From: Nicholas Bostrom (bostrom@mail.ndirect.co.uk)
Date: Tue Oct 07 1997 - 09:18:34 MDT
Damien R. Sullivan wrote:
> }[Nicholas Bostrom wrote:]
> } It doesn't even seem implausible to me that some
> } genetic algorithm or neural network architecture/learning rule that
> } might be so simple that it could be written on the back on an
> } envelope could achive superintelligence, given enough hardware and
> } unlimited interaction with the external world.
>
> I think we have one already. Do a bunch of things; see what happens;
> pick the thing that worked best; repeat. It works for creating
> lifeforms, building immune systems, running anthills, and Hofstadter's
> creative thinking programs.
Once you have a fitness function you can do gradient ascent and add
noise to avoid getting stuck on local maxima; in principle you could
even do an exhaustive search. However, if this is to work for the
average person with the superfast computer, as in my formulation of
the problem, then we would need to find a way to formalize the
success criteria. We must teach him to write a progam that evaluates
whether a given evolved network is intelligent or not; because it
would take too long if he were to do the evaluation manually. So how
can we formalize a criterion for intelligence? (A given IQ test like
the Raven Progessive Matricies wouldn't work in this case, since a
network would evolve that would answer those particular questions in
an ad hoc manner, but wouldn't generalize to other problems.)
Nicholas Bostrom
http://www.hedweb.com/nickb
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 14:45:00 MST