http://lesswrong.com/lw/42t/aixistyle_iq_tests/ discusses some research that attempts to turn Hutter's AIXI into an IQ test that could be uniformly applied to humans, AI programs, cats and dogs, etc. One issue with this line of attack is that simple Q-learning algorithms already outperform humans (see for example http://users.dsic.upv.es/proy/anynt/paper1-comparing.pdf). Many, but perhaps not all, would prefer a measure of intelligence for which current AIs are below the human level. -Thomas C On Tue, Jun 10, 2014 at 4:45 PM, Warren D Smith <warren.wds@gmail.com> wrote:
I'm much less optimistic than Dan Asimov re machine intelligence.
The difference with computer chess is, chess ratings allowed us to estimate progress and extrapolate it 10 (or whatever) years ahead. With the Turing test, I see no analogous thing and no way to extrapolate future progress.
I actually invented a vastly superior measure of "intelligence" to the Turing test, which by the way also was invented (essentially the same idea) by Marcus Hutter:
http://rangevoting.org/WarrenSmithPages/homepage/works.html paper #93
it to some extent would allow extrapolation of future progress, though it certainly is not as easily employed as chess ratings.
_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun