Re: answers I'd like from an SI

From: Wei Dai (
Date: Sun Nov 11 2007 - 17:41:53 MST

Stathis Papaioannou wrote:
> An intelligence must have *some* idea of what the criteria for a
> satisfactory answer to a question might be, even if it is something as
> vague as "it just seems right"; otherwise the question reduces to a
> request for random information. If these criteria could be formalised,
> that could be used in the objective function of an optimization
> process to answer the question, along with information on how the
> criteria should change in the light of new learning. Is it wrong to
> speak of an optimization process where the objective function is
> subject to revision?

Yes, I think it would be better to instead call it an intelligence that
makes use of an optimization subroutine, because much of its dynamics is now
outside the optimization process and in how it revises its objective

But to sidestep this semantic issue, let me restate my thesis this way: an
important part of intelligence, namely the ability to contemplate and make
progress on these types of questions, can not be modeled or implemented as
an optimization process with a fixed objective function.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:00 MDT