be indistinguishable from the original in any substantial way. Yet,

suppose that you fail to turn on the receiver during the transmission
process. The transmitter will scan and disassemble the victim and
send an unheard message to the inoperative receiver. The original
person will be dead. Doesn't the process, in fact, kill the original
person whether or not there is an active receiver? Isn't the duplicate
just that, merely a copy? Or suppose that two receivers respond to the

message from one transmitter. Which, if either, of the two duplicates

is the real original?
The body-identity position on this question is clear. A matter trans-
mitter is an elaborate execution device that kills you and substitutes a
clever impostor in your place. The pattern-identity position offers
a different perspective. Suppose that step into the transmission I

chamber. The transmitter scans and disassembles my jellylike body,

but my pattern (me!) moves continuously from the dissolving jelly,
through the transmitting beam, and ends up in other jelly at the
destination. At no instant was the pattern (I) ever destroyed. But what
about the question of duplicates? Suppose that the matter transmitter
is connected to two receivers instead of one. After the transfer there

will be a copy of me in each one. Surely at least one of them is a mere

copy: they cannot both be me, right? Wrong!
Rooted in all our past experience is the assumption that one person
corresponds to one body. In light of the possibility of separating mind
from matter and storing and transmitting it, this simple, natural, and
obvious identification becomes confusing and misleading. Consider
the message, "I am not jelly." As type it, it goes from my brain into
I

the keyboard of my computer, through myriads of electronic circuits,

and over great amounts of wire. After countless adventures, the
message shows up in bunches of books like the one you are holding.
How many messages were there? I claim that it is most useful to think
there is only one, despite its massive replication. If I repeat it here: "I

am not jelly," there is still only one message. Only if I change it in a

significant manner ("I am not peanut butter") do we have a second

message. And the message is not destroyed until the last written
 Grandfather (
^^

version is lost and until it fades sufficiently in everybody's m

to be unreconstructable. The message is the information conveye , ..v^i

the medium on which it is encoded. The "pattern" that I claim is the

realme has the same properties as this message. Making a momentary
copy of my state, whether on tape or in another functional body, does
not make two persons.
The "process" aspect is a little more complicated. At the instant that
a "person message" is assembled, it is just another copy of the original.
But if two copies are active, they will in time diverge and become two
different people. Just how far this differentiation must proceed before
society grants them unique identities is about as problematical as the
questions "When does a fetus become a person?" or "When does an
evolving species become a different species?" But if we wait zero time,
then both copies are the same person — if we immediately destroy one,
the person still exists in the other copy. All the deeds that that person
might have done, and all the thoughts she might have thought, are
still possible. If, instead, we allow both copies to live their separate
lives for a year and then destroy one, we are the murderers of a unique
human being. But if we wait just a short time before destroying one
copy, then only a little unique information is lost.

This rationale might be a comfort if you were about to encounter

danger but knew that a tape copy of you had been made recently.
Should you die, an active copy made from the tape could resume
your life. This copy would differ slightly from the version of you that
died, in that it lacked the memories since the time of copy. But a
small patch of amnesia is a trivial affair compared with the total loss
of memory and function that results from death in the absence of a
copy.
and recent backup of you exists
Intellectual acceptance that a secure
would not you from an extreme desire to preserve
necessarily protect
yourself if faced with imminent death, even in a worthwhile cause.
Such feelings would be an evolutionary hangover from your one-
copy past, no more in tune with reality than fear of flying is an
appropriate response to present airline accident rates. Old instincts are
not automatically erased when the rules of life are suddenly rewritten.
The pattern-identity position has clear dualistic implications — it

allows the mind to be separated from the body. Though mind is

entirely theconsequence of interacting matter, the ability to copy it

from one storage medium to another would give it an independence
120 Mind Children

and an identity apart from the machinery that runs the program.
The duahsm will be especially apparent if we consider some of the
different possibilities for encoding.
Some supercomputers have myriads of individual computers inter-

connected by a network that allows free flow of information among

them. An operating system for this arrangement might allow indi-
vidual processes to migrate from one processor to another in mid-
computation, in a kind of juggling act that permits more processes
than there are processors. If a human mind is installed in a future
machine of this variety, functions originally performed by particular
parts of the brain might be encoded in particular processes. The
juggling action would ensure that operations occurring in fixed areas
in the original brain would move rapidly from place to place within
the machine. If the computer is running other programs besides the
mind simulation, then the simulation might find itself shuffled into
entirely different sets of processors from moment to moment. The
thinking process would be uninterrupted, even as its location and
physical machinery changed continuously, because the pattern would
be maintained.
The most direct way of performing complex operations such as
occur in a human mind are often not the most efficient. A process that
is described as a long sequence of steps in a program can sometimes
be transformed mathematically into one that arrives at the same
conclusion in far fewer operations. As a young boy, the famous
mathematician Carl Friedrich Gauss was a school smart aleck. For
a diversion, a teacher once set him the problem of adding up the
all the numbers between 1 and 100. He returned with the correct
answer in less than a minute. He had noticed that the hundred
numbers could be grouped into fifty pairs, 1-t-lOO, 2+99, 3+98, 4+97,
and so on, each pair adding up to 101. Fifty times 101 is 5,050, the
answer, found without a lot of tedious addition. Similar speedups
are possible in complex computer processes. So called optimizing
compilers have repertoires of accelerating transformations, some quite
radical, to streamline programs they translate. The key may be a total
reorganization in the order of the computation and the representation
of the data.
One powerful class of transformations takes an array of values
and combines them in systematic ways to produce another array, so

that each element of the new array is a number formed by a unique

 —

Grandfather Clause 121

combination of all the elements of the old array. An operation on a

single element in the new array can then often substitute for a whole
host of operations on the original array, making enormous efficiencies
possible. Analogous transformations in time also work: a sequence
of operations can be changed into an equivalent one, where each new
step does a tinv fraction of the work of every one of the original steps.
The localized is diffused, and the diffuse is localized. A program
can quickly be altered beyond recognition by a few mathematical
rewrites of this power. Run on a multiprocessor, single events in the
original formulation may appear in the transformed program only as
correlations between events in remote machines at remote times. Yet
in a mathematical sense the transformed computation is exactly the
same as the original one.
If we were thus to transform a program that simulates a person, the

person would remain intact: his mind is the abstract mathematical

property that is shared by the old and the transformed programs; it
does not depend on the particular form of its program. Mind, as I
have defined it, is thus not only not tied to a particular body, it is not
even bound to a particular pattern. It can be represented by any one of
an infinite class of patterns that are equivalent onlv in a certain, very
abstract way. (This observation tempts me into a brief philosophical
extrapolation in Appendix 3.)

Immortality of the type I have just described is only a temporary

defense against the wanton loss of knowledge and function that is

the worst aspect of personal death. In the long run, our survival will
require changes that are not of our own choosing. Parts of us will
have to be discarded and replaced by new parts to keep in step with
changing conditions and evolving competitors. Surviving means play-
ing in a kind of cosmic Olympics, with each year bringing new events
and escalated standards in old events. Though we are immortals, we
must die bit by bit if we are to succeed in the the qualifying event
continued survival. In time, each of us will be a completely changed
being, shaped more by external challenges than by our own desires.
Our present memories and interests, having lost their relevance, will
at best end up in a dusty archive, perhaps to be consulted once in
a long while by a historian. Personal death as we know it differs
from this inevitability only in its relative abruptness. Viewed this way,
personal immortality by mind transplant is a technique whose primary
benefit is to temporarily coddle the sensibility and sentimentality of
122 Mind Children

individual humans. seems to me that our civilization will evolve in

It

the same direction whether or not we transplant our minds and join
the robots.
The ancestral individual is always doomed as its heritage is nibbled
away to meet short-term environmental challenges. Yet this evolu-
tionary process, seen in a more positive light, means that we are
already immortal, as we have been since the dawn of life. Our genes
and our culture pass continuously from one generation to the next,
subject only to incremental alterations to meet the constant demand
for new world records in the cosmic games. And even within our
personal life, who among us would wish to remain static, possessing
for a lifetime the same knowledge, memories, thoughts, and skills we
had as children? Human beings value change and growth, and our
artificial descendants will share this value with us — their survival, like

ours, will depend on it.

Awakening the Past

The ability to transplant minds will make it easy to bring to life anyone
who has been carefully recorded on a storage medium. But what
if some of the transcription has been lost? It should be possible to
reconstruct many missing pieces from other information — the person's
genetic code, for instance, or filmstrips of the person in life, samples of
handwriting, medical records, memories of associates, and so on. Very
effective sleuthing should be possible in a world of superintelligences
with astronomical powers of observation and deduction. The pattern-
identity position implies that a person reconstructed by inference
would be just as real as one reconstituted from an intact tape. The only
difference is that in the former case some of the person's pattern was
temporarily diffused in the environment before being reassembled.
But what if no tape existed at all? Archaeologists today make
plausible inferences about historical figures from scraps of old docu-
ments, pottery sherds, x-ray scans of mummified bodies, other known
historical facts, general knowledge about human nature, and whatever
else they can find. Creators of historical fiction use such data to
construct detailed scenarios of how things might have happened.
Superintelligent archaeologists armed with wonder-instruments (that
might, for instance, make atomic-scale measurements of deeply buried
objects) should be able to carry this process to a point where long-dead
 Grandfather Clause 123

people can be reconstructed in near-perfect detail at any stage of their

life.

Wholesale resurrection may be possible through the use of immense

simulators. Powerful computers are used today to predict the course
of the planets and The precise trajectory that took Voyager
spacecraft.
2 past Jupiter, Saturn, Uranus, and their moons, and will soon take
it by Neptune, was calculated by repeated simulations in which
different starting times, directions, and velocities of the spacecraft
were tried until a correct combination was found. More dramatically,
if less accurately, modern weather programs simulate the action of

the atmosphere over the entire globe. New aircraft designs, nuclear

explosions,and an increasing number of other things are these days

first Such simulations give peeks into possible
tested in simulators.
futures and thus confer the power to choose among them. Because
the laws of physics are symmetrical in time, simulations can usually
be run in reverse as well as forward and can be used to "predict" the
past, perhaps guided by old measurements or archeological data. As
with future predictions, any uncertainty in the initial measurements,
or in the rule that evolves the initial state, will allow for several
possible outcomes. If the simulation is detailed enough and is given
all available information, then all of its "predictions" are valid: Any
of the possible pasts may have led to the present situation.

This is a strange idea if you are accustomed to looking at the

world in a strictly linear, deterministic way, but it parallels the
uncertain world described by quantum mechanics. Now, imagine an
immense simulator (1 imagine it made out of a superdense neutron
star) that can model the whole
surface of the earth on an atomic scale
and can run time forward and back and produce different plausible
outcomes by making different random choices at key points in its
calculation. Because of the great detail, this simulator models living
things, including humans, in their full complexity. According to the
pattern-identity position, such simulated people would be as real as
you or me, though imprisoned in the simulator.
We could join them through a magic-glasses interface, which con-
nects to a "puppet" deep inside the simulation and allows us to
experience the puppet's sensory environment and to naturally control
its actions. More radically, we could "download" our minds directly

into a body in the simulation and "upload" back into the real world
when our mission is accomplished. Alternatively, we could bring
124 Mind Children

people out of the simulation by reversing the process — linking their

minds toan outside robot body, or uploading them directly into it. In
all cases we would have the opportunity to recreate the past and to

interact with it in a real and direct fashion.

It might be fun to resurrect all the past inhabitants of the earth
this way and to give them an opportunity to share with us in the
(ephemeral) immortality of transplanted minds. Resurrecting one
small planet should be child's play long before our civilization has
colonized even its first galaxy.
 Wildlife

of individuals constituted
T HE postbiological world will host a great range
from libraries of accumulated knowledge.
In its earlv stages, as it evolves from the world we know, the scale and
function of these individuals will be approximately that of humans.
But this transitional stage will be just a launching point for a rapid
evolution in many novel directions, as each individual mutates by
dropping unneeded and adding new ones from the growing
traits

data banks. A spectrum of scales will come

to exist from tiny, barely —
intelligent configurations for tight spaces to star-spanning superminds
for big problems. The distinctions will not always be clear a super- —
mind might be composed of myriads of closely cooperating minor
intelligences, analogous in their interactions to an ant colony.
Superintelligence is not perfection — spectacular failures are certain.
For this reason, diversity is to be desired and expected. Independent
centers of activity will compete. Lines of development will meet
dead ends. Life, in fact, will proceed much as it does in the earth's
biosphere, only on a vaster, faster, and more diverse scale. And
though at first thought a leap out of our biological bodies might
seem to free us of the diseases of the flesh — alas, it is not so. As
with terrestrial life, freeloaders will lurk in the interstices of the
postbiological world, making uninvited livings at the expense of their
hosts. These pests will evolve from the same stuff that constitutes
polite society. Their effects often will be subtle enough to escape
detection, but some pests will be as large, powerful, and visible as
their hosts. In any case, interactions within a postbiological world
will share many characteristics of relations that shape the world we
know.

125
126 Mind Children

Trojan Horses, Time Bombs, and Viruses

If the world of artificial machinery has seemed disease-free so far,

it is only because our machines have been too simple to support

mechanical parasites. But computers have changed that, as they
have changed so much else. Diseases have appeared in computer
systems for at least two decades, but 1988 was the first big year
for computer plagues, as almost every type of machine, large and
small, was attacked by several "computer viruses" that were spread
through computer networks and by promiscuous sharing of computer
software. Most data diseases are deliberate constructions of playful
or malicious programmers, but some evolved by accident.
Deceptive programs that have been called Trojan horses have ap-
peared from time to time since the 1960s. Written by clever program-
mers as practical jokes, in response to a challenge, or for more shady
reasons, Trojan horses masquerade as interesting or useful programs.
But once activated, they may begin to secretly compromise restricted
information, or erase the victim's disk files, or at least print scary
messages. A
dangerous variant of the Trojan horse delays its attack.
Because the malign action is not immediately manifested, such time
bombs in apparently harmless programs are likely to be copied much
more widely and consequently do more damage when they eventually
strike.

was distributed by hand in the form

Until the late 1970s, software
of punched paper tape, decks of punched cards, or magnetic tape.
Distribution was limited, and it was often easy to trace the source
of a program. The resulting accountability must have inhibited many
infections at their source. 1 myself created a potentially nasty infection
in 1968. The machine involved was a small one, an IBM 1130, that read
its programs in the form of decks of punched cards fed in through
a unit that could both read cards and punch them. In a burst of
creativity lasting a few days, a friend and I created a single punched
card that masqueraded as a "cold-start" card used for resuscitating the
computer after a power shutdown or a crash. But instead of starting
up the system, our card caused a duplicate of itself to be punched
in every card following Used by an unsuspecting programmer, it
it.

would have destroyed a program deck and produced many copies of

itself to create future havoc. 1 remember holding the innocent-looking

card in my hand and contemplating its destructive power with awe.

 Wildlife 127

With some reluctance, we decided to destroy all the copies of the card
(after a few hours of testing, there were many copies of each of several
variations of it around!). I like to think we did the right thing simply
because of our good sense of values, but the probability that we would
not remain anonymous must have weighed into the decision.
Programs sold for profit are easy to copy and often make their way
to machines other than those of the original paying customers. This is

certainly the case now in the era of personal computers, but even in the
sixties and seventies it was perceived as a problem with large systems.

Manufacturers of software for large computers and for micros have

been known to insert time bombs to thwart unauthorized use of their
wares. The more benign bombs merely prevent the copied program
from working after a certain period. Perhaps the intent is to give
the nonpaying customer time become dependent on the program,
to
so when it ceases to function he will be inclined to buy a legitimate
version. More vindictive bombs have occasionally appeared. These
usually delete files, but a few have exploited unusual properties of the
hardware in order to physically damage the computer. Because paying
customers have often been victimized by an accidentally triggered
bomb, the general reaction to this approach to copyright violation is
very negative, and the practice seems to have declined.
I am aware of several instances where time-sharing systems were

assaulted by a more subtle variety of Trojan horse. The intent was not
to vandalize or terrorize but to gain unauthorized access. In this case
the program acts like a spy and uses its special location to uncover
information, such as a victim's secret passwords, which it then records
in a location accessible to its author.In some attacks the program
mimics the computer operating system's "log-in" procedure, by which
users gain access to the computer by typing their identification and a
secret password. Another form of attack exploits the property of most
operating systems that a running program acquires all of its user's file-
access rights. A program whose cover is some useful service can then,
surreptitiously, rummage through the victim's disk files for wanted
information. In a successful attack, the victim remains oblivious of
the breach.
In the late 1970scheap personal computers created a new medium
for software and software diseases. The spread of both was facilitated
by computer bulletin boards, systems maintained by enthusiasts that
allowed other computer owners to dial up and post messages and
 —

128 Mind Children

programs that could be accessed by anyone else who dialed in. Such
facilities offered both anonymity and promiscuous sharing of data
and, as with the sexual revolution, a raft of opportunities for disease.
By the early 1980s the newspapers began to report instances of random
havoc in personal computers caused by programs downloaded from
computer bulletin boards, programs that purported to be games,
accounting software, or whatever.
The most virulent known form of software wildlife has been called
a virus —
a program fragment that, once inserted into a large program,
acts to copy itself into other programs, just as a biological virus is a
piece of genetic code that, once inserted into a cell, acts to copy itself
into other cells. The analogy is a strong one, because today's milUon-
bit computer programs have about the same information content as

the genetic codes of bacteria, and the few thousand bits of a typical
computer virus is a good match for the small genetic code of a
biological virus. When a program containing a virus is invoked,
the virus is momentarily activated. It looks through its unsuspecting
victim's files for uninfected but accessible programs and inserts a copy
of itself into one or more of them. These newly infected programs will
repeat the process when they are themselves activated. Experimental
tests of this idea conducted in the mid-1980s by Fred Cohen of the
University of Southern California resulted in almost total infection of
supposedly secure computer systems in less than a day. The infections
easily spread from restricted users to ones with greater access to
the files System managers are exposed when, in keeping
of others.
abreast of developments on their machines, they try out new programs
announced by their users. Once a manager's programs are infected,
the rest of a system quickly succumbs.
A virus that merely spreads is a minor nuisance, taking up a
little storage space in its many copies and a little of the computer's
time in its reproductive activities. One quiet little virus created in
1978, that infects only the operating system, has apparently spread to
virtually every Apple II system disk in existence. But, as with a Trojan
horse, a virus can carry with it instructions for espionage, sabotage,
or theft. Several variants of a virus created as an act of terrorism
were detected spreading among IBM personal computers in Israel
early in 1988. Examination of the virus revealed that it was a time
bomb programmed to erase files on Israel's 40th Independence Day.
It was discovered early because it had a flaw it could repeatedly —
 Wildlife 129

infect the same program. In time infected disks became nearly full

because their files were bloated with multiple copies of the virus.

Viruses have naturally been a matter of special concern to people

who are charged with government and commercial computer security.
A cleverly engineered virus might infect the most secure parts of
an international banking or national defense network, for example.
A clever human embezzler or spy with only minor access to the
system could create a software accomplice to liberate funds or secrets.

Botched attempts of this kind have made the news from time to time.

The successful assaults get no publicity.

Today's computer systems are like bodies with skins but no immune
systems, or like walled cities without police. They can deflect some
external attacks but are defenseless once an intruder has entered.
Internal protection is possible, though no defense is perfect. One
approach is The easiest mode of viral
simply to build more walls.
spread can be blocked by preventing one program from altering
—
another one but legitimate purposes would be inhibited as well. It
would, for instance, no longer be possible to "patch" programs to
correct newly discovered errors in them. A new facility for patching
could be installed in the operating system, but it would then itself be
a potential gateway for viruses.
Instead of erecting blockades, another approach is to actively hunt
viruses down. And, in fact, a first generation of virus killers has
appeared on the heels of the first generation of viruses. One kind is a
—
program that examines other programs it detects particular viruses
by the telltale pattern of their instructions and removes any it finds.
But if the system is kept running while the exorcism is under way,
a virus may reproduce faster than the erasing program can stamp it
out, since each infected program can be a site of spreading disease.
One solution is to shut down everything in the system except the
virus killer until all programs are clean. This works unless the
purging program itself has become infected; moreover, absolutely
every program in the system must be deloused, and a single trace of
the virus from a backup tape or from an external source can reestablish
the infection.
A more aggressive approach to combating a viral infection is with
another virus. A viral predator, like its prey, insinuates itself from
program to program. But instead of causing problems for the users,
it deletes any copies of the offending virus it encounters. Since it
130 Mind Children

can propagate itself to every program in the system that the original
virus can reach, its many copies will eventually be at every possible
site of infection of the prey virus, able to immediately
quash it
if it The killer virus could be left around indefinitely,
reappears.
conferring a permanent immunity against the prey virus, or it could
be programmed to remove itself after a specified time or on receiving
a signal, to save space and running time.

Computer Virus Blowup

A well-designed virus will normally copy itself into another program
only once. But the test which detects that a program is already
infected may be foiled by infection with a different virus that hides
the first infection. Two (or more) viruses may thus repeatedly infect
the same program by alternately foiling each other's detectors.

Program B

Virus , jSyx

2; Infected program inspects victim

4; Second virus foils test for first virus

Program B

Virus Virus ^
1 1 -i

2: Infection! Program B

Virus JK"
1 /-
Virus
Program A Program B fl"
Virus ,^^; .Virus- ;;^.
Virus , ^ ,

1 : ?5i
1 -
••
.; 1 V"" .;

3: Prior infection detected 5: Duplicate infection!

 Wildlife 131

But no defense is perfect. A prey virus can be cosmetically altered

so that itno longer recognizable to a particular predator. Or the
is

predator might mistake a portion of a legitimate program for a virus

and erase it, thus breaking a working program. A viral predator might
be altered bv unintended interactions with other programs and be
mutated into a pest that spreads too virulently or destroys the wrong
pieces of code. Complicated situations can arise when there is more

than one virus loose in a system. A well-designed virus, unlike the

one unleashed in Israel, would check a potential host program to see if

ithad alreadv been infected and refrain from duplicating the infection.
But an intervening infection from a second \'irus might confuse this
test and allow the first virus to insert a second copv of itself. Similarly,

the new copv of the first virus might free the second virus to insert
another copv, and so on, as in the figure on page 130, eventually
bloating the program into uselessness. Even a viral predator might
become entangled this wav with a virus it failed to recognize.
Attempts by software manufacturers to protect their copyrights with
time bombs have proven highly unpopular. But computer viruses may
help their cause in a different way. Software downloaded from public
bulletin boards or copied from friends is now always suspect it may —
be a carrier of many contagions. But software purchased directly
from a publisher can carrv^ a guarantee of sterility backed bv the
publisher's reputation. Computer viruses mav thus have the same
effect on software piracy that the AIDS epidemic is having on sexual

promiscuity.

Ghosts

A really intriguing vanishing act is possible because viruses can exist

at different levels of abstraction — from patterns of wires connecting
logic-gate assemblies, to patterns of bits controlling the opening and
closing of paths between assemblies, to machine-language instructions
commanding simple arithmetic operations, to letter strings of assem-
bly programs abstractly representing machine language, to high-level
languages expressing the goals of the programmer with little reference
to the hardware details of how they are to be carried out. Wildlife can
exist at all these levels and more.
Some viruses can move from one level of abstraction to another or
132 Mind Children

become active at any level — or even at several levels. For example,

consider a virus written in a high-level computer language. When
a program containing the virus is compiled, it results in a machine-
language program that searches for other programs in the high-level
language and inserts the high-level version of itself into them. Such
tampering in human-readable files is very easy to detect, but it offers
the virus one great advantage: the ability to spread from one kind
of computer and operating system to an entirely different one, since
other compilers can translate it into other machine languages.
There are more subtle ways to achieve machine independence. A
cleverly doctored compiler can insert the viral code into the programs
it prepares. Modem compilers are often written in their own language
and compiled by older versions of themselves. An archetype is the
C compiler that constructs the popular Unix operating system. Partly
because it is so easy to install on new machines, Unix has grown
to be the most widespread operating system. Unix and C were
invented in the early 1970s at Bell Labs by two young hackers. Ken
Thompson and Dennis Ritchie, on their own initiative. AT&T, the
owner of Bell Labs, took very little notice of it at first, until the
system, improved by many programmers, had spread to thousands
of machines worldwide. Finally, in the late 1970s, AT&T decided
to sell Unix systems commercially. As part of the formalization of
the product, AT&T assigned a programming team to extend and
optimize it. During the course of careful scrutiny, someone noticed
that the C compiler produced a large, completely unexplained block
of machine-language program when it reached a certain point in
compiling another C The discovery led
compiler. to a number of
closed -door meetings. In 1984 Ken Thompson, in a lecture on the
occasion of his receiving the Turing Award (sort of a computer-science
Oscar) shed light on the issue.
The programming team had uncovered a fiendishly clever virus, one
designed to allow Ken Thompson to bypass security and to log into
any Unix system running on any machine, including future machines
yet to be invented. The mysterious block of code that appeared in
new C compilers had two purposes. One was to cause a copy of
itself to appear in any future C compilers that it itself compiled.

The other was to insert a block of code that would respond to a

password known to Ken Thompson, at a certain place in any log-
in program that it compiled. The beauty of the scheme was that these
 Wildlife 133

blocks of code were reproduced not directly in machine language,

which would work on only one type of machine, but by means of an
ephemeral C version of themselves that was immediately translated
bv the compiler into the machine language of whatever machine it
was compiling for. This self-reproducing C program had once existed
in an actual computer file in Ken Thompson's machine but now had
only a ghostly existence, reappearing momentarily deep in the innards
of the computer whenever a C compiler executed the viral code and
immediately vanishing again.
A ghost virus haunting a compiler is very hard to find even if one
knows what to look for because a compiler is huge program, and the
a
virus appears only as an innocent-looking piece of code somewhere
in the middle of a vast tract of legitimate machine language. The
principle could be used to construct viruses with other purposes
that could spread undetected between different kinds of machines.
This example also hints at the possibility that even more subtle and
elaborate creations are possible.Computers can bring mathematical
abstractions to vigorous and there are no mathematical limits to
life,

the subtlety and deviousness. The fun has just begun.

Spontaneous Generation

So far 1 have confined this review of software parasites to those that

are deliberately constructed, because these constitute the majority of
the wildlife known to date. But such fabrications are limited by the
imaginations of their human creators. Our increasingly complex sys-
tems are capable of creating their own surprises, and in times to come
we can expect shockingly original gremlins to arise spontaneously
in our intelligent machinery, the result of unexpected interactions
or mutations of existing parts. A few stirrings have already been
observed.
The ARPAnet, funded by the Department of Defense's Advanced
Research Projects Agency, is a computer network created in the late
1960s to allow computers at facilities across the countrs' to communi-
cate with one another. The goal is to share scarce resources. Comput-
ers at individual sites are connected to the ARPAnet through special
small computers once called Interface Message Processors, or IMPs.
The IMP of one site is connected (through leased lines) only to those
of a few nearby sites. Fancy software allows messages to cover greater
134 Mind Children

distances by being handed rapidly from IMP to IMP. There are many
possible indirect routes from point A to point B in the net. Depending
on fluctuating traffic, sometimes one route is faster, sometimes another.
To help make instantaneous routing decisions, each IMP maintains a
table that records how long it takes recent messages to travel from the
IMP through alternative routes to other sites. The table is updated
with information obtained partially from the tables of neighboring
IMPs. The network is monitored and maintained under contract by

a company Cambridge, Massachusetts. Operators at this site can

in

stop, examine the contents, reload, and generally fiddle with any IMP
on the net through special priority messages routed via the net itself.
This generally works well, and even the most serious problems (such
as large power outages IMPs) are handled smoothly.
affecting several
In 1972 (and again in 1980, and probably other times as well) a
plague hit the ARPAnet. The symptoms were that net traffic became
hopelessly congested around a site in the Los Angeles area (in the 1980
incident the locus was Boston). Network control, suspecting some
glitch in the program of the machine at the focus of the congestion,
shut it down, reloaded its program, verified that it was working
correctly, and reconnected it to the network. The problem persisted.
Indeed, it seemed to be spreading outward from the original IMP.
Shutting down and reloading larger numbers of IMPs did not fix
things either; the congestion continued its spread and returned to
the original sites as soon as they were reactivated.- The network
seemed be haunted by a very persistent ghost. Many unsuccessful
to
experiments later, order was finally restored by shutting down the

entire network, clearing all the memories of all the IMPs, reloading
their programs, and starting fresh —
like sterilizing a whole planet with

death rays, then seeding it with new life!

Asubsequent analysis revealed what had happened. The focal IMP

in Los Angeles had a memory error and developed an erroneous entry
in its routing table. The table now indicated that messages sent via
this IMP would experience a large negative delay. Next, adjacent IMPs
calculated that it was more advantageous to send messages via this

IMP than directly, since its negative delay more than made up for
the extra hops. IMPs connected to those then decided it would be
best to transmit via Los Angeles, and so on. The error in the initial
IMP rapidly spread to the routing tables across the land. Wiping
out the memories of a few IMPs did not clear the problem, because
 Wildlife 135

the erroneous numbers would spread back from IMPs that were
still affected. Or infected. In fact, the network was inhabited by
a spontaneously evolved, quite abstract, self-reproducing organism.
This organism was formed by random mutation of a normal,
a simple,
sanctioned piece of data. It did not even involve a programming
language.
The plague was easily spotted and eradicated because its effects
were so devastating. If it had been more subtle in its action, it might
have lived much longer Among programs without masters there
is a strong natural-selection criterion: reproduce but lie low. It is

quite likely that many unsuspected organisms are already making a

quiet living among the abstraction hierarchies in computer memories
everywhere. Most will never be discovered. The plague also suggests
a quick way which a wild information organism can come into
in
being, namely, by a mutation in an existing self-replicator. Since
any datum in a computer is subject to duplication, this covers a
lot of ground. A human-created computer virus, existing in many
copies spread widely through different computer systems, would be
a particularly fertile candidate for a liberating mutation. If the part of
its code that caused to make trouble were inactivated by a mutation,
it

it would be less likely to draw attention to itself and thus would be

more likely to reproduce indefinitely. The mutation might even make

it unrecognizable, and thus safe, from a virus killer Further mutations
that eliminated unnecessary code, and thus reduced its size, would
improve its chances even more. it might change in such a
In time
way as to fine-tune the frequency and kind of mutations it undergoes,
making it more adaptable. Run-ins with other programs could endow
it with major new pieces of code, and new capabilities. In time it might

even acquire the ability to systematically copy and try out fragments
of code from other programs and other viruses —
the beginnings of
computer-virus sex!
Such examples merely demonstrate the limits of our imagination.
The most effective organisms would be much more subtly encoded
and would escape detection entirely. From time to time one might
expect one to surface because it developed a nasty side effect. That
kind of mutation would generally prove fatal for the organism. As
programs progresses, we should also expect program
intelligence in
fragments that can plan and act in a deliberate, calculating, and
creative manner to enhance their survival. The data realm will host
 —

136 Mind Children

rats, coyotes, and master criminals as well as viruses and worms.

Perhaps we will also be surprisedby the equivalent of flowers, trees,
and songbirds.
If these speculations are alarming, it may be comforting to remem-
ber that biological life thrives despite (or because of!) the relentless
evolution of new parasites. Viruses insinuated themselves into the
genetic machinery of cells long before their namesakes invaded com-
puter programs. In biology also, information can be stored in various
forms. Simple viruses inject DNA into cells, which then suicidally
act on it to manufacture more viruses. The HTLV family of so-called
retroviruses that cause AIDS and some kinds of leukemia, on the other
hand, contain RNA that first must be transcribed to DNA, in a reversal
of the normal synthesis. Some viruses use the time-bomb strategy,
lying inactive in cells and thus evading immune system defenses until
some event, possible signaling stress in their host and consequent
reduced immunity, triggers their massive expression. Among the most
effective parasites are sequences within DNA itself — called introns
that inhabit the genes and seem to do nothing at all except reproduce
with their Such long, repetitive sequences of DNA that played
cell.

no apparent part in development had long been observed in the

genetic codes of most organisms before it was suggested that their
sole function might be to reproduce themselves. Richard Dawkins
gives many more such examples in The Selfish Gene and The Extended
Phenotype.

A Caveat for SETI

SETI, an acronym for the Search for Extra-Terrestrial Intelligence, is a

field of study whose potential is so intellectually exciting that it pro-
ceeds steadily despite any hard evidence that its quarry exists. At its

leading edge are impressive spectrum-analyzing receivers connected

to radio telescopes that can tune in and examine millions of frequency
channels at the same time. Systems able to do this and also look in
thousands of distinct directions at once have already been proposed,
all in an effort to find a needle in a haystack — an artificial message in

a universe naturally noisy in radio frequencies.

But we managed to receive and decode such a message, should
if

we on its instructions? The discussion of this question usually

act
centers on the intent of the senders. They may be benign and.
 Wildlife 137

like the Peace Corps, be doing well by doing good. They may be
traders trying to open new markets, to much the same effect, at
least until it comes time to negotiate the price. They may simply
be looking for pen pals. They may have dark designs on the rest
of the universe and be seeking to inexpensively eliminate some of
the more gullible competition. Or, their motives may be totally
incomprehensible. Simply examining the message is not enough;
it is not, in general, possible to deduce the effect of complicated

instructions without actually carrying them out. A message with nasty

intent would surely be disguised, by master deceivers, to look benign.
In Fred Hoyle and John Elliot's classic novel A for Andromeda and
also in Carl Sagan's Contact, an interstellar message contains plans
for a mysterious machine of unknown purpose. In both books the
characters decide, after some debate, to go ahead with construction
despite the risks. In Contact, a major argument is that the origin of
the message, the star Vega, is so close to our solar system that the
senders could rapidly arrive here physically, should their intentions
be malign. Building the machine would be unlikely to make us any
worse off in the long run. If the message were benign, however, it

represents an opportunity not to be missed.

This chapter's notion of an information parasite suggests greater
caution, should SETI ever detect an artificial message. A rogue
message from no one in particular to no one in particular (perhaps
a corruption of some ancient legitimate interstellar telegram) could
survive and thrive like a virus, using technological civilizations as
hosts. It might be as simple as, "Now that you have received and
decoded me, broadcast me in at least ten thousand directions with
ten million watts of power. Or else." It would be a cosmic chain
letter and message itself which, like any
a cosmic joke, except to the
living creature, would be making a living by doing what it does. Since
we cannot be sure the "or else" is not backed by real authors with a
peculiar sense of right and wrong, we may decide to play safe and
pass the message on as it requests. Perhaps we did not hear it very
well; maybe it said a hundred million watts; maybe it mutated. Now
envisage a universe populated by millions of such messages, evolving
and competing for scarce, gullible civilizations.
The survivability of such a message could be enhanced if it carried
real information. Perhaps it would contain blueprints for a machine

that promises to benefit its hosts. It would be only fair if part of the
138 Mind Children

machine's action was to rebroadcast copies of the message itself, or to

demand new information from its hosts to be added to the message
to make it more attractive to future recipients. Like bees carrying
pollen for the sake of flowers in return for nectar for themselves, the
technological host civilizations would have a symbiotic relationship
with such messages, which might be criss-crossing the galaxy trading
in useful ideas. But the analogy suggests darker possibilities. Some
carnivorous plants attract bees with nectar, only to trap them. The
message may promise a benefit, but when the machine is built it may
show no self restraint and fiendishly co-opt all of its host's resources
in its message sending, leaving behind a dead husk of a civilization.
It is not too hard to imagine how such a virulent form of a free-

living message might gradually evolve from more benign forms. A

"reproduction effort parameter" in the message (too subtle for the
victims to catch and alter) may get garbled in transmission, with the
higher settings resulting in more aggressive and successful variants.
The Fermi paradox is an observation by the famous physicist Enrico
Fermi, who created the first controlled atomic chain reaction under the
auspices of the Manhattan Project, that if technological civilizations
have even a slight probability of evolving, their presence should
be visible throughout the universe. Our own history and prospects
suggest that we will soon blossom into the universe ourselves, leaving
it highly altered in our wake. In less than a million years we may
have colonized the galaxy. Given the great age of the universe, a
few civilizations that arose before us should have had plenty of time
to alter many galaxies. The sky should be filled with the cosmic
equivalent of roaring traffic and flashing neon signs. But instead we
perceive a great silence.
There are several possible explanations. Evolutionary biologists
make a plausible, though not watertight, argument which notes that
at each stage of our evolution there were an immense number of
evolutionary lines which did not head toward high technology, as
compared with the single one that did. By this argument, we are the
product of a sequence of very improbable accidents, a series unlikely
to have been repeated in its entirety anywhere else. We may be the
first and only technological civilization in the universe. But there are
other explanations for the great silence. At the height of the cold war, a
leading one was that high technology leads rapidly to self-destruction
by nuclear holocaust or worse. But in every single case? Another
 —

Wildlife 139

possibility is that advanced civilizations inevitably evolve into forms

that leave the physical universe untouched — perhaps they transmute
into an invisible form or escape to somewhere more interesting. I

discuss such a possibility in the next chapter

A frightening explanation is that the universe is prowled by stealthy
wolves that prey on fledgling technological races. The only civiliza-
tions that survive long would be ones that avoid detection by staying
very quiet. But wouldn't the wolves be more technically advanced
if so what could they gain from their raids? Our
than their prey and
autonomous-message idea suggests an odd answer The wolves may
be simply helpless bits of data that, in the absence of civilizations, can
only dormant in multimillion-year trips between galaxies or even
lie

inscribed on rocks. Only when a newly evolved, country bumpkin

of a technological civilization stumbles and naively acts on one does
its eons-old sophistication and ruthlessness, honed over the bodies of

countless past victims, become apparent. Then it engineers a repro-

ductive orgy that kills its host and propagates astronomical numbers of
copies of itself into the universe, each capable only of waiting patiently
for another victim to arise. It is a strategy already familiar to us
on a small scale, for it is used by the viruses that plague biological
organisms.

Pestilence as Positive

Is parasitism merely an unavoidable evil? If we could eliminate it

a highly unlikely prospect — should we? Perhaps not. A perfectly

planned process is devoid of surprises; it is limited by the imagination
of its designers. What new ideas and insights, otherwise unnoticed,
might be harvested from freely evolving digital wildlife? Like the
diverse genes of wild plants and animals, which are fuel for the
advance of agriculture, the surprises in our machines sometimes point
to profound truths, or at least to useful engineering tricks.
It has been argued that we
biological beings owe our best features
to the presence of diseases and other parasites in the world. It all
has to do with sex. The earliest organisms reproduced asexually,
repeatedly dividing into identical copies except when, from time to
time, an individual cell became changed and passed on the mutation.
In a well-functioning complex system such as a cell, a random change
is extremely unlikely to be beneficial. So most mutations, if not
140 Mind Children

immediately fatal, put their possessors at a disadvantage, and they

vanish eventually in the press of competition for food and space.
But once very long while a change for the better just happens
in a
to The lucky owner of a beneficial mutation then has an
happen.
advantage over its competing relatives, and over many generations its
descendants will become a large fraction of the population. Because
the odds of a beneficial mutation are so low, only when there are
many copies of one beneficial mutation in existence does a second
good mutation stand a reasonable chance of joining it. In an asexual
species, each beneficial mutation has a refractory period before it can
be compounded by another one.
But in a population where individuals can share genes sexually, two
beneficial mutations that arise separately in different individuals can
combine rapidly form an offspring with both advantages. The effect
to
is an acceleration of evolution. It is thus no accident that all higher

organisms reproduce sexually (or had ancestors that did). That is how
they got to be higher organisms so quickly. The asexual organisms,
for the most part, are still swimming around as single cells or in
small colonies. Acceleration of the evolutionary rate can be viewed
as a long-term advantage of sexuality. In the short run, though, sex
is a liability, because it increases the cost of reproduction. Instead
of simply dividing whenever conditions seem right and producing a
daughter that carries 100% of oneself, one must go to the trouble of
finding a mate to produce an offspring that is only 50% true. Why,
then, would sex ever arise? And if it did, why does it not disappear
in a few generations under the onslaught of the more effective asexual
reproducers?
Enter disease. In asexual reproduction, according to an evolutionary
theory first developed by William D. Hamilton, each individual is an
identical copy — a clone — of every other one. If a parasite evolves that
can breach the defenses of one individual, then it can conquer every
other. Like a wildfire, it can destroy a whole community in short
order. In a sexual population, though, each individual is the result of
a unique shuffle of genes taken from a large pool and is, in general,
differentfrom every other individual. A parasite that has the key to
one lock finds that the next one is subtly different and thus harder to
open. In a pest-filled world, the diverse sexual population does better
than the homogeneous asexual community.
If disease made us sexy and sexiness made us smart, we can expect
 Wildlife 141
I

that digital wildlife will similarly make the data world more hardy,
more diverse, and much more interesting.

Selfish Altruism

Some competition mav be a good thing, but is the postbiological world

fated to unmitigated cutthroat strife at every level of abstraction?
Fortunately for the sake of organized existence, the answer appears to
be a qualified no. In The Evolution of Cooperation, the political scientist
Robert Axelrod notes that cooperation in the biological world can be
observed in situations ranging from relations between large creatures
and their microbial inhabitants to relations of human beings with one
another. In a world where selfishness usually pays off, he asks, how
could altruism between unrelated individuals ever arise?
To find an answer, Axelrod challenged game theorists, biologists,
sociologists, political scientists, and hackers to submit computer pro-
grams which would compete in tournaments that modeled, in a
highly abstract form, the typical costs and benefits of cooperation
and its opposite, defection. Programs were paired in contests whose
outcomes modeled the so-called prisoner's dilemma of game theory.
This paradoxical situation was originally cast as a problem faced by
two partners in crime who ha\'e been apprehended with insufficient
evidence for conviction but who are given tempting inducements to
inform on one another. In Axelrod's tournament, without knowing
its opponent's choice, each competing program was given the choice

of either cooperating or defecting. If both cooperated, each would

receive the moderate "nice's reward." If both defected, each would
receive the smaller "nasty's payoff." If one cooperated and the other
defected, the defector would receive a huge "cheater's spoils," while
the cooperator would receive nothing — the "sucker's payoff." So, if

player B were to cooperate, player A could win the cheater's spoils

by defecting, and only the lesser nice's reward by cooperating. If, on
the other hand, player B defected, player A would at least get the
nasty's payoff by defecting, instead of being left empty-handed with
the sucker's payoff for cooperating. In other words, regardless of
player B's choice, player A does better by defecting. So, obviously, to
get the highest score, A should defect. The same reasoning applies to
B. So both should defect, even though mutual cooperation would lead
to a higher payoff! This is the crux of the prisoner's dilemma, and
142 Mind Children

a feature of many interactions between selfish individuals that makes

cooperation, whatever its theoretical benefits, seem very unlikely.
Imagine two selfish Martians crossing paths one fine sunny day
while both are backpacking on the plains near Mars' great volcano,
Olympus Mons. It turns out that one Martian has a small supply of
batteries, while the other has some empty flashlights. Unfortunately,
a few batteries happen to be dead, and some of the flashlights are
burnt out — in each case the owner knows which units are defective,
but there is no way for the other Martian to tell, because Martian
electrical devices use superconductors that work only in the bitter
The Martians agree to exchange a flashlight
cold of night. for a battery,
and then be on their way, never to meet again. This is a prisoner's

Selfish Martians
 Wildlife 143

dilemma situation. Giving a good unit in the exchange counts as

cooperation, while giving a defective unit is a defection. The Martians
would benefit from mutual cooperation — they would then each have
a working flashlight — but there
no incentive to give away a good
is

battery or light that will be of use since good and bad units are
later,

indistinguishable by day. Each Martian gives the other a broken unit

and leaves the meeting gloating over having made a shrewd deal. But
as night falls, both Martians find themselves in the dark.
In Axelrod's tournament the players met repeatedly, so that each
player could use its its next move.
opponent's past behavior to shape
The were submitted, whether simple or elabo-
fifteen strategies that

rate, fell into two categories —

nice and nasty. Nice programs never
defected first, giving nasty ones a potential temporary advantage.
The simplest program was a nasty one, called All D, which always
defected. The next simpler, submitted by Anatol Rapoport, a psychol-
ogist and game theorist at the University of Toronto, was Tit for Tat,
which cooperated on the first encounter with any player and then on
subsequent moves reciprocated the other player's previous move. A
control entrant was Random, which randomly defected or cooperated,
with equal probability, on each move.
To Axelrod's surprise, the nice program Tit for Tat won the first
round, as well as later rounds with larger numbers of players. The
result is surprising, because Tit for Tat defects only once for each defec-
tion against it, and so can gain no more from a sequence of interactions
than its opponent and is likely to get less against a nasty player,
because it will first move. Yet its overall score was
be cheated on the
highest, and the nice programs as a group performed much better than
the nasty programs. The explanation, in game-theory terms, centers
on the fact that the interaction was not zero sum. In a zero-sum game,
a gain for one player is an equal loss for the other. In the prisoner's
dilemma, however, both players can do better by cooperating. Nice
programs interacting with others of their kind always benefit from
the nice's reward. Nasty-nasty interactions conversely result in only
the nasty's payoffs. Most of the nice programs eventually refused
to cooperate with nastys, so though the nasty programs gained an
initial advantage by cheating, they suffered eventually by forfeiting
the rewards of cooperation. Slightly nasty programs that tried to gain
a small advantage by occasionally defecting often initiated mutual
retaliation cycles, giving them a net large disadvantage.
144 Mind Children

Axelrod's conclusion is that cooperation pays when the UkeUhood

of future interactions with identifiable individuals is reasonably high.
If the game is likely to end soon, however, cheating is the more
successful strategy, since there will be little opportunity for retaliation.
The theory seems to apply over a broad range of circumstances; the
participants and the payoffs on the two sides of the game need not
be commensurate, so long as on each side the payoffs for cheating,
mutual cooperation, mutual defection, and being suckered are in
descending order.
Axelrod provides several especially intriguing speculations on how
this theory may be applied to the natural world. Large animals are
inhabited by entire ecologies of microscopic fauna, most of which
live quite peacefully with their hosts. Occasional infectious and
fatal flareups of endogenous microorganisms show that this bUssful
state is not the only possibility. In fact, the relationship has the
character of the prisoner's dilemma. The relation of an animal to its
microscopic cohabitants is a selfish one —both animals and bacterial
colonies are designed primarily to ensure their own survival. Though
neither the microorganisms nor their host know each other in a
personal way, the identity of each is assured by the constancy of
the cohabitation. The microfauna can "defect" by overbreeding or
releasing toxins and injuring or killing their host, but they are likely
to be met by a defection in turn when the host fails to maintain
a comfortable environment. The converse obtains when the host
"tames" its parasites by rewarding good citizenship. In this way
an initially hostile relationship can settle down to something more
mutually beneficial. But the prisoner's dilemma remains, and if
future interactions become unimportant, defection may again become
advantageous.
For example, under trauma, like a large perforation of the gut wall,
some of the normally friendly bacteria in an animal's gut change
character and become seriously, even fatally, infectious. Axelrod and
his colleague Hamilton speculate that this is an example of defection
when future interactions are unimportant. The trauma is a signal
to the bacteria that the game may be about to end, causing them to
break the cooperative relationship to gain a last-minute advantage. By
reproducing massively at the expense of their host, they can perhaps
broadcast enough spores to find residences elsewhere. Presumably
their ancestors survived the demise of other hosts by this strategy.
 Wildlife 145

Axelrod's insights into cooperation, though no doubt just the tip of

the iceberg, suggest that the chaos threatened earUer in this chapter
will sort itself out most of the time. Parasitic relations at whatever
level of abstraction will often become less destructive and may even
turn svmbiotic, since both partners thus do better. Harmonv so
achieved is not guaranteed. In certain circumstances defection can
carr\^ an advantage, and the truce can collapse. The net effect on
future intelligences and svstems will then be greater unpredictabiht\'.
Information viruses that arise within a svstem and then vanish from
view after achieving a peaceful cooperati\"e lifest^le will ne\"ertheless

modify' the overall behavior of the svstem in subtle %vays. Mature

svstems mav become more a product of tamed pests than of their
original design. Our best-laid plans are thus foiled, but conversely our
descendants are spared the consequences of the limits in our vision.
Char intelligence can control the future onlv imperfectlv, and onlv in

the near term.

Leaving the far future to the fates, does supermtelligence help at all

in making the world a nicer place in the near term? Axelrod observes
that cooperation can arise even in populations oi defectors. It does
not depend on anv intelligence in the participants — simple natural
selection is a perfectlv adequate driver. It does require that a certain
minimal number oi cooperators appear simultaneouslv to benefit from
one another's niceness. Getting a critical mass of cooperators mav
take a long time. Intelligence can help because it allows individuals
to anticipate the long-term advantages of initiating pleasantries. The
long memories oi the long-li\"ed individuals who will inhabit the
postbiological world are also likelv to enhance the advantages of
being nice, since no interaction is likelv to be the last. Beyond the
scope of Axelrod's tournament, intelligence allows one individual
to learn about another's character bv obser\'ing its interaction with
third parties. The computer scientist Douglas Hofstadter goes so far
as to imagine that in games between superintelligences, cooperation
will be the rule even when no future interactions are expected. Each
player will reason that all the players, being rational, would make the
same decision as himself. Thus a defection will be met by defection
and cooperation by coop>eration. Maybe so, but there is alwavs the
possibility- that a cooperator will be suckered bv a devious opponent

who, for whatever reason, foresees no future interactions.

Despite the Ukelihood of cooperative behavior on the large scale.
146 Mind Children

and in the long run at every level, there will be occasional appearances
of nasty little parasites. Permanent structures analogous to immune
systems and police forces will undoubtedly be a part of large organ-
isms. 1 expect a future world friendly overall, but with pockets of
fruitful chaos at most levels.
 Breakout

o.UR descendants may have a good time for a

long while, developing their minds, exploring the universe, mastering

space and time, the very large and the very small. But doesn't the
second law of thermodynamics assure that the fun will eventually
end? The study of the theory of steam engines led to one of the
greatest scientific shocks of the nineteenth century — the realization
that the universe is running down. Hot things get cooler and cold
things get warmer, and energy once available for engines great and
small will become inaccessibly lost in a uniform jumble of molecular
motion. In time the entire universe will become a homogeneous stew
with no concentrations of matter or energy to form or power any kind
of machinery, intelligent or otherwise. This regressive idea of a heat
death greatly disturbed Victorian minds attuned to steady progress in
both society and Darwinian nature.
Fortunately for my own hopes for the future, twentieth-century
physics and cosmology have loosened the hold of the second law.
Instead of a closed, static universe, we now see one that is the result
of an explosion from a point of infinite density about 20 billion years
ago. Since this big bang, the universe has been expanding, and its

temperature, like that of any expanding gas, has been dropping. From
unimaginably high temperatures just after the big bang, the universe
has cooled to a very chilly average of four degrees above absolute
zero. If the universe continues its expansion, its temperature will
continue to fall, edging ever closer to absolute zero, a state where all

molecular motion would cease. This may not sound like progress,
but luckily for our superintelligent descendants, the energy required
to unambiguously send or record a signal also falls as temperature
drops. Molecules and radiation in the surroundings jostle less as they
cool, creating less background noise to be overcome. Therefore, the

147
148 Mind Children

energy required to do a computation is less at lower temperatures.

More and more thinking can be done with less and less power.
So here's the plan: Before it's too late (better hurry, there are only
some trillions of years left!) we take some of the remaining organized
energy in the universe and store it in a kind of battery. For the sake
of argument I imagine this battery to be a beam of photons bounced
back and forth between two mirrors, which in turn feel a pressure
from the light. Energy is extracted by allowing the light to push the
mirrors farther apart, like pistons in a car engine. The receding mirrors
will red-shift the light, slightly lowering its energy and increasing its
wavelength. The energy of the moving mirrors is used to power our
civilization. The idea is to use about half the energy in the battery

to do T amount of thinking, then to wait until the universe is cold

enough to permit half the remaining energy to support another T,
and so on indefinitely In this way a fixed amount of energy could
power an unlimited stretch of thought. As the machinery grows older
and colder, it becomes slower and larger as photons of ever-longer
wavelength do the work.
Whether the expansion of the universe will continue in this way
or eventually halt and reverse is a matter of debate and a debate of —
matter, in that gravity will be able to halt the expansion only if there is

enough total mass in the universe. But even if the universe turns out
to be fated for an eventual recompression, an inverse of the process
described above might be possible. Mirrors surrounding a stored
vacuum could derive increasing amounts of energy by shrinking
under the rising pressure of a collapsing cosmos. A subjective infinity
of thought might be done in the finite time to collapse by using this
growing power to think faster and faster as the end draws nigh. The
trick here is to repeatedly do an amount of thinking T in half the

remaining time. In an ever-expanding universe, time is cheap but

energy must be carefully husbanded. In a collapsing universe, energy
is cheap, but there is no time to waste! Both the expansion and the

compression scenarios exploit the size change of the universe as a

source of organized energy to counter heat death.
These suggestions are mere outlines for ideas that, at best, are
new and half-baked. In 1978 the physicist Freeman Dyson worked
out many details of a reach for immortality in an ever-expanding
universe and discusses them in his book Infinite in All Directions. The
astronomer John Barrow and the physicist Frank Tipler develop a
 —

Breakout 149

dramatic and encompassing kind of survival in a collapsing universe

in the last chapter of their book The Anthropic Cosmological Principle.

The Thinking Universe and Beyond

If our successors somehow manage to wrangle for themselves a
subjective infinity of time to think, will they eventually run out of
things to ponder? Will they be fated to repeat the same thoughts over
and over, in an endless and pointless cycle? At our present embryonic
stage of intellectual development, greater knowledge seems only to
expand our areas of ignorance. It is as if we were exploring a territory
—
from its interior as the mapped area enlarges, so does the length of its
perimeter. If the territory turned out to be finite, however, we would
eventually reach a point of diminishing frontiers. This is an unlikely
state of affairs because even if there turn out to be limits in some
areas of inquiry, for instance in space exploration, there seem to be
infinite hierarchies of ever more difficult and important mathematical
questions. Sometimes this kind of problem succumbs to a general
approach or algorithm that solves the entire hierarchy in one fell

swoop, but in other instances the problems must be solved one by

one, each harder than the last.

The mechanism of reasoning may itself be a fruitful line of inquiry.

The rules of inference by which conclusions are drawn from premises
have an apparent inevitability. Yet, viewed abstractly, they are but
rules for transforming strings of symbols into other strings. Entirely
different sets of transformation rules also produce consistent results.
It is at least possible that the way we reason is not a universal
absolute but merely an evolutionary expedient accidentally hit upon
by organisms unique Animals that thought more or less our
to earth.
way survived, while slightly different ways of thinking proved fatal.
But just as our intuitive grasp of physics does not encompass relativity
or quantum mechanics and is thus only an approximation applicable
within a narrow range of conditions, so may our reasoning processes
be fundamentally parochial and incomplete. Our perceptions of reaUty
are shaped by the inferences we draw, so new ways of reasoning may
effectively change our view of reality.

An eternity of pure cerebration, which may seem like heaven to an

academic, may be hell for the more actively inclined. Not to despair
doing as well as thinking will be an option within the machines of
250 Mind Children

the future. There will be worlds to explore and great engineering

projects to undertake. As a metaphor for the possibihties, consider
an invention of the mathematician and computer pioneer John von
Neumann.
Wishing to study the idea of self-replicating machinery without
having to deal with the messy details of real-world physics, von
Neumann devised a simple universe called a cellular automaton, in
which space was divided into an endless checkerboard of square cells.
Time advanced in discrete moments, and at any one moment each cell
was in one of 29 states. The state of a given cell in the next moment
depended simply on its current state and that of its four immediate
neighbors. The dependency was given by a "transition table" that
applied uniformly to every cell in the grid and listed the next state for
every combination of previous states.

By contriving a particularly helpful transition table, von Neumann

easily constructed "machines" in the grid that could carry out com-
mands to build other machines. Given the right instructions, they
could build copies of themselves. A typical machine consists of a
certain pattern of cell states (the machinery) in contact with another,
long pattern (the tape). A signal from the machinery causes the tape
pattern to march left or right one space in a wavelike fashion. The
machinery interprets the symbols on the tape as instructions that
control anarm protruding from one end of the machine. The arm
grows or shrinks by one square, bends left or right, or alters the state
of the cell at the end of its reach. A message might cause the arm
to sweep back and forth, shortening itself on each sweep and leaving
behind a desired "painting" made of quiescent states. When the end of
the tape is reached, it is commanded to rewind to its original position.
As it works its way backward, the machinery reads it a second time
and manufactures a copy of it in proximity to the painting. In a final

step a "breath of life" signal is transmitted to the painting that converts

its Depending on the tape message,
quiescent states to active ones.
the new machine might be a duplicate of the original, whichwould
then proceed to make another copy. From this model von Neumann
was able to prove that a cellular automaton that could contain general
self-replicating machines was universal, meaning it could be configured
to simulate (slowly) any other cellular automaton, or for that matter,
any other kind of computer. He was also able to show that in a
 Breakout 151

given universal cellular automaton a general constructor had to be

of a certain minimum size. About five years after von Neumann's
invention, Watson and Crick discovered that the DNA molecule acts
as the tape for a general constructor in the cells of living things.
Besides their theoretical importance, cellular automata turned out
to be fun. In 1969 John Horton Conway, a playful mathematician at

the University of Cambridge, invented an especially attractive one

that he called Life. It was presented in Martin Gardner's Scientific

American column (now collected in his book Wheels, Life and Other
Diversions) and sparked activity at scores of university computer cen-
ters. The Life automaton tended to produce certain easily recognizable

patterns, and these were rapidly given names: "blocks," "loaves," and
"beehives" are stable; "blinkers" flip back and forth between a short
horizontal line and a vertical one; "gliders" go through a series of
four contortions, ending up displaced diagonally one space, poised
to do it again; larger "spaceships" travel twice as far as ghders but
purely horizontally or vertically; the "R pentomino" starts out tiny
but grows to a writhing mass that peters out after 1,500 time-steps to
a collection of blocks, loaves, beehives, and blinkers, having shot off
five gliders.
Conway did not construct Life to embody von Neumann's goal of a
self-reproducing machine. Rather, Conway conjectured that Life was
not universal; specifically, he suspected that any finite pattern, though
it might grow innumber of active cells for a while, would eventually
exhaust itself, thus making replicators impossible. An especially
vigorous group of Life hackers at the MIT Artificial Intelligence lab
disproved by constructing patterns called "ghder guns"
this conjecture

that slowly oscillated and expelled a new glider at the end of each
long cycle, so producing an endless stream of gliders. Then they built
"puffer trains" that traveled while their patterns cvcled and which
left behind regular puffs of debris. Ultimately they were able to
combine these approaches into a large pattern that chugged along
like a puffer train, but whose puffs turned into glider guns that
immediately began to issue a stream of gliders. After a time this
pattern produces a wedge of space filled with gliders. In these studies
the group developed methods components of
for constructing all the
a von Neumann replicator in the Life space, though no one has yet
built such a huge machine in toto.
252 Mind Children

Newway and the Cellticks

Imagine now a huge Life simulation running on an enormously large

and fast computer, watched over by its programmer, Newway. The
Life space was seeded with a random pattern that immediately began
to writhe and froth. Most of the activity is uneventful, but here and
there small, growing, crystalline patterns emerge. Their expanding
edges sometimes encounter debris or other replicators and become
modified. Usually the ability to spread is inhibited or destroyed in
these encounters, but once in a while there emerges a more complex
replicating pattern, better able to defend itself. Generation upon
generation of this competition gradually produces elaborate entities
that can be considered truly alive. After many further adventures,
intelligence emerges among the
Life inhabitants and begins to wonder
about its and purpose. The cellular intelligences (let's call them
origin
the Cellticks) deduce the cellular nature and the simple transition rule
governing their space and its finite extent. They realize that each tick
of time destroys some of the original diversity in their space and that
gradually their whole universe will run down.
The Cellticks begin desperate, universe-wide research to find a way
to evade what seems like their inevitable demise. They consider the
possibility that their universe is part of a larger one, which might
extend their life expectancy. They ponder the transition rules of their
own space, its extent, and the remnants of the initial pattern, and find

too little information to draw many conclusions about a larger world.

One of their subtle physics experiments, however, begins to pay off.

Once in a long while the transition rules are violated, and a cell that

should be on goes off, or vice versa. (Newway curses an intermittently

flashing bulk-memory error indicator, a sign of overheating. It's time
to clean the fan filters again.) After recording many such violations,

the Cellticks detect correlations between distant regions and theorize

that these places may be close together in a larger universe.
Upon completing a heroic theoretical analysis of the correlations,
they manage to build a partial map of Newway's computer, includ-
ing the program controlling their universe. Decoding the machine
language, they note that it contains commands made up of long se-
quences translated on the screen similar to the cell patterns
to patterns
in their universe. They guess that these are messages to an intelligent

operator. From the messages and their context they manage to decode
 Breakout 153

a bit of the operator's language. Taking a gamble, and after many false
starts, the Cellticks undertake an immense construction project. On
Newway's screen, in the dense clutter of the Life display, a region of

cells is form the pattern, slowly growing in size: LIFE

manipulated to
PROGRAM BY J. NEWWAY HERE. PLEASE SEND MAIL.
A bemused Newway notices the expanding text and makes a
cursory check to rule out a prank. This is followed by a burst of
hacking to install a program patch that permits the cell states in the
Life space to be modified from keyboard typing. Soon there is a dialog
between Newway and the Cellticks. They improve their mastery of
Newway's language and tell their story. A The
friendship develops.
Cellticks explain that they have mastered the art of moving themselves
from machine to machine, translating their program as required. They
offer to translate themselves into the machine language of Newway's
computer, thus greatly speeding their thoughts. Newway concurs.
The translation is done, and the Celltick program begins to run. The
Life simulation is now redundant and is stopped. The Cellticks have
precipitated, and survived, the end of their universe. The dialog
continues with a new vigor. Newway tells about work and life in the
larger world. This soon becomes tedious, and the Cellticks suggest
that sensors might be usefulto gain information about the world
directly. Microphones and television cameras are connected to the
computer, and the Cellticks begin to listen and look. After a while the
fixed view becomes boring, and the Cellticks ask that their sensors and
computer be mounted on a mobile platform, allowing them to travel.
This done, they become first-class inhabitants of the large universe, as
well as graduates of the smaller one. Successful in transcending one
universe, they are emboldened to try again. They plan with Newway
an immense project to explore the larger universe, to determine its
nature, and to find any exit routes it may conceal. This second great
escape will begin, as the first, with a universe-wide colonization and
information-gathering program.
At this stage of our development we have hardly a clue as to the
nature and purpose of our universe. Physical theories like relativity
and quantum mechanics, and the particle theories and cosmologies
woven from them, are the most powerful methods now available
for fathoming reality far beyond our experience. But there is no
reason to be confident that these theories are more reliable beyond
the limits within which they have been experimentally tested than
154 Mind Children

Newtonian n^echanics is at describing objects moving near the speed

of light. But though incomplete and rooted in pedestrian laboratory
measurements, our theories already hint at universes beyond the 40-
billion-light-year-diameter sphere of stars thatwe perceive when we
gaze skyward. Quantum mechanics makes highly accurate predictions
about the outcomes of experiments by summing up the effects of an
infinity of possible ways the unobserved parts of the experiment may
behave. In one successful interpretation of quantum mechanics, these
alternativeshappen in an infinity of parallel worlds, each equally real.
I discuss some implications of this idea in Appendix 3. Strange brews

of general relativity and quantum mechanics are required to think

about the universe when it was very dense and very hot. Some of
these describe a universe that repeatedly collapses and expands, each
cycle producing a new world with a unique arrangement of matter
and energy and even physical laws. Other concoctions describe a
super universe in which our own 40-billion-light-year sphere is but
a bubble, like a tiny expanding pocket of steam in a boiling liquid
containing many, many others. Obviously we have yet much to learn.
A recent remarkable development in Life programs hints at how
subtle the problem of figuring out a universe from the inside may be.
It concerns the nature of space, time, and reality.

HashLife

The MIT group that showed Life to be universal worked with a clever
and efficient simulation program. The ease and speed with which
they could examine the evolution of Life patterns was one of their
advantages over other communities of Life hackers. Instead of simply
mapping an entire Life grid into an array of bits in the computer's
memory, the MIT program stored a large space as small patches and
simply skipped empty regions. The computation to advance each
—
patch to its next state depended on the pattern patches holding
common predictable patterns like blocks or gliders were done by swift
looks in a table. Only in uncommon or complex areas did the program
resort to the laborious application of the transition rules. It worked
quite well, as witnessed by the group's discoveries. Yet there was
an annoying sense of things undone. The entries in the fast-update
table were all specified in advance, by hand. What if some important
patterns had been overlooked? Could a program be devised that
 Breakout 155

learned such things from its own experience? In 1982, a decade after
the Life-hacking at MIT had ceased. Bill Gosper, the premier theorist
of the group, now in California, devised a solution.
The state of a Life cell depends only on its own state and that of
its immediate neighbors at the last time instant. Thus patterns creep

over the surface no faster than one cell per instant, a velocity called
the speed of light. The future of the interior of a large square portion of
a larger Life space can be predicted up to a certain time simply from
its past The predictable area shrinks with time as the pattern is
state.

corrupted by information creeping inward from the edges at the speed

of Hght. If the two-dimensional Life space is plotted horizontally,
and if successive instants are stacked vertically, the predictable region
forms a pvramid with the original square portion at its base, as in the

figure on page 156. Gosp>er's method depends on slicing this pyramid

in two. The pattern on the large bottom square is used to predict the
half-sized square at the cut.
Square Life patterns are associated with unique numbers called
hash addresses (hashing is an old and effective computer technique
for turning long and complicated items like names into relatively
small numbers, so entries can be stored and looked up rapidly in

a table). The hash number for a given square is found by cutting it

into four smaller squares and combining the hash numbers of each
with a certain formula. This subdivision stops when the squares get
verv small (4 cells which point the basic pattern itself
on a side), at
gives the number Gosper's program keeps a table with an entry for
each hash number Each entr\' is itself five hash numbers, one for
each of the four smaller squares making up the pyramid's base, and
one for the "answer plateau." Whenever a pattern is encountered the
second or subsequent time in a Life simulation, its answer is simply
looked up in the table. Even when an entry is not in the table, it can be
built quickly if partial answers are known, as illustrated in the figure
on page 157. As more and more configurations are encountered and
stored, the program can go faster and faster, effectively taking bigger
and bigger steps.
Some interesting issues arose when the program was first tried on a
computer The effectiveness of the method was clear: in a typical run,
the first 100 ticks of the Life simulation might take as long as a minute
ofcomputer time, the next 1,000 ticks could happen in ten seconds,
and the next 1,000,000 might consume only a second. But how does
156 Mind Children

one display such an accelerating simulation? Given a hash number

that encodes an answer (that is, the future of an initial pattern), its
subparts, sub-subparts, and so on can be found by tracing through

Spacetime Pyramid
A "spacetime" diagram way of presenting the evolution of a
is one
cellular automaton (or any other physical system). Here, the initial
state of a Life world is represented by the square base of the pyramid.
Successive layers on this base represent the world at successive times.
The state of a cell in Life depends only on its own state and the
states of its immediate neighbors at the last instant. A large square

of Life cells completely determines its own next state except for an
outer boundary one cell thick. If we trim away these corrupted outer
cells, what remains is a smaller square, that again fully predicts a
still smaller square one time-step later. If we continue this way, the
ever-smaller squares form a pyramid of spacetime. Every cell in the
pyramid is an indirect consequence of the pattern at the pyramid's

base. Gosper's hashlife program stores away the half-size square

halfway up a pyramid, to avoid having to recompute the intertnediate

steps when a given base pattern is encountered more than once.

Pyramid of predictability

Gosper's
answer plateau'

/I » V
Time

Initial configuration
of Life array

Space
 Breakout 157

the hash table. A complete picture of the Life pattern can thus be
built. Since only the portion to be displayed must be constructed, the
program can handle extremely large spaces. Gosper often simulated
Life universes a billion cells on a side!
But what if one wants to see the calculation in progress, as in
the Newway story? At first Gosper tried simply displaying the
partial answers as they were computed. The results were bizarre.
The program advances the simulated time in different portions of
the space at different rates. Sometimes it even retreats in places,
because some regions are described by more than one pyramid,
and the different pyramids are not computed at the same times.
A single glider advancing across the screen would cause a display
where gliders would appear and disappear in odd places almost at
random, sometimes several in view, sometimes none. Constraining
the program so it never reversed time in any displayed cell improved
things only slightly.

Large Spacetime Pyramids from Small

Two layers of smallpyramids can, with considerable overlap, be used
to construct a pyramid twice as large. The first layer has nine small
pyramids, the second has four, for a total of thirteen. In this way
answers for small regions of space and short times can be assembled
by the hashlife program into solutions for large spaces and times.

Answer plateaus of four small pyramids are

fused to form answer plateau of large pyramid

Answer plateaus of first layer

become initial configurations
for second layer
Four small pyramids
overlap on second layer

Nine small pyramids

overlap on first layer
158 Mind Children

The best solution turned out to be not to display at all until the
calculation was finished. The pattern might start out a billion cells

on a side (necessarily mostly and its future would

empty space!)
be calculated for a half-billion time-steps. The full history of the
calculation would end up compactly encoded in the hash table. A
separate program could then invoke the table entries to view the
universe at any given time. The viewing program allowed Gosper
to scan, godlike, forward and backward in time through the evolving
Life pattern. Where did that glider come from? Here it is at time 100,000.

It wasn't there at 50,000. Nor at 75,000. Aha! This collision just before
80,000 generated it. Let's look at that step by step...

But what if Cellticks were to evolve in a hashlife space? By encoding

their universe and its evolution in such an efficient way, Gosper has
played them a dirty trick. What they perceive as the steady flow of
time for the most part does not exist. The hashlife program skips
over large chunks of spacetime without going through all the tedious
intermediate steps. The Cellticks may have memories of things that
never actually happened, though they were mathematically implied
from their past.
Figuring out about, and affecting, the larger world would be much
harder in the hashlife universe than in the simple Newway scenario.
Harder, but not impossible. The human race owes its present success
to the many small problems solved during our development, first

by biological and recently by cultural evolution. To build on these

successes and, like the Cellticks, exceed our universe, we will surely
have overcome much more difficult problems. What better way
to
to meet the challenge than by improving our minds using the most
powerful means that come to hand!

The Road Ahead

We are at the start of something quite new in the scheme of things.
Until now we have been shaped by the invisible hand of Darwinian
evolution, a powerful process that learns from the past but is blind to
the future. Perhaps by accident, it has engineered us into a position
where we can supply just a little of the vision it lacks. We can choose
goals for ourselves and steadfastly pursue them, absorbing losses in
the short term for greater benefits further ahead. We see the road
before us only dimly — it hides difficulties, surprises, and rewards
 Breakout 159

far beyond our imaginings. Somewhere in the distance there are

mountains that may be difficult to cHmb but from whose summit the
view mav be clearer. In the metaphor of Richard Dawkins, we are the
handiwork of a blind watchmaker. But we have now acquired partial
sight and can, if we choose, use our vision to guide the watchmaker's
hand. In this book I have argued for the goal of nudging that hand
in the direction of further improved vision. New worlds may then

reveal themselves, to our vision and to our reach.

Appendixes

Bibliography

Acknowledgments &
Illustration Credits

Index
 Al Retinas and
Computers

circuitrv
T HE discussion in Chapter 2 comparing neural
with computer calculations makes many assumptions and
may have raised some questions. It is difficult enough to compare
different electronic computers, let alone such fundamentally dissimilar
systems.

Justhow representative of the whole brain are the structures in the retina?
As suggested in Chapter 2, the size constraints and great survival
value of the retinal circuitrv have probably made it more efficient
than the average brain assembly. In efficiency, it mav be similar to
the wiring found in animals with small nervous svstems, which have
been mapped in recent years, where each neuron seems to play an
important role. A lot of evolutionary design time has been available
to get the most out of a relatively small number of neural connections.
The larger, newer structures in the human brain are likelv to use
their neurons less effectively, on the average. The same considerations
can apply to artificial intelligences. Small subsystems can be highly
optimized, but larger, less structured processes may have to loaf along
with more fat; there simply is not time to optimize huge pieces of
program so well.

Your analyses are based on a partial undcrstafuiijig of the main kinds of

retinal neurons. But other types are found occasionally. Also, the neurons
respond to chemical messages from several sources. Don't these extra effects
throw off your calculations? Rare connections are probably important,
but because they are few in number, their effects add little to the
computational quantity. In a similar way, broadcast chemical messages
are slow and contain only a relatively small amount of information. In
a program their effect can probably be mimicked by a modest number
of global variables that are referenced by other computations.

163
264 Mind Children

It has happened many times in computer science that a mathematical

discovery has reduced astronomically the amount of computation required
to get a certain answer. Could operations in the brain be candidates for

such improvements, designed as they are by a process that is unable to

perform large-scale restructurings? Maybe so. My retinal calculation

already benefits from a modest gain of this type (see below). Still,

some computations are not greatly reducible, and we have no sure

way of finding optimizations for all those that are. If we manage
to collapse half of what goes on in the brain to almost nothing, we
are still left with the other half, and the conversion ratio changes
insignificantly. Only if almost every process could be shrunk would
the effect seriously accelerate my predictions.

In your ratio, 10'^ calculations per second does the job of about 20" neurons.
This budgets only 100 calculations per second for each neuron. Surely this
is an underestimate. Many neurons integrate thousands of inputs and can
respond in hundredths of a second. It would be an underestimate if

we were attempting to simulate the brain by simulating each of its

neurons. But the computer can be used more efficiently when single
optimized programs do the functions of large groups of neurons. For
example, consider a retinal horizontal It makes thousands of
cell.

connections to the photocells in a large and computes the average

field

brightness of the field. The analogous job for a robot might be done
by a computer program that adds together thousands of pixels from
the robot's TV camera. If done for every horizontal cell, that would
be a lot of adding. But there is a way to avoid most of the effort.
The following idea works well on a two-dimensional image, but I will
present it in one dimension because it is easier to explain.
Let's say we have a million photocells all in a row, with a horizontal
cell connected to every adjacent group of 1,000 of them. Thus, the
first horizontal cell attaches to photocells 1 through 1,000, the second
covers 2 through 1,001, and so on, making a total of slightly fewer
than a million horizontal cells. Each horizontal computes the
cell

average brightness of its field. A naively written program would

do 1,000 additions for each of the million horizontal cells. A clever
program would exploit the fact that the sum produced by the second
horizontal cell differs from that of the first only in that it includes
photocell 1,001 and that it excludes photocell 1. Thus the second
sum can be calculated from the first one by merely subtracting the
 Retinas and Computers 165

first photocell value and adding the 1,001st. Similarly, the third sum
is calculated from the secondby subtracting photocell 2 and adding
photocell 1,002. And so on. Instead of a thousand calculations, each
additional horizontal cell costs only two. This technique works for
computers, but it cannot be exploited in a nervous system design
for two reasons. Since each sum depends on the one before it,

the finalsum depends on a chain of almost a million steps. At a

minimum of a thousandth-of-a-second delay per neuron, the rightmost
cell would not respond correctly to a change in the input for many

minutes! Even then the answer would be wrong because small errors
in the sum would rapidly accumulate down the long chain. On the
computer, however, the technique works very well because each step
takes only about a microsecond, and the arithmetic is done entirely
without error. Our robot vision programs are filled with shortcuts of
this kind, which exploit the great speed and precision of computer
operations. Nervous systems, on the other hand, are filled with rich,
overlapping interconnections, exploiting the power of self-replicating
genetic construction machinery.

Maybe those tricks work for the retina, but there is no guarantee that they
will work for all the diverse structures in the brain. It is possible that
some parts of the brain use their neurons so cleverly that a computer
program cannot do better than to simulate individual neurons and
synapses, but it is unlikely. The retina example illustrates two general
principles. The first is that the slow switching speed and limited
signaling accuracy of neurons rules out certain solutions for neural cir-

cuitry that are easy for computers. Second, a smooth function applied
repeatedly on overlapping inputs can be decomposed into subparts in
such a way that the subparts can be used more than once. Many
neural structures in the brain involve regular cross connections of
many inputs to many outputs, making them candidates for this kind of
economy. The human cerebral cortex, one of the largest structures, is
a crumpled disk about 2 millimeters thick and 20 centimeters in diam-
eter, containing 10 billion neurons arranged into a half-dozen layers,
wired quite repetitively. The well-studied one or two percent of this
sheet that handles vision carries further the processing begun in the
retinas, apparently using similar methods. Edges and motion in differ-
ent directions are picked out cleanly in the first few layers, and those
feed layers that respond to more complex patterns such as corners.
166 Mind Children

Some regularity is to be expected in the nervous system because

there is insufficient information in the 10^° bits of the human genome
to custom-wire many of the 10^'* synapses in the brain. Interestingly,

thisargument may not hold for small nervous systems such as that of
the much-studied sea slug Aplysia, which has about 100,000 neurons
clumped into 100 ganglia. Several of the ganglia have been mapped,
and the neurons and their interconnections seem to be exactly the
same from animal to animal, with each junction playing a unique
and important role in the animal's behavior. It is plausible that the
few billion bits in Aplysia's genetic code contain special instructions
for wiring each of its several million synapses. If it should turn out
that a direct neural simulation is necessary for particularly irreducible
parts of the vertebrate brain, it would still be possible to stay on
my time track. A general-purpose computer suffers a thousandfold
handicap over my retinal conversion number if forced to simulate
individual neurons. The speed could be regained, however, at the cost
of flexibility, by building special-purpose neuron-simulating machines
using about the same amount of circuitry as the general-purpose
computer. I'm betting it won't be necessary.

There's something asymmetrical about equating the retina with a computer.

Doesn't a computer programmed to emulate a neural circuit have considerable
potential not found in the neural arrangement itself? The computer is, after

all, general-purpose and can be reprogrammed for radically different tasks.

The retina is forever stuck with doing its one computation. The difference
isonly one of convenience and speed of reprogramming. The retina
can and has been reprogrammed many times during the course of
our evolution. The computation done by the retina of a particular
individual organism is fixed in the same sense that the computation
of a running computer is fixed by the program it happens to contain at
that moment. The set of all possible programs that the computer may
contain is analogous to the set of all possible ways a given amount
of neural tissue could be connected. Evolution selected a certain
configuration of neuron properties and interconnections within one
set in the same sense that our research is selecting a certain program.

Of course there is a difference in programming time. The neural

configuration is controlled by genetic instructions, and each change
requires the growth of a new organism, a process that may take years.
 Retinas and Computers 167

A corresponding change can often be tested on a computer in minutes,

about a million times as fast.

This is a great advantage, and one reason I believe that special-

purpose computers will find a major place in robotics only after the
basic research is almost complete. A special-purpose computer is an
arrangement of arithmetic, memory, and control circuitr\' optimally
configured to do one particular task. A special-purpose computer
can be as much as one thousand times faster than a general-purpose
machine of similar and cost doing the same task. But it takes
size
about as long to design and build a specialized machine as it takes to
grow a new organism. So, I think the research will happen on general-
purpose machines, but once the requirements for human equivalence
are well understood, it will be possible to build specialized thinking

machines that are much cheaper than my projections. On the other

hand, I also think that the self-improvement possibilities inherent in a
general machine will be too valuable to give up. A mature, intelligent
robot will probably contain some special machinery supporting a
general-purpose superstructure.

Aren't neurons, the product of a billion years of ei^olution, highly complex

and optimized devices that we are unlikely to improve upon? No. First,
much of the neuron's mechanism has to do with growing and building
itself from inside out. Present and foreseeable computer components
dispense with this baggage bv being constructed from the outside.
This is a huge advantage — all the structure can be used for con-
and thought. Second, the neuron's basic
trolling perception, action,

—
information-passing mechanism the release of chemicals that affect
the outer membranes of other cells —
seems to be a very primitive one
that can be observed in even the simplest free-swimming bacteria.
Animals seem to be stuck with this arrangement because of limitations
in their design process. Darwinian evolution is a relentless optimizer
of a given design, nudging the parameters this way and that, adding
a step here, removing one there, in a plodding, tinkering, way. It's
not much of a redesigner, however. Fundamental changes at the
foundation of its creations are out of reach, because too many things
would have to be changed correctly all at once. By contrast, human
designers are quite good at keeping the general shape of an idea, while
changing all its parts. Calculators were once built of cogs and levers.
168 Mind Children

then of relays, then of vacuum tubes, then transistors and integrated

circuits. Soon Hght or supercurrents may flow in their wiring. In
all this time the fundamental operations carried out were much the
same, and design principles and software developed for one type of
hardware are usually easily transferred to the next.

What assumptions went into the placement of the animal nervous systems

in the figure on page 61 in Chapter 2? The data are listed in Table 1.

Table 1. Nervous Systems

Animal Brain mass Neurons Power Capacity

grams bits/sec bits

Snail 10' 108 10«

Bee 10^ 10' 10'

Hummingbird 0.1 107 10'° 10'"

Mouse 1 10« 10" 10"

Human vision 100 lO'o 10" 10"
Human 1,500 10" 10^4 10'*

Elephant 3,000 2x10" 2x10^^ 2x10'*

Sperm whale 5,000 5x10" 5x10'^ 5x10'-'
 A2 Measuring
Computer Power

B,EC A USE of its effect on computer sales, compar-

ing the relative power and cost-effectiveness of different computers
has always been a contentious affair. But the range of disagreement
among manufacturers about the power of one another's machines is
usually less than a factor of ten, and such a small ratio does not

materially affect the appearances of the diagrams nor the conclusions

of Chapter 2, where scales of a trillion are exhibited. Yet any particular
formula for estimating power may be grossly misled by an unlucky
or diabolic counterexample. For instance, if a computer's power were
defined simply by how many additions per second it could do, an
otherwise useless special circuit made of an array of fast adders,
and nothing else, costing a few hundred dollars, could outperform a
$10-million supercomputer. My intuition about useful computation
has suggested a trickier but, I believe, safer measure. Be assured that
for reasonable machines my formulas give almost exactly the same
numbers for processing power as simpler approaches.

Things that compute massively can alter their internal variables, and
their outputs, in unexpected ways. We can say a stationary rock, or
even a rolling one, or the adder array described in the last paragraph,
does littlecomputing because it is so predictable, while a mouse
scurrying in a maze must be doing quite a bit. Claude Shannon's
information theory is built on a way of quantifying the amount of
surprise, or information, in a message. The more unexpected the next

piece of message, the greater the amount of information it contains. 1

will use this approach to measure the information in a computation.

Computing power will be defined as the amount of information,
or surprise, exhibited per second as a machine runs, that is, as it

repeatedly changes from one internal state to another. The more

unexpected the next state of the machine, the greater the amount of

169
270 Mind Children

information contributed by the transition to that state. Quantitatively,

if to the best of our knowledge there is a probability p that the machine

will go into a certain state, then if it does go into that state, the
transition will have conveyed -logjp bits of information (log2 means
logarithm to the base 2, and bits are binary digits). If the probabilityp
was 1/2, the transition will convey just 1 bit of information. If p was
1/2", the n bits will have been conveyed. A p of 1/1,000 gives about
10 bits of information.
The average information conveyed in a transition is found by
multiplying the information conveyed in a transition to each possible
next state by the probabihty of going to that state, and then summing
over all the possibilities:

N
Information per transition = 2^ - Pi logj p, hits
i= 1

where N is the number of possible states and p, is the probability

that the ith state will be the next one. The computational value of a
given transition can be different for different observers because they
assign different probabilities to the outcomes. The information reaches
a maximum of log2 N in a totally ignorant observer, for whom all the
p, are equal.At the other extreme, an all-knowing witness can be
certain that the next state will be ;, and thus let p, = \ and and all
the other p, = 0, making the information 0. A machine does a useful
computation for you only if you don't already know all the answers
in advance!
Computing requires long sequences of these kinds of transitions
from one state to another. The total information capacity of a system
is log2of all the states it can ultimately reach. In a general-purpose
computer this is simply the total memory size. A powerful machine
is able to step through states quickly. I measure processing power by
dividing the transition information by the average time required for a
transition. This gives us the formula

I Pit,

The units are bits per second. This measure also is reduced by
predictability.
 Measuring Computer Power 171

The formulas capture a number of ideas. A computer endlessly

repeating a program loop becomes totally predictable, and its comput-
ing power drops to zero. Programs written in high-level languages or
using interactive environments often run much more slowly than if
written directly in the computer's machine language. High-level lan-
guage constructs are converted to stereotyped sequences of machine
instructions, making the program more predictable than one written
in pure machine language, thus lowering the effective processing

power. Adding memory to a computer modestly increases its power

even without any increase in its raw speed. Among the techniques
for using memory to enhance computation are tables of previously
computed results and reorganizations of data structures that take up
more space but are faster to access. This effect, of memory increasing
computational power, appears in my measure because in a computer
instruction the identity of the memory location referenced is as much
a surprise as which operation is to be performed. As the number of
possible locations increases, so does the amount of surprise, though
modestly. Doubling the memory increases the power by only one bit

per instruction time.

In highly parallel machines, especially those using a single instruc-
tion stream to control numerous processing units, most of the surprise
is in the parallel data, not the instructions. Although the total number
of bits flowing in the data streams represents an upper bound to the
processing power of this kind of machine, the real power may be
considerably less because of intrinsic redundancies or predictabilities.
Estimating these in disparate machines designed for finite element
analyses, symbolic processing, cellular automata, or playing chess is

difficult. The on page 64 is not much affected by this difficulty,

figure
because almost all the machines on it are of the conventional von
Neumann type, where only one datum is processed per instruction.
Future versions of the graph may have to deal with massively parallel
machines, some of which are just now being tested with real problems.
Even with conventional architectures, the power measurements get
messy when the formulas are applied to real computers. How can
probabilities be assigned to different instruction types when each
kind of program exhibits its own statistics? In computers with large
instruction sets, many operations are almost never used. Besides this,

detailed descriptions are hard to come by for many old machines. My

compromise is to assume every machine uses 32 distinct operations
 —
171 Mind Children

(six bits worth) that are mixed in equal proportion. If each memory
location is equally likely to be addressed in an instruction, then the
information it contributes is equal to the logarithm of the memory
size. an upper bound. Since the contents of memory locations
This is

themselves can change, the data stored there are also a source of
surprise, but only if the data are read rather than being simply
overwritten. If we assume that half the instructions read data, this
channel contributes a maximum of half of a word size of information.
In a parallel machine controlled by a single instruction stream, the
aggregate word size of the parallel data streams would be considered,
and this would be the major component of total information.
Another issue is timing. Once again, obtaining great detail is
difficult. Two readily available numbers, however, are the average

times to do an addition and to multiply. Addition is typical of the

fastest computer operation, while multiplication is slow. 1 assume

that the instruction mix contains seven instructions taking as long as

an add for every one taking as long as a multiply.
With these approximations the power formula becomes

memory + word /2
Power = —+
6 logT

(7 X
1,
+
-^ -—
T^ multiply If O
Tadd

where memory is the capacity of the machine's fast memory in indi-

vidually addressable words, and word is the size of a data word in

bits. The capacity of the machine is found by multiplying memory by
word. For decimal machines, the number is approximated by
of bits
multiplying the number of decimal digits by four. This formula was
used to plot the points in the figure on page 64, and the data for that
figure is listed in Table 2.

A Nautical Metaphor

I have defined two qualities essential for interesting computation

or thought. They are computational power and capacity. Basically,
power is the speed of the machine, and capacity is its memory size.
Computing can be compared with a sea voyage in a motorboat. How
fast a journey can be completed depends on the power of the boat's

engine. The maximum length of any journey is limited by the capacity

of the boat's fuel tank. The effective speed is decreased, in general.
 Measuring Computer Power 173

if the course of the boat is constrained, for instance if the boat must
sail due east/west or north/south instead of being able to make a
beeline to its destination. Some computations are like a trip to a
known location on mapless search
a distant shore; others resemble a
for a lost island. computing is like having a fleet of small
Parallel
boats: it helps in searches and in reaching multiple goals, but may
not help very much in solving problems that require a sprint to a
distant goal. Special-purpose machines trade a larger engine for less
rudder control. Attaching disks and tapes to a computer is like adding
secondary fuel tanks to the boat. The capacity, and thus the range, is
increased, but if the connecting plumbing is too thin, it will limit the
fuel flow rate and thus the effective power of the engine. Input/output
devices are like boat sails. They capture power and capacity in the
environment. Outside information is a source of variability and thus
power, by our definition. More concretely, it may contain answers
that would otherwise have to be computed. The external medium can
also function as extra memory, increasing capacity.

Table 2. Calculating Machines, by Year

Cost Memory Word "^add ^mult Power Capacity Power/cost

1988$ words bits sec sec bits/sec bits b/s/$

Human
1x10^ 2x10' 40 6x10' 6x102 2x10-' 8x102 2x10-6

1891 — Ohdner (mechanical)

1x105 6x10-2 20 1x102 6x102 7x10-2 1x10" 5x10-7

1900 — Steiger Millionaire (mechanical)

IxlO' 1x10-' 24 5x10' 1X102 3x10' 3x10" 2x10-6

1908 —
Hollerith Tabulator (mechanical
5x105 8x10' 30 5x10' 2x102 4x10' 2x103 7x10-7

1910 —
Analytical Engine (mechanical)
9x10^ 1x10' 200 9x10'^ 6x10' 8x10" 2x105 8x10-7

1911 — Monroe Calculator (mechanical)

4x105 1x10" 24 3x10' 1x102 4x10' 2x10' 1x10-6

1919 — IBM Tabulator (mechanical)

1x105 5x10" 40 5x10" 2x102 8x10' 2x102 9x10-6

continued
174 Mind Children

Table 2 (cont.)

Cost Memory Word T^j^ Tj^^j^ Power Capacity Power/cost

1988$ words bits sec sec bits/sec bits b/s/$

1920 —Torres Arithmometer (relay)

1x105 2x10° 20 1x10^ 1x10^ 7x10-1 4x10^ 7x10-^

I92S — National-Ellis 3000 (mechanical)

1x10^ 1x10° 36 1x10' 6x10' 1x10° 4x10' IxlO'^

1929 — Burroughs Class 16 (mechanical)

1x10^ 1x10° 36 1x10' 6x10' 1x10° 4x10' IxlO-^

1938 —Zuse-1 (mechanical)

9x10* 2x10' 16 1x10' 1x10^ 8x10"' 3x10^ IxlO'^

1939 —Zuse-2 (relay & mechanical)

9x10* 2x10' 16 1x10° 1x10' 8x10° 3x10^ IxlO"*

1939 —BTL Model 1 (relay)

4x105 4x10° 8 3x10' 3x10' 4x10' 3x10' 9x10-5

1941 — Zuse-3 (relay & mechanical)

4x105 6x10' 32 5x10' 2x10° 4x10' 2x10^ 1x10"^

1943 —BTL Model 2 (relay)

3x105 5x10° 20 3x10-' 5x10° 2x10' 1x10^ 6x10-5

2943 — Colossus (vacuum tube)

6x105 2x10° 10 2x10^ 2x10-2 4x10^ 2x10' 7x10-3

1943 — BTL Model 3 (relay)

1x10^ 2x10' 24 3x10-' 1x10° 6x10' 4x10^ 4x10-5

1944 — ASCC (Mark 1) (relay)

2x10^ 7x10' 70 3x10' 6x10° 5x10' 5x10^ 2x10-5

1945 — Zuse-4 (relay)

3x105 6x10' 32 5x10' 2x10° 4x10' 2x10^ 1x10"^

2946 —BTL Model 5 (relay)

3x10^ 4x10' 28 3x10' 1x10° 7x10' 1x10^ 2x10-5

2946 — ENIAC (vacuum tube)

3x10^ 2x10' 40 2x10-^ 3x10-^ 6x10* 8x10^ 2x10-^

2947 — Harvard Mark 2 (relay)

1x10^ 1x102 40 2x10' 7x10' 1x10^ 4x10^ 9x10-5

2948 — IBM SSEC (vacuum tube & relay)

2x10^ 8x10° 48 3x10^ 2x10-^ 1x10" 4x10^ 6x10-^

continued
 Measuring Computer Power 175

Table 2 (cont.)

Cost Memory Word T^^^ ^mult Power Capacity Power/cost

1985$ words bits sec sec bits/sec bits b/s/$

1949 — EDSAC (vacuum tube)

4x105 5x102 35 3x10^ 3x10-3 5xW 2x10-* 1x10-'

1950 — SEAC (vacuum tube)

3x10^ 1x103 45 2x10"* 2x10-3 8x10-* 5x10" 2x10-2

1951 — UNIVAC I (vacuum tube)

4x10^ 1x10^ 44 1x10-* 2x10-3 1x105 4x104 3x10-2

1952 — Zuse-5 (relay)

4x1 0^ 6x10' 32 1x10-' 5x10-' 2x102 2x103 5x10^

1952 — IBM CPC (vacuum tube & relay)

4x105 9x10° 144 SxlO"* 1x10-^ 4x10* 1x103 9x10-2

1953 — IBM 650 (vacuum tube)

8x105 1x103 40 7x10-^ 1x10-2 2x10" 4x104 2x10-2

1954 — EDVAC (vacuum tube)

2x10^ 1x10^ 44 9x10"* 3x10-3 3x1 0^ 5x104 2x10-2

1955 — Whirlwind (vacuum tube)

8x105 2x103 16 2x10-5 3x10-5 1x10^ 3x104 2x10°

1955 — Librascope LGP-30 (vacuum tube)

1x105 4x103 30 3x10-^ 2x10-2 1x10" 1x105 1x10-'

1955 — IBM 704 (vacuum tube)

8x10^ 8x103 36 1x10-5 2x10-4 1x10^ 3x105 1x10-'

1959 — IBM 7090 (transistor)

1x10^ 3x10^ 36 4x10-^ 2x10-5 7x10^ 1x10* 6x10'

1960 — IBM 1620 (transistor)

7x105 2x10^ 5 6x10-4 5x10-3 2x104 1x105 3x10-2

I960 — DEC PDP-1 (transistor)

5x105 8x103 18 1x10-5 2x10-5 2x1 0^ 1X105 5x10°

1961 —Atlas (transistor)

2xW 4x103 48 1x10-^ 5x10-^ 3x1 0^ 2x105 2x10"

1962 — Burroughs 5000 (transistor)

4x10^ 2x10^ 13 1x10-5 4x10-5 2x10^ 2x105 5x10'

2964 — DEC PDP-6 (transistor)

1x10^ 2x10'' 36 1x10-5 2x10-5 3x1 0^ 6x105 3x10°

continued
176 Mind Children

Table 2 (cont.)

Cost Memory Word *add ^muit

Power Capacity Power/cost
1988$ words bits sec sec bits/sec bits b/s/$

1964 — CDC 6600 (transistor)

2x107 5x105 64 3x10-7 5x10- 2x108 3x107 1x10'

1965 — IBM 1130 (hybrid chip)

4x105 8x1 0^ 16 8x10-* 4x10-5 2x10* 1x105 6x10°

1966 — IBM 360/75 (hybrid chip)

2x107 2x10* 32 8x10-7 2x10-* 5x107 6x107 3x10°

1967 — IBM 360/65 (hybrid chip)

1x10^ 1x10* 32 2x10-* 3x10-* 2x107 3x107 2x10°

1968 — DEC PDP-10 (integrated circuit)

2x10* 1x105 36 2x10-* 1x10-5 1x107 5x10* 8x10°

1969 — CDC 7600 (transistor)

3x10^ 1x10* 64 1x10-7 2x10- 5xl0« 6x107 2x10'

2970 —GE-635 (transistor)

6x10* 1x105 32 2x10-* 1x10- 1X107 4x10* 2x10'^

1971 — SDS 920 (transistor)

3x105 6x10' 32 2x10-5 3x10-5 2x10* 2x10* 7x10°

1972 — IBM 360/195 (hybrid chip)

2x1 0^ 1x105 32 1x10-7 2x10- 3x1 0» 4x10* 1x10'

1973 — Data General Nova (integrated circuit)

3x10^ 8x103 16 2x10-5 4x10-5 1x10* 1x105 5x10'

1974 — IBM-370/168 (integrated circuit)

4x10" 3x105 32 2x10-7 4x10-7 2xl0« 8x10* 4x10'

1975 — DEC-KL-10 (integrated circuit)

1x10* 1x10* 36 8x10-7 2x10-* 5x107 4x107 5x10'

1976 — DEC PDP-11/70 (integrated circuit)

3x105 6x10" 16 3x10-* 9x10-* 8x10* 1x10* 3x10'

1976 — Apple II (integrated circuit)

6x103 8x103 8 1x10-5 4x10-5 2x10* 6x10" 3x102

1977 — Cray-1 (integrated circuit)

2x1 0^ 4x10* 64 2xl0-« 2x10-^ 3x10" 3x108 2x10^

1979 — DEC VAX 11/780 (microprocessor)

3x105 2x10* 32 2x10-* 3x10-* 2x107 6x107 8x10'

continued
 Measuring Computer Power 177

Table 2 (cant.)

Cost Memory Word T^jj ^mult Power Capacity Power/cost

2988$ words bits sec sec bits/sec bits b/s/$

2980 —Sun-2 (microprocessor)

4x10^ 3x1 05 32 3x10-* 1x10-5 1x107 8x1 0^ 3x102

2982 — CDC Cyber-205 (integrated circuit)

1x107 4x10* 32 3x1 0-« 3xl0-« 1x10' 1x108 1x10^

2982 — IBM PC (microprocessor)

3x10^ 2x10-* 16 4x10-* 2x10-5 5x10* 4x105 2x103

2982 — Sun-2 (microprocessor)

2x10^ 5x10^ 32 2x10-* 6x10-* 1x107 2x107 6x1 0^

2983 — Vax 22/750 (microprocessor)

6x10^ 1x10^ 32 2x10-* 1x10-5 2x107 3x107 3x102

2984 — Apple Macintosh (microprocessor)

2x1 0^ 3x10^ 32 3x10-* 2x10-5 8x10* 1x10^ 3x103

2984 —Vax 22/785 (microprocessor)

2x10^ 4x10^ 32 7x10-7 1x10-6 5x107 1x105 2x102

2985 — Cray-2 (integrated circuit)

1x107 3xl0« 64 4x10-^ 4x10-^ 2x10^0 2x10'° 1x103

2956 — Sun-3 (microprocessor)

1x10^ IxlO'' 32 9x10-7 2x10-* 4x107 3x107 4x103

1956 — DEC VAX 8650 (microprocessor)

1x1 05 4x10* 32 2x10-7 6x10-7 2xl0» lxlO« 1X103

2987 —Apple Mac II (microprocessor)

3x10^ 5x105 32 1x10^ 2x10-* 4x107 2x107 1x10^

1987 — Sun-4 (microprocessor)

1x10* 4x10* 32 2x10-7 4x10-7 2xl0« 1X108 2x10^

2989 — Cra\/'3 (gallium arsenide)

1x107 1x107 64 6x10-'° 6x10-'° IxlO'i 6x108 IxlC
 A3 The Outer Limits
of Computation

Mind without Machine?

In Chapter 4 I suggested that a mind is a pattern that can be impressed
on many different kinds of body or storage medium. went I further
to say that a mind could be represented by any one of an infinite
class of radically different patterns that were equivalent only in a
certain abstract mathematical sense. A person's subjective expenences
are an abstract property shared by all patterns in this class, so the
person would feel same regardless of which pattern she was
the
instantiated in. This leads to the question that if a mind is ultimately
a mathematical abstraction, why does it require a physical form at
all? Don't mathematical properties exist even if they don't happen
to be written down anywhere? Doesn't the billionth digit of pi exist
even if we haven't yet managed to compute it? In the same sense,
don't the abstract mathematical relationships that are the feelings of
a person exist even in the absence of any particular hardware to
compute them? This happens to be an old philosophical conundrum;
and though I think it must be true, do not see how to draw any
I

meaningful conclusions from it, since it seems to imply that everything

possible exists. So, instead, let's consider a slightly less sweeping line
of thought.
Suppose that a program describing a person is written in a static
medium like a book. A superintelligent being who reads and un-
derstands the program should be able to reason out the future de-
velopment of the encoded person in a variety of possible situations.
Existence in the thoughts of an intelligent beholder is fundamentally
no different than existence in a computer simulation, and we have
already suggested that a mind can be satisfactorily encoded in a
computer. But an intelligent being can do more with the simulation

178
 The Outer Limits of Computation 179

than simply carry it out rigidly —

it need not model accurately every

single detail of the beheld and may well choose to skip the boring
parts, to jump to conclusions that are obvious to it, to approximate
other steps, and to lump together alternatives it does not choose to
distinguish. Human authors of fiction do this every day as they create
adventures for the characters in their books — our superintelligent
being is different only in that its imagination works at such a level
of detail that its simulated people are fully real. Like an author of
fiction, the being can think in a time-reversed way; it may choose a
conclusion and then reason backwards, deciding what must have pre-
ceded it. Perhaps the superintelligent being prefers to imagine certain
kinds of situations and contrives to maneuver its mental simulations
to make them happen. By failing to flesh out unimportant details
in the simulation and steering events toward particular conclusions,
our intelligent being may create enough peculiarities in the simulated
world to attract the notice of the simulated person.
Quantum mechanics, a cornerstone of modern physics, seems to
imply that in the real world as we know it, unobserved events happen
in all possible ways (another way of saying no decision is made as to
which possibility occurs), and the superposition of all these possibili-
ties itself has observable effects, including mysterious coincidences at
remote times and places. Is there any connection between these ideas?
Once again have argued myself into a conundrum, though
I this time
one that has some possibility of being answered eventually.

Nondeterministic Thinking

The cellular automata universes featured in Chapter 6 have a clock-

work rigidity that makes them easy to think about. But modern
physics has revealed much more interesting foundations for our own
universe. Just what it all really means is still a matter for fascinat-
ing speculation; the only consensus is that the truth is very weird.
The following is a little self-indulgence — an attempt to harness the
strangeness to a practical end. Of course reality will turn out to be far
stranger.
Before computers, serious scientific calculations were all difficult,

lengthy, and error prone. Practical engineering problems were ap-

proached in an ad hoc manner with numerical tables, graphical meth-
ods, and clever devices like slide rules and, in this century, mechanical
180 Mind Children

calculators. Automatic computers dramatically changed the situation.

Although they could handle astronomically large problems, they
required, in advance, an absolutely precise and detailed step-by-step
specification of the procedure. Writing such programs, and getting
them was slow and tedious (though not nearly as slow as doing
right,

the calculations themselves by hand). It was a great advantage if the

same program could be used for many problems of the same tvpe,
and often such generalization made the structure of the calculation

I Think, Therefore I Am
A simulated Descartes correctly deduces his own existence. It makes
no difference just who or what is doing the simulation — the simulated
world is complete in itself.
 The Outer Limits of Computation 181

clearer. It also spawned the new mathematical field of computational

complexity, the study of the intrinsic difficulty of computing answers

to different kinds of problem.
A general program can handle problems of different sizes, and
usually its running time will grow as its problem size is increased.
The simplest sorting programs, for instance, take times proportional to
the square of the number Doubling the number
of items to be sorted.
of items quadruples the sorting time. More complicated programs
have been discovered whose times grow more slowly, and students of
computational complexity were able to prove that the fastest possible
sortingprograms would require a number of steps proportional to
the number
of items times its logarithm —
sorting one million items
should take about about ten thousand times as long as sorting one
thousand items. The time grows faster than the number of items but
not as fast as its square. The difficulty of other kinds of problem was

shown grow as the cube of the size, or fourth, or some other power.
to
Problems whose solution can be computed in times proportional to
(or less than) some fixed power of the problem size are said to be of
polynomial time, or P, type. In the computer era they are considered
easy.
Another important class of computations turned out to be much
harder. One example is the so-called traveling-salesma7i problem, which
involves finding the shortest path that passes through each of a given
set of cities exactly once. The best exact solutions that anyone has
found require that the program check almost every possible route,
then pick the best one. The possible routes are enumerated by
generating all permutations of the cities. Each permutation itself

can be checked in polynomial time, but the number of permutations

grows exponentially, multiplying manyfold each time a single city is
added. The problem becomes astronomically difficult for quite modest
numbers of cities. A hypothetical computer able to explore all possible
paths simultaneously (a mere mathematical abstraction known as a

nondeterministic machine because it does not make up its mind at

branchpoints but splits into two machines and goes both ways) could,
in principle, solve the problem in polynomial time. For that reason

these hard problems are called NP, for nondeterministic polynomial.

It many NP problems can be mathematically converted
turns out that
into one another, and that a fast (P) solution for one would solve all.
Alas, no fast solution that works on a real, deterministic computer
182 Mind Children

has been found, and it is not known if such a solution exists. This
so called P=NP? question is a central one because some extremely
important problems, for instance in the design of optimal hardware
and software, and automatic reasoning, are NP. Since exact solutions
are much, much too slow, present design systems limp along with
approximate methods that do not guarantee the best answer, and
sometimes do quite badly.
Let's suppose, as seems likely, that there is no shortcut to NP
problems. No matter how fast the conventional machine we use
to solve them, small increases in problem size will swamp it. An
intelligence designing improvements for itself will encounter many NP
issues. The efficiency of its designs, and thus its future, will depend
directly on how well it can handle these problems, so heroic methods
are warranted.

Replication and Quantum Nondeterminism

A nondeterministic computer, able to spawn indefinitely many ver-
sions of itself to examine alternative answers, is a mathematical fabri-
cation. It can be approximated, to a limited extent, by a multiprocessor
that contains a fixed number of distinct processors. At each branch-
point in the computation a new processor is invoked. Eventually,
however, the last processor will be occupied, and subsequent branches
must be evaluated The computation can be sped up only
sequentially.
by as many times making hardly
as there are individual processors,
a dent in the astronomical growth rate of NP problems. But what if
the number of processors could somehow be increased as the problem
size grew?
The ability to reproduce is a sine qua non for the members of any
race that aspires to immortality. Our machines do it now, though
biological humans are an essential part of the process. As the future
portrayed in this book dawns, more and more of the steps will be taken
over by machines, until fully automatic manufacture of automatic
manufacturing machines is the norm. In the superintelligent future,
reproduction of small thinking units will be child's play. It is thus
reasonable to imagine small computers that can both compute and
make is limited by the
copies of themselves. Reproduction eventually
finite amount and energy, but the limits could
of available material
be made astronomically high by breeding very small machines in a
 The Outer Limits of Computation 183

very large, energetic nutrient medium (bacteria-sized machines in the

oceans of Jupiter, or in pulsar-pumped interstellar dust clouds, let's

say!). One such computer would reproduce, to become two, the two
would become four, the four eight, the eight sixteen, and so on, in an
exponential growth to astronomical numbers matching, up to a point,

the exponential growth of an NP problem. If, in one "generation"

time step, a machine can do either a certain amount of computation
or reproduce itself once, the best strategy would be to reproduce like

mad until there are as many machines as there are alternatives to

examine, then for each resulting machine to examine one alternative
answer. Having computed a possible answer, the machines would
then engage in a "mine's bigger than yours" tournament with nearby
machines. Two machines would compare each other's answers, and
then both would adopt the better of the two results and go off to
compete with other machines. The best answer will rapidly spread
to all the machines, and any one can then be harvested to learn
the final result. In this process, the reproduction phase takes a
time linearly proportional to the size of the problem; working out
the individual answers is polynomial in the size, and the answer
tournament phase, like the reproduction, is linear in the size, allowing
moderately sized NP problems to be solved in polynomial time. But
very large problems would swamp any amount of available space
and would thus still be out of reach. (Earth's biosphere, of course, is
performing a computation of just this kind.)
Truly nondeterministic computers are a mathematical fiction. Or
are they? Quantum mechanics, a cornerstone of modern physics, has
indeterminism at its heart and soul. Outcome probabilities in quantum
mechanics are predicted by summing up all the indistinguishable
ways an event might happen, then squaring the result. One strange
consequence of this is that some otherwise possible outcomes are ruled
out by the existence of other possibilities. An excellent example is the
two-slit experiment. Photons of light radiate from a pinpoint source to
a screen broken by two slits, on page 184. Those that
as in the figure
make it through the
slits encounter an array of photon detectors (often

a photographic film, but the example is clearer if we use individual

sensors that click when struck by light). If the light source is so
dim that only one photon is released at a time, the sensors register
individually —
sometimes this one, sometimes that one. Each photon
lands in exactly one place. But, if a count is kept of how many photons
184 Mind Children

have, landed on each detector, an unexpected pattern emerges. Some

detectors see no photons at all, while ones close to them on either side
register many, and a little farther away there is again a dearth. If the
number of clicks from each detector is charted, the result is identical to
the banded interference pattern that would occur if steady waves had
been emitted from the location of the light source and passed through
the two slits, to be split there into two waves that interfered with
each other at the screen, constructively at some places, destructively
at others, as in the figure on page 185.

Two-Slit Experiment
A photon picked up by a detector at screen S might have come through
slit A or through slit B — there is no way to distinguish. In quantum

mechanics the "amplitudes" for the two cases must be added. At some
points on the screen they add constructively, making it likely that a

photon will end up there; at nearby points the amplitudes cancel, and
no photons are ever found.

j+trj+tr II
 The Outer Limits of Computation 185

But waves of what? Each photon starts from one place and lands in
one place; isn't it at just one place on every part of its flight? Doesn't
it go through one slit or the other? If so, how does the presence of the
other slit prroent it from landing on the screen? For if
at a certain place

one slit is blocked, the total number of photons landing on the screen
is halved, but the interference pattern vanishes, and some locations

that received no photons with both slits open begin to register hits.
Quantum mechanics' answer is that during the tlight the position of
the photon is unknown and must be modeled bv a complex valued
wave describing all its possible locations. This ghostlv wave passes
through both slits (though it describes the position of only a single
photon) and interferes with itself at the screen, canceling at some
points. There the wave makes up its mind, and the photon appears in
just one of its possible locations. This wave condition of the photon

Two Slits and Waves

When sound waves are passed through the two slits, an interference
pattern results. But no individual clicks are heard. Each wave gently
affects all the detectors.

^^
Waves out of phase
— quiet

Waves in phase
- loud
186 Mind Children

before it hits the screen is called a mixed state or a superposition of states.

The sudden appearance of the photon in only one detector is called
the collapse of the wave function.
This explanation profoundly disturbed some of the same physicists
who had helped to formulate the theory, notably Albert Einstein
and Erwin Schrodinger. To formalize their intuitive objections, they
constructed thought experiments that gave unlikely results according
to the theory. In some, a measurement made at one site caused
the instant collapse of a wave function at a remote location — an
effect faster than light. In another, more frivolous, example, called
Schrodinger's Cat, a radioactive decay that may or may not take place
in a sealed box causes (or fails to cause) the death of a cat also in the
box. Schrodinger considered absurd the theory's description of the
unopened box as a mixed state superimposing a live and a dead cat.

He suggested that the theory merely expressed ignorance on the part

of an observer: in the box the cat's fate was unambiguous. This is

called a hidden-variables theory; that is, the system has a definite state
at all times, but some parts of it are temporarily hidden from some
observers.
The joke is on the critics. Many
most "absurd" thought
of the
experimental results have been observed in mind-boggling actuality
in clever (and very modern) experiments carried out by Alain Aspect
at the University of Paris. These demonstrations rule out the simplest
and most natural hidden variables theories, local ones, in which, for
instance, the hidden information about which slit the photon went
through is contained in the photon itself, or ones in which the state
of health of Schrodinger's cat is part of the fehne.
Nonlocal hidden-variables theories, where the unmeasured informa-
tion is distributed over an extended space, are a possibility. It is easy to
construct theories of this kind that give results identical with ordinary
quantum mechanics. Most physicists find them uninteresting: why
introduce a more complicated explanation with extra variables when
the current, simpler equations suffice? Philosophically, also, global
hidden-variables theories are only slightly less puzzling than raw
quantum mechanics. What does it mean that the "exact position" of a
particle is spread out over a large chunk of space? This question was
the subject of a lively controversy among the founders of quantum
mechanics in the early part of this century. It has recently become of
widespread interest again.
 The Outer Limits of Computation 187

Many Worlds and the Doomsday NP Computer

In the two-slit experiment, a photon destined for the screen might

go through sht A or it might traverse sHt B. The interference pattern

suggests it somehow manages to do both at the same time. In 1957
Hugh Everett at Princeton published in his PhD thesis what may
be the most profligate nonlocal hidden-variables explanation of this

puzzle. In Everett's model, the photon does go through both slits,

in different universes. At each decision point the entire universe, or

at least the immediate portion of it, splits into several, like multiple
pages from a copying machine. measurement
made, the
Until a is

different versions of the universe lie in close proximity and interfere

with each other, causing banded patterns on screens, for instance. A
measurement that can distinguish one possibility from another causes
the universes to diverge (alternatively, the divergence is the definition
of "measurement"). The interference stops, and in each, now separate,
universe a different version of the experimenter can contemplate a
different unambiguous result.

The word "astronomical" hardly begins to capture the number of

distinct universes created every instant under this idea. Such numbers
alone may make the thought unappealing, vet we should not be
intimidated by mere scale. Shock also greeted the first suggestions
of the number of atoms in a speck of matter, or the distance to the
nearest stars, or the size and age of the universe. Practical physicists
object to the profligacy of universes for a different reason. Once a
measurement made, the universes in which the result was other
is

than our own no longer influence us. Postulating their continued

existence is an unnecessary complication. Thus stripped, the manv-
worlds idea is reduced to conventional quantum mechanics. The
unnecessary hidden variables (that identify which universe you are
talking about) are removed.
Can quantum mechanical indeterminism help solve difficult prob-
lems? So-called holographic methods have been demonstrated that
use coherent laser light to simultaneously search for certain patterns
(fingerprints, or enemy radar signals, for instance) in large fields of
other patterns. For the right problem, holographic methods are fast
and efficient, and the action can be interpreted as the effect of mixed
states. It can also be viewed in a more classical way as light-wave
interference. Waves are interesting because their spread is a kind of
188 Mind Children

reproduction, and different parts of a waveform can be used to per-

form different parts of a computation. But as yet only linear speedups
have been proposed or achieved. NP problems remain difficult.
But the many-worlds idea has other consequences. John Gribbin
is a writer and physicist who has expanded on its more bizarre
possibilities in several stories, articles,and books. One of his stories
has the following plot.

The Doomsday Device

Two builders of a future super (immensely expensive) particle accelerator
have a problem. The machine has been completed for months, but so far
has failed on each attempt to use it. The problem is not in the design but
seemingly just in the designer's bad luck. Lightning caused a power outage
just at turn on, or a fuse blew, or a janitor tripp^ed over a cable, or a little

earthquake triggered an emergency cutoff; each incident was different, and

apparently unrelated to the others.

But perhaps the failures are an enormous stroke of luck. New calculations
suggest that the machine is powerful enough to trigger a collapse of the

vacuum to a lower energy state. A cosmic explosion might radiate out at the
speed of light from the accelerator's collision point, eventually destroying the
entire universe. Wliat a close call!
Or was it? If the universe had been destroyed, there would be no o}ie left

to lament the fact. Wliat if the many-worlds idea were correct? In some
universes the machine would have worked. For all practical purposes those
worlds would have ceased to exist. Only in the remaitider would a pair of
puzzled physicists be scratching their heads, zvondering what had gone wrong
this time. Given so many nearly identical uniz^erses, the destruction of a few
seems of small consequence. An idea strikes them. Wliy not reinforce the
weak points in the machine so that a random failure within it is extremely
unlikely, then wire it to a detector of a nuclear attack, like the doomsday
machine in Stanley Kubrick's film Dr. Strangelove.^ Ati attack would be
met by the destruction of the offending universe. Only those universes in

which the attack had not happened, for some reason (the commanding general
had a heart attack, the missile launch system failed, the piremier had a fit

of compassion...), would live to wonder about yet another close call. The
machine in Strangelove was ineffective as a deterrent unless the other side

was aware of it. Not so the many-worlds version. No attack (that anyone will

notice) can occur so long as it operates, no matter how secret its existence.
 The Outer Limits of Computation 189

Preventing nuclear war is a laudable objective perhaps worth the

destruction of untold numbers of possible universes. But can we
use the same approach to solve day-to-day problems? Getting back
to the NP problem, could parallel universes somehow be used to

search alternative possible solutions simultaneously, allowing one

conventional computer to act like a nondeterministic one? Let's begin
by wiring a universe-destroying device to a computer, so that it can
be activated by a certain computer instruction. While we are at
it, let's also connect a true random number generator, based on a
hiss from a hot resistor or clicks from a geiger counter. Potential
solutions to NP problems are characterized by a number, for instance

the total path length in the traveling-salesman problem, that we seek

to minimize. Guess and write a program to
a probable length,
choose a path at random (using the generator you added). Have
the program trigger the doomsday device if the randomly generated
path is longer than your guess. Run the program. Most likely you
will find that the random path your program has generated is shorter
than your guess, because universes in which a longer path came up
are wiped out. But suppose there is no path that short. Well, then
your computer must have broken somehow and failed to destroy the
universe.
It is a nuisance to have your computer break in an uncontrolled
way, so why not provide an easily fixed weak link? Install another
thermally or radioactively controlled device that, with low probability,
can interrupt your computer's calculation. If you run your program
now, it produce an answer shorter than your guess, or else
will either
be interrupted before it finishes. In the former case, choose a shorter
path length; in thelatter, choose a longer one. Then run the program

again. Keep doing this until you find a path length L such that the
machine finds an answer with a guess of L, but interrupts itself when
the guess is L-1. The L-length solution is an optimal answer to your
problem. The search for a proper L can, of course, be incorporated into
the program itself to make the process fully automatic. An optimal
strategy for doing this, called a binary search, can finish the search
in anumber of steps proportional to the logarithm of L, a mere 20
stepswhen L is a million. Each step, which involves generating each
random answer, takes polynomial time. So the answer is yes, a many-
worlds version of quantum mechanics can be used to solve arbitrarily
hard NP problems in modest times.
190 Mind Children

The cost in destroyed universes is staggering, however. In the

nuclear-war-preventing use of the doomsday device, if the annual
chance of an attack were 50%, about half of the possible universes
would be destroyed each year. But in a traveling-salesman problem
with 100 cities only a few solutions out of 100^"^ (a 1 followed by 200
zeros) will be optimum, so a doomsday computer solving the problem
would destroy about 100'°^ worlds for each one it let survive! The

growth rate of new universes is much larger than that, so perhaps it

does not matter.
Only one component of this solution to the NP problem cannot
be immediately constructed today, and that is the doomsday trigger
itself. Gribbin's design depends on highly speculative physics that
has lost favor as of this writing. But is it really necessary to destroy
the whole universe if the answer comes out wrong? Of course not.
Subjectively, it is just as good to merely destroy yourself! Now that

outcome can be achieved today. Wire your computer's doomsday

connection to a cranial explosive charge, for instance. you run the If

traveling-salesman program you will, with overwhelming odds, blow

your brains out. But it's quick and should be painless. And, in one
world out of 100^°^, you will, by a tremendous stroke of luck, survive
and have the right answer. Your cranial explosive will be intact, ready
to solve the next problem.
The above idea would work with other problems of life. If an
outcome you desire, however unlikely, fails to materialize, destroy
yourself. In some universe, vou will survive, having won your bet.
If the many-worlds interpretation of quantum mechanics is correct,

as it might be, why isn't suicide a common solution to everyday

problems? The demographics of multiple universes may have some
bearing on the answer. If you are careless about losing your life, there
will be fewer copies of you among the universes. A universe picked
at random will contain mostly individuals who successfully struggle
to avoid death whenever possible. If the universes are truly infinite
in number, this has little meaning — one trillionth of infinity is still the

same infinity. The consequences are serious, however, if the number

is merely huge. And if not all possible outcomes are pursued, then
destroying yourself with high probability could, in fact, truly end your
existence.
 The Outer Limits of Computation 191

One World, Not Many?

It may be that the indeterminate nature of quantum mechanics is

simply a kind of illusion and that there is only one world. Here
is an outline of a model where the uncertainties at any location, or the
hidden variables, are simply "noise" from the rest of the universe.
Imagine, somewhere, there is a spherical volume uniformly filled
with a gas made up of a huge but finite number of particles in motion.
Pressure waves pass through the gas, propagating at its speed of
sound, s, and suppose no faster signal can be sent. The sphere has
resonances that correspond to wave trains passing through its entire
volume at different angles and frequencies. Each combination of a
particular direction and frequency is called a wave mode. There is
a mathematical transformation called the (spatial) Fourier transform
that arranges these wave modes very neatly and powerfully. The
Fourier transform combines the pattern of pressures found over the
original volume of the sphere {V) in various ways to produce a new
spherical set of values (f ). At the center of F is a number representing
the average density of V. Immediately surrounding it are (complex)
numbers giving the intensity of waves, in various directions, whose
wavelength just spans the diameter of V. Twice as far from the center
of F are found the intensities of wave modes with two cycles across
V; these are surrounded by another shell containing modes whose
wavelength is one third the diameter of V, and so on. Each point
in F describes a wave filling V with a direction and a number of
cycles given by the point's orientation and distance from the center
of F. Another way of saying this is: direction in F corresponds to
direction in V; radius in F is proportional to frequency in V. Since each
wave is made of periodic clusterings of gas particles, the interparticle
spacing sets a lower bound on the wavelength, thus an upper bound
on frequency, and a limit on the radius of the F sphere. The closer the
particles, the larger F must be.
A theorem about Fourier transforms states that if sufficiently high
frequencies are included, then F contains about as many points as
V has particles, and all the information required to reconstruct V is

found in F. In fact, F and V are simply alternative descriptions of

the same thing, with the interesting property that every particle in V
 —

192 Mind Children

contributes to the value of each point in F, and every point in F is

reflected as a component
motion of every particle of V.
of
If the particles in V bump into one another, or interact in some

other nonlinear way, then energy can be transferred from one wave
mode to another that is, one point in f can become stronger at
the expense of another. There will be a certain amount of random
transference among all wave modes. Besides this, there will be a
more systematic between "nearby" wave modes those
interaction —
very similar in frequency and orientation, thus near each other in the
F space. In V, such waves will be in step for large fractions of their
length. Because the gas is nonlinear, the periodic bunching of gas
particles caused by one mode will influence the bunching ability of a
neighboring mode with a similar period.
So nearby points in F interact systematically, distant points do not.
The interaction can be considered a physics of the f world. If the
physics is rich enough, it may be able to support the basis of complex
structures, life and intelligence, just as does ours. Imagine a physicist
made of f stuff, for whom points in f are simply locations, not
complicated functions of another space. We can deduce some of the
"laws of physics" this inhabitant of F will find by reasoning about
effects in V, and translating back to F. In the following list, such
reasoning is in italics:

Dimensionality: F is three dimensional. // V is three dimensional,

each wave train will be described by its orientation, given by two angles,
say azimuth and elevation, and by its frequency. Frequency in V becomes
radius in F, while the two angles in V remain angles in F. If V were an
n dimensional sphere, F would also have n dimensions.

Locality: Points near to each other in F can exchange energy in

consistent, predictable ways while distant points cannot. Two wave
trains in V that are very similar in direction and frequency are in step for

a long portion of their length, and the nonlinear bunching effects will be
roughly the same cycle after cycle along the length. Wave modes distant

from one another, on the other hand, whose crests and troughs are not
correlated, will lose here, and gain there, and in general appear like mere
random buffetings to each other.

Interaction Speed: There is a characteristic speed at each point in

F. Points far away from the center of F interact more quickly than
 The Outer Limits of Computation 193

those closer in. An interaction is the nonrandom transfer of energy from

one wave mode to another. The smallest repeated unit in a wavetrain is

a cycle. An effect which happens in a similar way at each cycle can have

a consistent effect on a whole wave train. Effects in V propagate at the

speed of sound, so a whole cycle can be affected in the time it takes sound
to traverse it (this is the time period of the wave). The outermost parts
of F correspond to wave modes with the highest frequencies, and thus the
fastest interaction rates.

Uncertainty Principle: The energy of a point in f cannot be

determined precisely in a short time. The best accuracy possible
improves linearly with duration of the measurement. The energy
at a point in F is the total energy of a particular wavetrain tlmt span^
the entire volume V. As no signal in V can travel faster than the speed
of sound, discovering the total energy in a wavetrain would involve
waiting for signals to arrive from all over V, a time much longer than
the basic interaction time. In a short time, the summation is necessarily
over a proportionately small volume. Since the observer in F is itself

distributed over V, exactly which smaller volume is not defined —and thus
the measurement is uncertain. As the time and the summation volume
increase all the possible sums converge to the average, and the uncertainty
decreases.

Superposition of States: Most interactions in F will appear to be the

sum of many possible ways the interaction might have happened.
YJhen two nearby wavetrains interact, they do so initially on a cycle-by-
cycle basis, since information from distant parts of the wavetrain arrives
only at the speed of sound. Each cycle contains a little energy from
the wavetrain in question and a lot of energy from many other waves
of different frequency and orientation passing through the same volume.
This "background noise" will be different from one cycle to the next along
the wavetrain, so the interaction at each cycle will be slightly different.

When all is said and done — that is, if the information from the entire
wavetrain is collected — the total interaction can be interpreted as the sum
of the cycle-by-cycle interactions. Sometimes energy will be transferred one
way by one cycle and the opposite way by a distant one, so the alternatives
can cancel as well enhance one another.

These and other properties of the F world contain some of the

quantum mechanics but are the consequence only
strangest features of
194 Mind Children

of an unusual way of looking at a prosaic situation. There are a few

differences. The superposition of states is statistical, rather than a
perfect sum over all possibilities, as in traditional quantum mechanics.
This makes only a very subtle difference if V is very large but might
result in a very tiny amount of "noise" in measurements that could
help distinguish the F mechanism from other explanations of quantum
mechanics.
The model as presented does not exhibit the effects of special
any obvious way, and this
relativity in is a serious defect, if we hope
to wrestle it into a description of our world. There is something
wrong in the way it treats time. It does have one property that mimics
the temporal effects of a general relativistic gravitational field. Time
near the center of F runs more slowly than at the extremes, since
the interactions are based on lower frequency waves. At the very
center, time is stopped. The central point of F never changes from
its "average density of the whole sphere" value, and so is effectively

frozen in time. In general relativity the regions around a gravitating

body have a similar property: time flows slower as one gets closer.
Near very dense masses (that is, black holes), time stops altogether at
a certain distance.
A few of modern physics' more exotic theories have a possible
explanation in this model. Although energy mainly flows between
wave modes very similar in frequency and direction (such as points
adjacent in F), nonlinearities in the V medium should permit some
energy to flow systematically between harmonically related wave
modes, for instancebetween one mode and another on the same
direction, but twice as high in frequency. Such modes of energy flow
in F provide "degrees of freedom" in addition to the three provided
by nearby points. They can be interpreted, when viewed on the small
scale, as extra dimensions (energy can move this way, that way, that
way and and that way...). Since a circumnavigation
also that way,
from harmonic harmonic will cover the available space in fewer
to

steps than a move along adjacent wave modes, these extra dimensions
will appear to have a much smaller extent than the basic three. The
greater the energy involved, the more harmonics may be activated and
the higher the dimensionahty. Most physical theories these days have
tightly looped extra dimensions to provide a geometric explanation
for the basic forces. Ten and eleven dimensions are popular, and
new forces suggested by some theories may introduce more. If
 The Outer Limits of Computation 195

something like the F explanation of apparent higher dimensionality

is correct, there is a bonus. Viewed on a large scale, the harmonic
"dimensions" are actual links between distant regions of space and,
properly exploited, could allow instantaneous communication and
travel over enormous distances.

Big Waves

Now, forget the possible implications of the F idea as a mechanism

for quantum mechanics and consider our universe on the grand scale.
It is permeated by a background of microwave radiation with a

wavelength centered around 1 millimeter, a length slowly increasing

as the universe expands and cools. It affects, and is affected by,
clouds of matter in interstellar space and thus interacts with itself

nonlinearly. If we do a universe-wide spatial Fourier transform of

this radiation (that is, treat our world as V), we end up with an F
space with properties much like those of the previous section. The
expansion of the universe adds a new twist. As our universe gradually
expands, the wavelengths of the background radiation increase. As
the wavelengths get longer and longer, the relative rate of time flow
in the F world slows down. Any inhabitants of F would be ideally
situated to practice the "live forever by going slower and slower as it
gets colder and colder" strategy proposed by Freeman Dyson. By now
—
they would be moving quite slowly their fastest particle interactions
would take several trillionths of a second. In the past, shortly after the
big bang, when the universe was dense and hot, the F world would
have been a lively place, running millions or billions of times faster.

In the earliest moments of the universe, the speed would have been
astronomically faster.

The microsecond of the big bang could represent eons of

first

subjective time in F. Perhaps enough time for intelligence to evolve,

realize its and seed smaller but eventually faster life in
situation,
the V space. the large scale F and V are the same
Though on
thing, manipulation of one from the other, or even communication
between the two, would be extraordinarily difficult. Any local event
in either space would be diffused to nondetectability in the other. Only
massive, universe-spanning projects with long-range order would
work, and these would take huge amounts of time because of the
speed limits in either universe. Real-time interaction between V and
196 Mind Children

F is ruled out. Such projects, however, couW affect many locations

in the other space as easily (in many cases more easily) as one, and
these could appear as entropy-violating "miracles" there. If I lived in
F and wanted to visit V, I would engineer such a miracle that would
condense a robot surrogate of myself in V, then later another miracle
that would read out the robot's memories back into an f-accessible
form.
The Fourier transform that converts V into F is identical except for
a minus sign to the inverse transform that converts the other way.
Given just the two descriptions, it would not be clear which was
the "original" world. In fact, the Fourier transform is but one of
an infinite class of "orthogonal transforms" that have the same basic
properties. Each of these is capable of taking a description of a volume
and operating over it to produce a different description with the same
information, but with each original point spread to every location in
the result. This leads to the possibility of an infinity of universes, each
same underlying stuff, each exhibiting
a different combination of the
quantum mechanical behavior but otherwise having its own unique
physics, each oblivious of the others sharing its space. I don't know
where to take that idea.
 Bibliography

Prologue

Cairns-Smith, A. G. 1982. Genetic Takeover and the Mineral Origins of Life.

Cambridge: Cambridge University Press.

1985. Seven Clues to the Origin of Life: A Scientific Detective Story.

Cambridge: Cambridge University Press.

1985 (June). "The First Organisms." Scientific American 253:90-98.

Calder, Nigel. 1983. Timescale: An Atlas of the Fourth Dimension. New York:
Viking Press.
Desmond, Kevin. 1986. A Timetable of Inventions and Discoveries. New York:
M. Evans.
Sagan, Carl. 1977. The Dragons of Eden: Speculations on the Evolution of Human
Intelligence. New York: Random House.

Chapter 1: Mind in Motion

Ashby, W. Ross. 1963. An Introduction to Cybernetics. New York: John Wiley

& Sons.
Asimov, Isaac. 1950. I, Robot. New York: Doubleday
Bakker, Robert T. 1986. The Dinosaur Heresies. New York: William Morrow.
Braitenberg, Valentino. 1984. Vehicles: Experiments in Synthetic Psychology.
Cambridge: MIT Press.

Brooks, J., and G. Shaw. 1973. Origin and Development of Living Systems. New
York: Academic Press.
Buchsbaum, Ralph. 1948. Animals without Backbones. 2nd ed. Chicago:
University of Chicago Press.
Feigenbaum, Edward A., and Juhan Feldman, eds. 1963. Computers and
Thought. New York: McGraw-Hill.
Fichtelius, Karl-Erik, and Sverre Sjolander. 1972. Smarter than Man? Intelli-

gence in Whales, Dolphins and Humans. New York: Ballantine Books.

Griffin, Donald R. 1984. Animal Thinking. Cambridge: Harvard University
Press.

197
198 Bibliography

Lane, Frank W. 1962. Kingdom of the Octopus: The Life History of the Cephalopoda.
New York: Pyramid Publications.
Marsh, Peter. 1985. Robots. New York: Crescent Books.
Mayr, Ernst. 1982. The Growth of Biological Thought. Cambridge: Harvard
University Press.
McCorduck, Pamela. 1979. Machines Who Think. San Francisco: W. H.
Freeman.
Minsky, Marvin, ed. 1985. Robotics. New York: Doubleday
1986. The Society of Mind. New York: Simon & Schuster.
Moynihan, Martin. 1985. Communication and Noncommunication by Cephalopods.
Bloomington: Indiana University Press.
Nilsson, Nils, ed. 1984. Shakey the Robot: Artificial Intelligence Center Technical
Note 323. Menlo Park: SRI International.
Pawson, Richard. 1985. The Robot Book. London: W H Smith & Son.
Pratt, Vernon. 1987. Thinking Machines: The Evolution of Artificial Intelligence.

Oxford: Basil Blackwell.

Raphael, Bertram. 1976. The Thinking Computer: Mind inside Matter. San
Francisco: W. H. Freeman.
Reichardt, Jasia. 1978. Robots: Fact, Fiction and Prediction. Middlesex: Penguin
Books.
Thorson, Gunnar. 1971. Life in the Sea. New York: McGraw-Hill.
Walter, W. Grey. 1961. The Living Brain. Middlesex: Penguin Books.
Wiener, Norbert. 1965. Cybernetics, or Control and Communication in the Animal
and the Machine. Cambridge: MIT Press.

Chapter 2: Powering Up
Babbage, Charles. 1961. Charles Babbage and His Calculating Engines. Edited
by Philip Morrison and Emily Morrison. New York: Dover Publications.
Berkeley, Edmund C. 1949. Giant Brains or Machines That Think. New York:
John Wiley & Sons.
Booth, Andrew, and Kathleen Booth. 1956. Automatic Digital Calculators.
London: Butterworths.
Cale, E. G., L. L. Gremillion, and J. L. McKenney. 1979. "Price/performance
patterns of U.S. computer systems." Communications of the ACM 22(4):
225-230.
Dertouzos, Michael L., and Joel Moses, eds. 1979. The Computer Age: A
Twenty-Year Review. Cambridge: MIT Press.
Dowling, John E. 1987. The Retina: An Approachable Part of the Brain. Cam-
bridge: Harvard University Press.
 Bibliography 199

Drexler, K. Eric. 1986. Engines of Creation. New York: Doubleday.

Fames, Charles, and Ray Eames. 1973. A Computer Perspective. Cambridge:
Harvard University Press.

Hubel, David, ed. 1979 (September). The Brain — Scientific American, special
issue, 241(3).

Kandel, Eric R. 1976. Cellular Basis of Behavior: An Introduction to Behavioral

Neurobiology. New York: W. H. Freeman.
Kandel, Eric R., and J. H. Schwartz. 1982. "Molecular basis of memory."
Science 218:433-436.
Kuffler, Stephen W., and John G. Nichols. 1976. From Neuron to Brain:
A Cellular Approach to the Function of the Nervous System. Sunderland,
Massachusetts: Sinauer Associates.
Lazou, Christopher. 1986. Supercomputers and Their Use. Oxford: Clarendon
Press.
Minsky, Marvin, and Seymour Papert. 1969. Perceptrons: An Introduction to
Computational Geometry. Cambridge: MIT Press.
Moreau, Rene. 1984. The Computer Comes of Age. Cambridge: MIT Press,
von Neumann, John. 1958. The Computer and the Brain. New Haven: Yale
University Press.
C^eisser, Hans. 1988. The Conquest of the Microchip. Cambridge: Harvard
University Press.
Randell, Brian, ed. 1973. The Origins of Digital Computers: Selected Papers.
Berlin: Springer-Verlag.

Rosen, Saul. 1971. ACM 71: A Quarter-Century View. New York: Association
for Computing Machinery.
Squire, Larry R- 1986. "Mechanisms of memory." Science 232:1612-1619.
Turn, Rein. 1974. Computers in the 1980s. New York: Columbia University
Press.
Weik, Martin H. 1957. A Second Sun^ey of Domestic Electronic Digital Computing
Systems. Aberdeen, Maryland: Army Ballistic Research Laboratories,
report no. 1010.
Wolfe, Jeremy M. 1986. The Mind's Eye: Readings from Scientific American. New
York: W. H. Freeman.
Wolken, Jerome J. 1975. Photoprocesses, Photoreceptors, and Evolution. New
York: Academic Press.

Chapter 3: Symbiosis

Boorstin, Daniel J. 1983. The Discoverers. New York: Random House.

Time-Life Books, eds. 1985. Computer Images: Understanding Computers Series.

Arlington: Time-Life Books.

200 Bibliography

Chapter 4: Grandfather Clause

Forward, Robert L. 1980. Dragon's Egg. New York: Ballantine Books.

1984. The Flight of the Dragonfly. New York: Simon & Schuster
Hofstadter, Douglas R., and Daniel C. Dennett, eds. 1981. The Mind's I. New
York: Basic Books.
Rucker, Rudy. 1983. Software. Middlesex: Penguin Books.
Sperry, Roger 1982. "Some effects of disconnecting the cerebral hemispheres."
Science 217:1223-1226.
Vinge, Vernor 1984. True Names. New York: Bluejay Books.

Chapter 5: Wildlife

Axelrod, Robert. 1984. The Ez'olution of Cooperation. New York: Basic Books.
Cohen, Fred. 1984. Computer Viruses. Los Angeles: University of Southern
California.
Dewdney A. K. 1984 (May), 1985 (March), and 1987 (January). "Core wars."
Scientific American 250:14-17, 252:14-23, 256:14-18.
Dawkins, Richard. 1976. The Selfish Gene. Oxford: Oxford University Press.
1982. The Extended Phenotype. Oxford: Oxford University Press.

1986. The Blind Watchmaker. New York: W. W. Norton.

Hoyle, Fred, and John Elliot. 1962. A for Andromeda. New York: Harper
Oliver, B. M., and J. Billingham. 1971. Project Cyclops: A Design Study of a
System for Detecting Extraterrestial Intelligent Life. Moffett Field, California:
NASA/Ames Research Center
Rosen, Eric C. 1981. Vulnerabilities of Netivork Control Protocols: An Example.
Cambridge: Bolt Beranek and Newman, Inc.

Sagan, Carl. 1985. Contact. New York: Simon & Schuster

Shoch, John F, and Jon A. Hupp. 1982. "The 'worm' programs: early
experience with a distributed computation." Communications of the ACM
25(3):1 72-180.

Thompson, Ken. 1984. "Reflections on trusting trust." Communications of the

ACM 27(8):761-763.

Chapter 6: Breakout

Barrow, John D., and Frank J. Tipler 1986. The Anthropic Cosmological Principle.
Oxford: Oxford University Press.
Dyson, Freeman. 1979. Disturbing the Universe. New York: Harper & Row.
1988. Infinite in all Directions. New York: Harper &c Row.
Gardner, Martin. 1983. Wheels, Life and Other Mathematical Amusements. New
York: W. H. Freeman.
 Bibliography 201

Hawking, Stephen W. 1988. A Brief History of Time: From the Big Bang to Black

Holes. New York; Bantam Books.

Weinberg, Steven. 1977. The First Three Minutes. New York: Basic Books.

Appendix 3: The Outer Limits of Computation

S., and Neill Graham, eds. 1973. The Many-Worlds Interpretation

DeWitt, Bryce
ofQuantum Meclianics. Princeton: Princeton University' Press.
Feynman, Richard P., Robert B. Leighton, and Matthew Sands. 1965. The
Feynman Lectures on Physics. Vol. 3. New York: Addison-VVesley.
Gribbin, John. 1984. In Search of Schrodinger's Cat: Quantum Physics and Reality.
New York: Bantam Books.
1985. "Doomsday device." Analog Science Fiction-Science Fact 105:22-
127.
Robot Pals
Some of the author's mechanical collaborators, with the years of the
association.
 Acknowledgments &
Illustration Credits

ory isn't up
T
to the task of
HIS book has roots deep in my
acknowledging them individually, but
childhood. My mem-
my thanks
go to the authors of science and science fiction, teachers, librarians, science
fair organizers and friends who helped shape my mental world through
four decades. My memory is good enough to recall that my younger sisters

Elizabeth and Alice were the long-suffering sounding boards for many years
of my long-winded speculations.
In late 1971, when I arrived as a graduate student at the Stanford Artificial
Intelligence Laboratory', a lively debate was just winding down that had been
sparked bv a proposal from Dick Fredericksen in his self-published newsletter,
A Word in Edgeunse. Over several he had developed the concept of
articles

achieving immortality, and by replacing a human nervous system,

much else,

bit by bit, with a more durable artificial equivalent. The exchange interested

me greatly because the idea had occurred to me years before, in high school,
but had suffered from lack of a receptive audience. At SAIL, Fredericksen's
proposal had polarized those who took it seriously. Bruce Baumgart was
its chief proponent, while Larry Tessler found it dehumanizing. My own
thoughts about the future of intelligent machinerv' crystallized in discussions
with Rod Brooks, Bruce Bullock, Mike Farmwald, Bob Forward, Don Gennery,
Erik Gilbert, Bill Gosper, David Grossman, Brian Harvey, Marc Le Brun,
Robert Maas, John McCarthy, Ed Mcguire, Dave Poole, Jeff Rubin, Clem Smith,
Russ Taylor, Lowell Wood, and quite a few others. In 1975 I wrote an essay
on the subject that evolved over the years into several articles and, eventually,
into this book.
The discussions continued when I came to Carnegie Mellon University
in 1980. Here it is my pleasure to acknowledge mind-stretching exchanges
with Mike Blackwell, Kevin Dowling, Alberto Elfes, Larry Matthies, Pat Muir,
Gregg Podnar, Olin Shivers, and Richard Wallace. I would also like to thank
the administration of the Robotics Institute, especially Raj Reddy and Takeo
Kanade, for maintaining an environment that allows me to pursue long-range
goals. I am equally grateful to the Office of Naval Research, and my program

203
 —

204 Acknowledgments & Illustration Credits

director, Alan Meyrowitz, for providing the steady funding that has supported
my basic research since 1981.
Though I vaguely intended to develop my ideas to book length in 1975,
it was only in 1985 that 1 seriously began to work on a manuscript. By an
amazing coincidence, within two weeks of undertaking the project I received a
letter from Howard Boyer, the newly arrived Editor for Science and Medicine

at Harvard University Press, inviting me to write just such a book. In the

three years since then, Howard has whipped the book through a grueling
series of writing and publishing hurdles. 1 am deeply grateful for his interest
and insight.

The first hurdle — to produce a detailed outline that would pass editorial
muster at Harvard Press — was surmounted with the aid of extensive reviews
written by Vernor Vinge. The many drafts of the manuscript which followed
were greatly improved by the Press's referees, whose comments were var-
iously encouraging, informative, and stern. 1 thank Rod Brooks, Richard
Dawkins, Kee Dewdney, Bruce Donald, John Dowling, Bob Forward, John
McCarthy, Pamela McCorduck, and others whose identity 1 have not learned.
The most dramatic improvements in the book occurred when it was put in
the hands of my manuscript editor, Susan Wallace. Susan reorganized the text
from a ragged collection of ideas into a cohesive whole, setting the stage for
—
and coaching to completion a rewrite that made a night-and-day difference
in the book's quality.
With exceptions noted below, the line art in the book was drawn by me
on a Macintosh II from Apple Computer. The programs used were Cricket

Draw from Cricket Software and SuperPaint from Silicon Beach Software, with
occasional dips into digitized clip-art collections — the McPic! packages from
Magnum Some art
Software and ClickArt packages from T/Maker Graphics.
was scanned from hand drawings and photographs with the Thunderscan
program and hardware from Thunderware. Earlier versions of many of
the drawings had been produced on smaller Macintoshes with MacPaint,
MacDraw, FuUPaint, and MacDraft.
Mike Blackwell redrew "Intelligence on Earth" (page 18) in Cricket Draw
from my MacDraw original. "The Retina" (page 54) was drawn on a Macintosh
SE by Mary Jo Dowling, using Adobe Illustrator from Adobe Systems. She
worked from an illustration which appeared in The Retina by John Dowling (no
relation). The gear and integrated circuit icons in "A Century of Computing"

(page 64) were drawn by Gregg Podnar using MacPaint. "A Robot Bush"
(page 103) was produced by a program which wrote with help from Mike
1

Blackwell, in Apple Computer's MPW C language, running on a Macintosh

II. It contains one quarter of a million line segments and took ten hours to

compute. The "Selfish Martians" cartoon (page 142) was drawn by Kimberlee
Faught with Cricket Draw. The pictures of the Cart, Pluto, and Neptune in
 Acknowledgments & Illustration Credits 205

"Robot Pals" (page 202) are digitized and touched-up renderings of pencil
drawings by Bill Nee. The picture of Uranus is a digitized photograph touched

up by me and Gregg Podnar.

The following organizations provided photographs and granted permission
to reproduce them: "Walking Machines" (page 27), courtesy of Odetics, Inc.;

'Three Fingers" (page 30), courtesy of David Lampe, MIT; "Autonomous

Navigation" (page 33), courtesy of Denning Mobile Robotics, Inc.; "Object
Finding" (page 35), courtesy of SRI International; "The Retina" (page 54),
courtesv of John Dowling; "ENIAC" (page 76), © Smithsonian Institution;
"Magic Glasses" (page 87), courtesy of United Technologies /Hamilton Stan-
dard; "Robot Proxy" (page 88), courtesy of Naval Ocean Systems Center;
"Unreal Estat^The Road to Point Reyes" (page 92), © 1986 Pixar.
The book was designed by Joyce C. Weston, Marianne Perlak, and Mike
Blackwell. It was typeset by Mike Blackwell in the TgX document preparation
system created at Stanford University by Don Knuth, as instantiated in the
Texturesprogram by Addison-Wesley, running on a Macintosh II. The typeface
is from Adobe Systems. Camera-ready copy was generated
Palatino, obtained
on a Linotronic 300 digital typesetter owned by the Robotics Institute of
Carnegie Mellon University, driven by a Macintosh SE. Early drafts of the
book were printed on Apple laserwriters.
 Index

A for Andromeda, 137 Bacteria, 18-19, 43, 61, 69, 144, 167, 183
Abstract thought, 16 Barrow, John, 148
Abstraction, 39, 131-135, 141, 145, 178 Basic, 40
Addition, 120, 164-165, 172 Batch mode, 80
Address, 78, 155, 172 Bees, 43, 61, 138, 151, 168
Adept, 12 Bell Laboratories, 66, 71, 82, 132

AI, see Artificial intelligence Bin-picking problem, 36

Algol, 40 Binary search, 189
Algorithms, 31, 149 Biotechnology, 72, 108
Alternative actions, 39, 43, 47-49, 123, Bipolar cell, 54-57
179 Birds, 18-19, 97
Alto, 85 Bit, 60, 77, 131, 171-172
Altruism, 141 Bivalves, 18-19
Amacrine cell, 54, 57 Black holes, 101, 194
Amphibians, 18-19 Blinkers, 151
Analog computer, 6-8 Blocks world, 14
Analytical Engine, 65-66, 69 Blue-green algae, 18-19
Animal Thinking, 43 Body-identity position, 116-118
Aplysia, 44, 60, 166 Body/mind problem, 4, 116
Apollo, 51, 101 Boeing, 87
Apple Computer, 11-12, 37, 85 Boredom, 47, 63, 91, 114, 153, 179
Apple 11, 128 Bottom-up evolution, 17, 20, 50
AppUcations software, 24, 37, 45-46 Brain, 18-19, 52-54, 59-61, 108-115, 118-
Arithmometer, 66 120, 163-165, 190
ARPAnet, 133 Brainstem, 110-111
Arthropods, 18-19 Bulk storage, 77
Artificial intelligence (Al), 7-10, 13-16, Bulletin boards, computer, 81, 127-128,
21, 50-51, 61, 75, 151, 163 131
Asimov, Isaac, 10
Aspect, Alain, 186 C compiler, 132-133
Assemblers, 78-81 Cairns-Smith, A. G., 3
Attention span, 114 Calculation, 66-67, 77, 113, 158, 164,
Automatically Guided Vehicle (AGV), 180
22 Calculators, 2, 11, 61-62, 65-66, 167, 180
Axelrod, Robert, 141-145 Calculus, 8-10
Axons, 53-58 Camera, 21, 31-32, 48, 88-90, 153, 164
Camera eye, 13, 88
Babbage, Charles, 65-66, 69 Capacitor, 7

