"Superhuman" Go AIs still have trouble defending against these simple exploits

arstechnica.com

31 pointsamichail5 comments

A trained vision system may recognize sofa with a tiger pattern as a tiger. Because what it learned during training was to recognize tiger stripes, not tigers.

It is important to not anthropomorphize machine learning.

Has similar research happened against AlphaZero or other pure-NN Chess engines?

Curious to see what chess exploits would look like.

If Google had made Alpha Zero available for the public, I’m sure we would see exploits very quickly.

I saw one exploit where hikaru made the position very locked up, with essentially no way to progress the game without sacrificing. I don’t know if there are any exploits against alpha zero though.

I found the video: https://youtu.be/DTMnrUdd8Tw?si=dJxB1aq_TOfMPCk8

From what I can tell, the engine he plays, Rybka, is an alpha beta search engine that is rated around 3000 elo? I'm not sure how to compare it with Hikaru being ~2900, but it seems like a fairly "even" match-up without a crucial exploit being the reason for it losing.

Reply on news.ycombinator.com