Well, ideally we should drop the automatic hash() and only accept:
1) ints/longs
2) buffer-like objects

(and tell people to hash() explicitly if they want to)

If that's too disruptive, we should document it.
And, for 3.x, provide the following recipe to hash from a bytes object without losing entropy, and keeping the same results under 32-bit and 64-bit builds:

>>> import random
>>> random.seed(int.from_bytes(b'foo', 'little'))
>>> random.random()
0.08384169414918807