From: Eric Watt Forste (arkuat@pobox.com)
Date: Tue Jan 14 1997 - 12:08:52 MST
Eliezer writes:
>Um, I could be wrong, but I think that most of the data controlling how
>the human mind works is inside the genome... even if it needs external,
>environmental triggers. You are correct in that even if we understand
>the genome, we might still need some understanding of external
>environment in order to build an AI from it... but I think that external
>understanding could be hacked up over a few hours.
Why don't you try running some numbers? I don't have my reference books
handy, but I vaguely recall that the bandwidth of long-term memory
storage for human beings is about 2 bits per second. Don't trust me on
this: I wouldn't be at all surprised if I was off by several orders of
magnitude (it's only my long-term memory I'm consulting, after all).
But actually, I don't even trust those measurements. The way that
information passes back and forth between human beings in live
conversation, through tone of voice, facial expression, gesture,
word, etc seems very-high-bandwidth to me, and measuring the exact
bandwidth seems like a very difficult project. But my unreliable
introspective observation is that what makes me unique, what makes
me Eric Watt Forste, has a lot more to do with the unique constellation
of information that my brain has ingested since I was born than it
has to do with my genome. Not that I believe in tabula rasa or
anything, but quite some time ago I decided that AIs will probably
have to be raised like children, educated slowly (though we'll
probably be able to make multiple copies of the first few that are
properly educated). I could be wrong about this if they have a
*much* higher input bandwidth than we do, but I still expect that
the first few AIs are going to be quite stupid for the first several
years they are running. Isn't this Lenat's expectation with CYC
(or whatever his current project is)? I don't think you can just
copy a workable system of ethics into someone's head, at least not
with current technology. I think the information content of the
total set of values our culture has evolved to this point is very
high, and memetic replication has nothing like the error-correction
mechanisms that genetic replication has. Communication is difficult,
but we have managed to muddle through so far somehow.
I do take it that you've read Hofstadter's essay "Waking Up from the
Boolean Dream", have you not? I think it's one of his more important
pieces of writing, especially considering how long ago he wrote it. It's
rather prophetic of much of the work that has been done with genetic
algorithms and artificial neural networks since then.
Oh, and I'm still not going to participate in your Poll, because you
left out Art again, even after Natasha reminded you.
-- Eric Watt Forste ++ arkuat@pobox.com ++ http://www.pobox.com/~arkuat/ Naked is a state of mind. -- Luscious Jackson
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 14:44:00 MST