Friday, December 08, 2017

Changing My Mind on AI

This report by kottke on the advances in AI, particularly the advances in learning, makes me change my mind.  I've been relatively conservative on my expectations for AI.  I remember back in the late 80's getting excited by the possibility of using AI to make "person " determinations for payment limitation purposes.  That evaporated under the pressure of other demands on time and resources and the wide gap between us program specialists and the private consultant types we were talking to.

Over the years I've followed with some interest the progress of chess playing software, which finally beat the best human player a few years ago.  But the slowest of the progress and the narrow limits of the field meant to me I should take the dramatic predictions of the future of AI with a big dose of salt.

But now I've changed.  Why?  Because of the advance in AI in learning how to do AI. If I understand it, the key is setting a desired criteria--what it means to "win" a chess game--provide starting conditions and letting the computer teach itself, by playing itself repeatedly and changing the program used based on the outcome--if a difference in the program brings the outcome closer to the desired criteria, incorporate it.

So the important thing is the improved strategy for AI, and presumably a strategy which can be applied to any situation where you can identify a desired criteria, a definite outcome.  It's "learning how to learn" applied to software. 

[Update: a piece in Technology Review on the subject. Perhaps a bit more balanced than the Kottke piece.]

No comments: