Anyone know if there’s a human-executable adversarial attack against LeelaKnightOdds pr similar? Seems like the logical next piece of evidence in the sequence
AI is massively superhuman, if you’re playing chess against Stockfish you can’t predict what move it will make but you can predict that it’ll win.
These adversarial-to-humans chess AIs necessarily play weaker chess than would be optimal against an approximately perfect chess player. It seems likely that there are adversarial strategies which reliably win against these AIs. Perhaps some such strategies are simple enough to be learnable by humans, as happened with Go.
A cursory google search didn’t turn anything up though. But my Google-fu is not what it used to be, so “I didn’t find when I googled” is not strong evidence that it doesn’t exist.
Anyone know if there’s a human-executable adversarial attack against LeelaKnightOdds pr similar? Seems like the logical next piece of evidence in the sequence
AI is massively superhuman, if you’re playing chess against Stockfish you can’t predict what move it will make but you can predict that it’ll win.
Actually humans can beat AI with a pretty small material advantage
No, that’s just because the AI hasn’t trained with a large material disadvantage, and models that optimally exploit human weaknesses can overcome quite large material handicaps
is
These adversarial-to-humans chess AIs necessarily play weaker chess than would be optimal against an approximately perfect chess player. It seems likely that there are adversarial strategies which reliably win against these AIs. Perhaps some such strategies are simple enough to be learnable by humans, as happened with Go.
A cursory google search didn’t turn anything up though. But my Google-fu is not what it used to be, so “I didn’t find when I googled” is not strong evidence that it doesn’t exist.