[ExI] intelligence and generalization

[John Clark]

I originally sent this to the list a week ago but it kept bouncing so I
sent it directly to Bill and we discussed it off list, I'll try to send it
to the list again:
=====

On Mon, Jan 7, 2019 at 10:36 AM William Flynn Wallace <foozler83 at gmail.com>
wrote:

> How do we teach creativity to an AI when we don't even know what it is or
> how to measure it?


Einstein didn't know what creativity was or how to measure it and nobody
taught him to be creative, and yet he was, although he couldn't teach
others to be as creative as he was. There is considerable evidence
something similar will be true of AI. The exact same program, AlphaZero,
was able to play 3 entirely different board games GO, Chess, and Shogi at a
superhuman level. Previous programs could only play one game and they had
to have many thousands of games played by human grandmasters in their
database before they could do anything, but AlphaZero was different it
started with nothing in its database except for the basic rules of the
games, it then started playing against itself and used Neural Network
Algorithms to get better and better.

In just a few hours AlphaZero went from being a know nothing beginner to
easily defeating Deep Blue who in turn defeated the best human Chess player
in 1996. The downside of Neural Networks is after a while they become so
complicated nobody understands how they work anymore, or at least not in
any detail; and AlphaZero can't explain why it made the moves it did
either. But then the same was true for Einstein. People are always asking
geniuses how they can do what they do but they can never give satisfactory
answers because they don't know they just do it. If they did know and could
explain it we'd all be geniuses just like them.

John K Clark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20190114/d9fec695/attachment.html>