AI: Relative difficulty (Was: SCI:BIO: raw genome length)

Eliezer Yudkowsky (
Wed, 15 Jan 1997 19:13:32 -0600

[Anders Sandberg:]
> Was it Marvin Minsky or another AI guru who said something like: "Computer
> vision seems fairly easy, relative to figuring out general planning" and
> set a graduate student to solve problem during the summer? :-)

Computer vision *is* easy... relative to general planning! I mean we
now have some vague idea of how the visual cortex does what it does
[thanks to David "2-1/2 D" Marr, RIP], even if we don't have the raw
power to implement it, and we can even say what tasks specific layers of
neurons are performing... it's incredibly easy. Relatively speaking.

While general reasoning is still WAY beyond us; we have vague ideas
about how it *might* work but nothing more, and nobody is even *close*
to drawing neural correspondences.

So I stand by my statement that if we have the software, computing
power, and raw intelligence to draw lines from DNA to the brain, hacking
up a virtual world takes an afternoon. The environment accounts
probably for most of our personalities, but very little for our
problem-solving abilities (as opposed to strategies). Though I could be
wrong; visual visualization might derive its rules from the environment,
with the genes just setting up the visual cortex to learn from it. But
I still think that the major difficulty is going to be understanding the
preprogrammed stuff, not how it evolves.

--       Eliezer S. Yudkowsky

Disclaimer:  Unless otherwise specified, I'm not telling you
everything I think I know.