Robin Hanson wrote:
>
> Peter McCluskey wrote:
> >I question the appropriateness of using Flynn effect theories to predict
> >machine intelligence increases (but I suspect you are right that super-
> >intelligence will result from widespread breakthroughs).
>
> Well you obviously have a lot of company, as long threads continue on
> about how AIs will explode in intelligence, and no one yet seems to think
> my position plausible enough to be worth offering an argument against it.
>
> I find it striking that people seem to think that the speed and nature of
> learning is one area where AIs will be very different from humans, even
> though they seem to think AIs will be similar in so many other ways,
> including having consciousness, wanting to squash bugs, being unable to
> clearly articulate their goals, etc. In contrast, I think that while many
> other details can differ a lot, the speed and nature of learning is not
> one of them. Yes, maybe AIs can redesign their hardware and low level
> code, but I don't see that giving big wins after the first easy wins are
> taken.
A good and interesting point. One thing that seems less than certain to
me concerning future AIs and learning is the notion that they will be
almost totally able to share learned knowledge with one another
relatively fully and instantaneously. But if the AI has a lot of its
mentality modeled along the lines of neural nets (or the equivalent)
then its learning is likely to be a matter of a huge number of
interconnnections being made in its neural matrix including connections
intertwined with what it already knew and the individual being it is.
This may first of all be not much faster to acquire in subjective time
than it is for a human when learning a difficult subject. And once
learned it may be no easier to extract the learned knowledge cleanly
than in the case of a human because that net of information is quite
entangled with other things that have little to do with the subject
except to that particular individual. Assuming it can be more or less
disentangled I would doubt the result is anywhere near as rich or that
it would give the same overall depth of knowledge and experience of the
subject to a recipient of the knowledge. And that recipient would have
to go through yet another process of integrating the information.
Perhaps this is an argument for not using things that resemble such
networks for general knowledge acquisition and association.
- samantha
This archive was generated by hypermail 2b30 : Mon May 28 2001 - 09:50:16 MDT