[ExI] Vint Cerf on AI, at ORNL

John Clark johnkclark at gmail.com
Thu Sep 5 11:25:43 UTC 2019


On Wed, Sep 4, 2019 at 6:22 PM William Flynn Wallace via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> Seemingly missing from all the intelligence definitions in your post is
> the ability to adapt to novel situations, which to me is really important.
>

If you had never even known there was a game called "chess" but then one
day you were shown the basic rules of the game (and nothing else) and then
told you had 24 hours to teach yourself to become the best Chess playing
entity on the planet, and you did exactly that, and then you did the same
thing with GO and Shogi (both more complex than Chess) wouldn't you say
that proved you had the ability to adapt to novel situations? Well
AlphaZero did all that and it didn't use brute force to do it either.

Stockfish is another Chess program but unlike AlphaZero it didn't teach
itself humans did. Stockfish could easily beat any human player but it
couldn't beat AlphaZero despite the fact that Stockfish was running on a
faster computer and could evaluate 70,000,000 positions a second
while AlphaZero's much smaller computer could only do 80,000.

And AI is becoming less narrow and brittle every day. GO is played on a 19
by 19 grid, but if you changed it to 20 by 20 or 18 by 18 and gave it
another 24 hours to teach itself AlphaZero would be the best player in the
world at that new game too, and all without any human help. It's true
AlphaZero is not infinitely adaptable, but then humans aren't either.

 John K Clark
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20190905/98e871bd/attachment.htm>


More information about the extropy-chat mailing list