I've tested this a bit on my own arch linux install, and it seems that this has to do with how python3 handles bytestrings and your use of sys.stdin to request filenames. From the comments for genfromtxt:

fname : file or str
File, filename, or generator to read. If the filename extension is
`.gz` or `.bz2`, the file is first decompressed. Note that
generators must return byte strings in Python 3k.

"Note that generators must return byte strings in Python 3k." I suspect that sys.stdin is not returning the byte string that python3 requires, but I'm not familiar enough with it to know why.

Solution
Replace sys.stdin with input, which is a builtin function for python3 (raw_input for python2). This should work almost identically to your current implementation.

data = np.genfromtxt(input(), names=True)

Or for a more interactive experience, you can have it print a string to the console first:

data = np.genfromtxt(input("Please input a filename here: "), names=True)

Thanks @nigeil for your comment, but that solves a different problem. I want to call my program as

$ ./test.py < test.dat

That way I don't need to bother with the filename in python.
I believe you are right that sys.stdin returns strings, whereas genfromtxt requires bytestrings. I will need to read up on how to get a file-like object that returns bytestrings.

Ah, I see the difference. Darned python3. You could replicate the functionality you're looking for with my solution by instead passing a file containing only the filename of the data file - but that is pretty lame. If you find a solution, please share it!

With Python 3.3.2 and numpy 1.8, I can reproduce your error, and get it to work using sys.stdin.buffer instead of sys.stdin, you may want to give it a try.

Thanks @jaimefrio. sys.stdin.buffer does indeed work. Personally, I think genfromtxt() should support sys.stdin directly, as it feels more natural to me. What do you think?

Yes, I think I agree: it seems like an important shortcoming in Python 3 support. That said, I don't know that code enough to understand what the implications would be.

My general feeling is that numpy's read-from-file functionality needs a serious overhaul to bring it up to the standard that pandas has set, and I am not sure that small incremental changes will take us there.

But PRs are always welcome! ;-)

Fixed in 1.14