207
208 Index

Capek, Karel, 10 Darwinian evolution, 3, 17, 44, 158,

Cart, Stanford, 20, 31-33 167
Cells, 18-19, 34, 56, 73, 117, 136, 139, Data, 135, 139, 171-172
150-152, 155-156, 167 Dawkins, Richard, 136, 159
Cellticks, 152-153, 158 Death, 4, 18-19, 110-112, 115-121, 138
Cellular automaton, 150-151, 171, 179 Defense Advanced Research Projects
Center-surround, 54, 57, 59 Agency (DARPA), 21, 133
Cephalopods, 18-19 Dendrite, 53
Ceramic chips, 67 Denning Mobile Robotics, 32
Cerebral cortex, 165 Denning Sentry, 33
Checkers, 8, 15, 150
Descartes, Rene, 180
Chess, 8-9, 13, 78, 171
Determinism, 123, 181
Cohen, Fred, 128
Devol, George, 10
Colossus, 8
Dexterity, 29-30, 102
Common sense, 9, 13, 20, 83
Digital Equipment Corporation, 51
Communication, 7, 43, 90, 112-115, 133,
Divergence, 187
195
Diversity, 125, 141
Compiler, 79, 120, 132-133
Computation, 58, 120, 148, 163, 166,
DNA , 2, 18-19, 61, 108, 116, 136, 151
Dolphins, 19, 115
169, 188
Computational power, 57, 65, 68 Dowling, John, 52
Computer bulletin board, 81 Drive, 26

Computer evolution, 36 Dualism, 119-120

Computer image, 35 Dynabook, 84-85
Computer language, 40 Dyson, Freeman, 148, 195
Computer model, 35
Computer power, 169 Earnest, Les, 20
Computer software, 25 Echinoderms, 18-19
Computer system configuration, 9 Einstein, Albert, 74, 93, 186
Computer virus, 126-131, 135 Electric charge, 77
Computer vision, 20, 36, 52 Electric current, 117
Computer word, 60 Electric motor, 66
Computer modeling, 72 Electronic brain, 7
Conditioning, 45-50 Electronic computer, 62
Consciousness, 37-40, 43-44, 93, 110-111 Electronic library, 86
Contact, 137
Electronic map, 89
Control language, 40
Electronic switch, 56
Control theory, 11
Electronic turtle, 7
Convergent evolution, 37-39, 42, 57
Electronics, 6, 10, 56
Conway, John Horton, 151
Electrons, 70
Cooperation, 141-145
Elephants, 19, 61, 115, 168
Core dump, 80
Elfes, Alberto, 34
Corpus callosum, 111-113
Elliot, John, 137
Cray 2, 60-61, 71
Cray 3, 71 Emergence, 44
Crick, Francis, 151 Emotion, 37-39, 44, 111

Crystals, 3, 71-72, 152 Engelberger, Joseph, 10

Cultural evolution, 2-4 EN AC

I , 8, 75-78
Culture, 4, 18-19, 101, 115-116, 122, 158 Enigma, 8
Cybernetics, 7-8, 11, 14-16 Everett, Hugh, 187
 Index 209

17, 48-50, 55, 82, 100, 115, Hand, 9, 29, 34, 86-88, 102, 109-111,
Evolution,
119, 122, 136-140, 149, 154, 158, 163, 158-159, 180

166 Hardware, 31, 89, 168, 178, 182

Existence, 141, 158, 178, 190 Hash, 155-158

Expert systems, 9 Hashlife, 154-158

Extinction, 101 Heat death, 147-148

57-58, 86-88, 105, Hidden variables, 186-187, 191-194
Eye, 9, 19, 32, 39, 52,

110-112 High-level language, 51, 79-81, 131-

132, 171
Hitachi, 26-28, 37
Feelings, 44, 49, 93, 112
Hofstadter, Douglas, 145
Fermi, Enrico, 138
Hollerith, Herman, 67
Fermi paradox, 138
Horizontal cell, 54-57, 164-165
Fifth Generation, 52, 68
Hoyle, Fred, 137
Forgetfulness, 42
FORTRAN, 79
Human equivalence, 68, 72-73, 100,
109, 167
Fourier transform, 191-196
Human intelligence, 17
Fourth generation, 68
Human mind, 1
Fovea, 58
Human nature, 122
Function table, 76
Human vision, 168
Fungi, 18-19
Hybrid chip, 64, 67

Galileo, 97
IBM, 8, 67, 80, 128
Gallium arsenide (GaAs), 64, 71
IBM 1130, 126
Game theory, 141-143
IBM 650, 60
Games, 122, 128
Icon, 82-84
Ganghon, 19, 54, 57-59, 166
Idea, 116, 138-139, 196
Gardner, Martin, 151
Identity, 109, 115-120, 171
Gauss, Carl Friedrich, 120
Imagery, 47
General Motors, 11, 24, 36
Imagination, 48, 133-135, 139, 179
General-purpose robot, 22-25, 38 Imaging eye, 19
General relativity, 154 121-124, 148, 182
Immortality, 5,
Genetic code, 128 Indeterminism, 183, 187, 191-194
Genetic engineering, 108 Industrial manipulator, 29, 40
Genetic information, 3 Industrial revolution, 2, 8-10, 65
Genetic takeover, 3-4 Industrial robot, 11, 23, 36, 47
Genetics, 72, 110, 115, 136 Industrial vision system, 11
Genome, 166 Infinite in All Directions, 148
Germanium, 71 Information, 63, 114-116, 119, 137, 163,
Ghosts, 131-134, 185 166-173
Giant squid, 19, 115 Information retrieval, 81
Gosper, Bill, 155-158 Information theory, 63, 169
Gravity, 95-97 Integrated circuit, 7, 55, 64, 67-73, 104,
Gribbin, John, 188-190 168
Griffin, Donald, 39, 43 Intelligence, 16-19, 106, 114, 136, 145,

Gripper, 23, 29 152, 192

Growth of Biological Thought, 44 Interactive program, 80
InterfaceMessage Processor (IMP), 133
Hackers, 81-83, 132, 141, 151-155 Interim Dynabook, 85
HamUton, William D., 140, 144 Internal storage, 77
210 Index

Intron, 136 Magnetic memory, 10

Invertebrate, 42, 56 Magnetic resonance, 109
Magnetic tape, 80, 112, 126
Jacquard loom, 10 Mammals, 18-19
Japanese, 26, 36, 52, 71 Manipulation, 38, 84, 195
Jet Propulsion Laboratory (JPL), 21 Manipulator, 31, 36, 90, 110
Jobs, Steve, 85 Manual labor, 25, 101

Johns Hopkins Beast, 7 Manufacturer's learning curve, 23

Many worlds, 187-189
Kandel, Eric, 60 Mapping, 31-34, 37, 43, 48, 89-90, 152
Kay, Alan, 84-85 Mass, 94-97, 148
Kepler, Johannes, 98 Mass production, 66
KL-10, 51 Mathematical logic, 14
Kubrick, Stanley, 188 Mathematical notation, 79
Kuffler, Stephen, 52 Mathematics, 82, 93, 108, 121, 149, 158,
164, 178, 181-183
Language, 2, 9, 83, 93, 113, 132, 135, Matter, 118
153, 171 Matter transmitter, 117-118
Laser rangefinder, 21 Matthies, Larry, 32-34
Lasers, 72, 114 Mayr, Ernst, 44
Leg, 22, 26-28, 102, 107 McCarthy, John, 8, 20
Leonardo da Vinci, 6 Memory,' 57, 60-66, 77-81, 85, 96-97,

Life, 44, 111, 115-116, 123-125, 134-137, 114-116, 119-122, 135, 145, 158, 167,

151, 192 170-172, 196

Life expectancy, 152 Menu, 82-84
Life extension, 108, 116 Microbes, 141
Life program, 151-158 Microcomputer, 36, 127
Light, 6, 54, 72-73, 93, 154-155, 168, Microelectronics, 72
183-184, 187-188 Microfauna, 144
Lindauer, Martin, 43 Micromachining, 104
Lisa, 85 Microorganism, 144
Location, 31, 34, 37, 45, 48, 96, 171-173 Microphone ear, 7, 13
Locomotion, 25-26, 38 Microprocessor, 64, 68, 71
Logic, 13 Microtechnology, 73
Logic gate, 7, 131 Mind, 44, 109-121, 124, 158, 178

Logic Theorist, 8 Mind transferral, 115, 117

Mind/body problem, 4, 116
Machines, 7, 49, 110-112, 120, 126, 137- Minsky, Marvin, 8-10
139, 147-150, 167-171 Mitochondria, 18-19
Machine intelligence, 16 Mobile platform, 153
Machine language, 77-78, 131-133, 152- Mobile robot, 14, 20-21, 29, 32, 37
153, 171 Mobility, 15-17, 28, 31, 40, 85, 153

Macintosh, 11, 37, 61, 68, 85, 91, 94 Mollusks, 18-19

MACSYMA , 9 Moore, Gordon, 68
Magic glasses, 85-94, 110, 123 Motion detector, 59
Magic gloves, 86-88 Motivation, 116
Magic wardrobe, 90-91, 97 Mouse, 61, 83, 168-169

Magnetic core, 77-78 Multicelled animals, 15, 18-19

Magnetic disk, 77 Multifingered gripper, 29

Magnetic field, 70 Multiprocessor, 62, 121, 182
 Index 211

Nanocomputer, 73 Pattern-identity, 117-119, 122-123

Nanotechnology, 73, 104 PDP-10 , 51
NASA, 21 Perception, 15-17, 55, 75, 167
Natural selection, 44, 145 Perfection, 125, 131
Naval Ocean Systems Center (NOSC), Person, 117-119, 178
88 Personal computer, 24, 37, 68, 83-84,
Neptune robot, 32 127-128
Nerve, 57, 111, 117 Personal death, 121
Nerve cell, 53 Personal identity, 109
Nerve network, 16 Personality, 111, 115

Nervous system, 4, 7, 16-17, 39, 42-44, Perspective, 110

52-55, 116, 163-166 Pestilence, 139
Network, 56, 84-86, 90, 98, 120, 129, Pests, 125, 145
133-134 Phobia, 46
Neuroanatomy, 111 Photocell, 7, 54-58, 117, 164-165
Neurobiology, 52, 57 Photon, 70, 148, 183-186
Neuron, 16-19, 44, 53-56, 59-60, 108- Photoreceptor, 56
110, 163-167 Physics, 15, 71, 74, 93, 96, 107, 147-152,

Neurotransmitter, 55 183, 190-192

Neutron, 114, 123 Pixar, 92
Neutron star, 74, 114, 123 Pixel, 58
Newell, Allen, 8 Planck, Max, 73
Newton, Isaac, 94, 97-98, 154 Planck's constant, 73
Newway, 152-153, 157-158 Planets, 101, 114, 123-124
Nilsson, Nils, 14 Planning, 9, 14, 36, 139
Nondeterminism, 189 Plants, 16-19, 139
Nondeterministic polynomial (NP) Pocket calculators, 100
problems, 106, 179-183 Pocket televisions, 86
Nonverbal knowledge, 25 Point Reyes, 92
Nucleated cell, 18-19 Polaroid, 32
Postbiological world, 1, 5, 125, 141, 145
Object finding, 35 Power, 26-29, 49, 61, 93, 106, 134, 148,
Obstacle avoidance, 38 169-172
Octopus, 42, 56 Power per unit cost, 52
Odetics, 26-27 Predators, 129-131
Odex, 27 Predictability, 39, 63, 154-156, 171
Office of Naval Research, 21 Prediction, 48, 123, 155-156, 164
Optic nerve, 57-58, 111 Pressure, 86-88, 108, 148, 191-194
Optics, 72, 86 Prey, 47, 129-131, 139
Optimizing compiler, 120 Price, 62, 65, 137
Organisms, 105, 135, 139-140, 146, 166- Primitive action, 15, 40
167 Prisoner's dilemma, 141-144
Probability, 34, 45, 143, 170
Pain, 42, 45-49 Problem size, 181
Parallel world, 154 Problem solving, 14, 51
Parallel-jaw gripper, 29 Process, 117-120
Parasites, 126, 133, 136, 139-140, 144- Processing power, 51, 63, 170-171
146 Program, 62-63, 76, 79-82, 109-114, 120,
PARC, 83-85 123, 126-136, 141-143, 153-154, 163-
Pattern, 117-122, 150-158, 165, 178, 184 166, 171, 178-180
212 Index

Programmer, 60, 79-81, 126, 131-132 Robotics, 10-17, 20-22, 167

Programming, 93, 135 Robotics industry, 48
Protein, 72-73 Robotics Institute, 32
Proxy robot, 88-91 Robotics Research Vehicle (RRV), 21
Pulsars, 183 Rosen, Charles, 14
Punched card, 10, 65-67, 80, 126 Roving burglar alarm, 32
Punched tape, 126 Roving robot, 39-40, 43-44
Punishment, 81
Purpose, 152-153 Sabotage, 128
Sagan, Carl, 137
Quantum, 70-73, 123, 149, 153-154, 179, Salisbury Hand, 30
183-187, 191-196 Salisbury, Ken, 29-31, 37
Quantum dot, 71 Samuel, Arthur, 8
Scanning tunelling microscope, 72
Radiation, 70, 108, 147, 195 Schrodinger, Erwin, 186
Radio, 61, 67-69, 84, 106 Science fiction, 10, 23, 117
Radio technology, 6 Screen, 84, 90-91, 152, 184-186

Radio telescope, 136 Screen editor, 51

Radio tube, 7 Sea slug, 44, 60, 166
Radioactivity, 186, 189 Sea urchin, 19
RAND, 8 Second law of thermodynamics, 147
Randell, Brian, 65 Self replication, 135, 150, 165
Random, 123, 135, 139, 143, 152, 157, Self reproduction, 102, 133-135, 151
189 Selfishness, 141-144
Range, 32-34 Semiconductor, 71-72, 102
Rangefinder, 21 Sense, 93-94
Rapoport, Anatol, 143 Sensorimotor ability, 15-17
Reasoning, 13, 50, 75, 113, 149, 182 Sensors, 6, 23, 32, 39-42, 45, 87, 104
Reasoning program, 13-15, 20 Serrey, Bruno, 34
Receiver, 117-118, 136 SETl', 136-137
Reflex arc, 105 Seven Clues to the Origin of Life, 3
Relativity, 99, 149, 153 Sex, 18-19, 115, 128, 131, 135, 139-140

Relay, 62, 66-67, 168 Shakey, 14-15, 20

Remote control, 6, 14, 21 Shannon, Claude, 8, 63, 169

Replication, 151-152, 182 Shaw, John, 8
Reproduction, 3-4, 18-19, 129, 136-140, Shellfish, 19
144, 182-183, 188 Sikorsky, 87
Reptiles, 18-19 Silicon, 68, 71, 78

Retina, 19, 52-59, 163-166 Silicon carbide, 71

Retrovirus, 136 Simon, Herbert, 8
Ribosomes, 72 Simulation, 48, 91, 98, 108-110, 120,
Ritchie, Dennis, 132 123-124, 152-157, 164-166, 178-179
RNA , 72, 136 Simulator, 48-50, 123
Robot arm, 9-14, 20-22, 29, 40, 73 Single-cell animal, 19
Robot bush, 102-108 Skyhook, 94
Robot control program, 24 Slug, 60
Robot hand, 13, 22, 29 Snails, 61, 168

Robot proxy, 90 Social insects, 19

Robot security guard, 32 Software, 89, 127-133, 168, 182
Robot servant, 23 Solar power, 19, 102
 Index 213

Solar system, 101, 116 Sujjerstrings, 74

Sonar, V, 22, 32-34, 37, 41 Superstructure, 167
Songbirds, 136 Surprise, 17, 37, 40, 44, 63-65, 139, 158,

Soul, 183 169-172

Sound, 6, 86, 106, 185, 191-194 Surreptitiousness, 127
Space, 94, 108, 130, 140, 147-150, 154, Symbiosis, 75, 138, 145
157-158, 196 Symbolic programming, 78
Spacesuit, 89 Synapse, 44, 53-61, 165-166
Spacetime, 156-158 Synchotron, 70
Special relativity, 194 Synthetic imagery, 86
Species, 119, 140 System manager, 128
Spectrum, 136
Speech, 86, 96 Tape, 66-67, 81, 119, 122, 129, 150-151,
Speed, 32, 62, 112, 154, 165, 172, 188, 173
195 Television, 14, 23, 33-37, 56-61, 69, 86,

Speed of light, 155 153, 164

Sperm whale, 19, 61, 168 Television camera, 7-9, 12-14, 31-33, 86
Sperry, Roger, 111 Teraops, 59-62, 68, 72
Spider, 43 The Anthropic Cosmological Principle,
Spinal cord. 111 149
Split brain. 111 The Extended Phenotype, 136
Spontaneous generation, 133 The Selfish Gene, 136
Spontaneous mutation, 17 The Walk, 96
Spores, 144 Thermodynamics, 147
Sputnik, 21 Thompson, Ken, 132-133
Stanford Artificial Intelligence Lab Thorpe, Chuck, 32
(SAIL), 12 Thought, 93, 112, 116, 119, 122, 148-
Stanford Artificial Intelligence Project, 149, 167
20 Tune, 115, 123, 130, 147-150, 154-158,
Stanford Cart, 20, 31 166, 172, 179-183, 194-195
Stanford Research Institute (SRI), 14, Time bomb, 126-128, 131, 136
20 Time sharing, 51, 80-84
Star, 85 Tipler, Frank, 148
Starfish, 19 Tit for tat, 143
Steam engine, 147 Tools, 2, 18-19, 29, 73-75, 102, 105-106
Steam power, 65 Top-down approach, 16, 20, 50
Stem cell, 53 Torres Arithmometer, 66
Stereoscopic vision, 38 Torres y Quevedo, Leonardo, 66
Storage, 77, 122, 178 Touch, 86, 105
Stored program, 77 Transcription, 1 22
STRIPS, 14-15 Transformation, 13, 120-121, 149, 191-
Subminiature radio tube, 7 194
Supercomputers, 36, 52, 60, 68-71, 120, Transistor, 64, 67, 71, 78, 168
169 Transition, 1 70
Superconductors, 70-71, 142 Transition rule, 152-154
Supercurrent, 168 Transition table, 150
Superintelligence, 102, 107-108, 122, Transmigration, 108
125, 145, 147, 178-179, 182 Transmitter, 117-118
Superlattice, 70 Transplant, 109, 121-124
Superposition, 179, 186, 193-194 Transportation, 29, 90
214 Index

Traveling salesman problem, 181, 189- von Frisch, Otto, 43

190 von Neumann, John, 8, 75-76, 150-151,
Trojan horse, 126-128 171
Tubes, vacuum, 64, 67, 78, 168 Voyager, 123
Turing, Alan, 8, 132
Turing Award, 132 Walking robot, 26
Two-slit experiment, 183-184, 187 Walter, W. Grey, 7
Wasp, 47
Uncertainty, 123, 191-194 Watson, James, 151
Unicelled ancestor, 19 Waves, 184-187
Unimation, 10 Wave mode, 191-194
UNIVAC I 67 , Waveform, 188
Universal robot, 29, 34, 45, 48 Wavelength, 71, 148, 191-195
Universality, 150-151, 154 Wavelike, 150
Universe, 101-102, 108, 116, 138-139, Whales, 19, 61, 115

147-154, 158, 179, 187-196 Wheels, 26

Unix, 82, 132 Wheels, Life and Other Diversions, 151
Unpredictability, 145 White dwarf, 74
Uranus robot, 34 Wiener, Norbert, 7
Utilitarianism, 44 Wildlife, 125, 128, 131-133, 139-141
Word-recognizer, 46
Vacuum tube, 64, 67, 77, 168 Word processor, 24
Verbal interface, 39 World knowledge, 13
Verbal knowledge, 25 World model, 39
Vertebrate, 18-19, 42-45, 53, 56, 111 World simulator, 48
Video recorder, 61 Worms, 17, 136
Viking, 21, 101
Virulence, 128, 138 X-rays, 70, 122

Virus, 61, 101, 126-132, 135-137, 145 Xerox, 85

Vision, 36-41, 52, 105, 145, 158-159, 165 Xerox Palo Alto Research Center
Vision system, 11-12 (PARC), 83
Visual acuity, 58
Visual field, 111-112 Zuse, Konrad, 66
$8.95

"One would be making a mistake to let Mind Children recede unopened into
a guiltless oblivion. It's a tonic book, thought-provoking on everv page. And
it reminds us that, in our accelerating, headlong era, the future presses so
close upon us that those who ignore it inhabit not the present but the past."
— Brad Leithauser, Nezv Yorker

"A comprehensive and highly readable survey of the state of the art in
robotics."
—M. Mitchell Waldrop, New York Times Book Review
"Moravec, by his ownis an intellectual joyrider, and riding his
admission,
runaway an exhilarating experience
trains of thought is This is an intel-
. . .

lectual party that shouldn't be pooped, no matter how much it may disturb
the neighbours and encourage over-indulgence."
—
Brian WooUey, Guardian

"[Mind Children] has the accuracy of a college text and the can't-put-it-down
appeal of a good novel. Moravec has turned the flights of mind of one of the
world's foremost roboticists into hard copy. And he has written a tremen-
dously good book in the process."
— Eric Bobinsky, Bifte

"A dizzying display of intellect and wild imaginings by Moravec, a world-

class roboticistwho has himself developed clever beasts Undeniably, . . .

Moravec comes across as a highly knowledgeable and creative talent —

which is just what the field needs."
—
Kirkus Reviews

Hans Moravec is Director of the Mobile Robot Laboratory of Carnegie

Mellon University.

Harvard University Press ISBN D-b7M-57t3lfl-7

Cambridge, Massachusetts
and London, England
90000
Jacket design bif Joyce C. Weston. Illustrations cour-
tesy of: (front left) © Marjorie Nichols '87/Boston
Computer Museum/Maxell Tapes; (front right)
Christopher Croioley and The Media Laboratory
of MIT, ©MIT 9 780674"576186