Comment by joe_the_user
10 hours ago
I don't think it's solely a matter of raw strength, but as Shin said, a willingness not to play to the program's strengths. I mean, one thing that rankled me about original Lee Sedol match was that Lee had no access to the program's "record" while the machine by the nature of the AI training process had effectively studied Lee's games in great detail.
I recall a while back someone came up with a set of "anti-computer" strategies that allowed even an amateur to defeat a strong go program. These moves weren't anything like ordinary go moves (and perhaps the "loophole" has been closed now) but imo, their existence suggests that a study of programs may reveal other unexpected weakness.
In a march 2026 interview David Wu (lightvector, Katago’s creator at Jane Street) noted that he doesn’t have a systematic solution for the cyclic group problem, but adding examples to the training set mostly ensures Katago during MCT rollout figures it out. I don’t think there has been a post mid 2024 verified exploit.
https://gomagic.org/david-wu-on-building-katago/
I saw the same things when the OpenAI Dota bots could eviscerate humans 1v1 - even pros lost!
Until a more average player confuses the AI with an unseen behaviour (pulling creeps between the towers etc) to get an advantage.
We saw this with AlphaStar too, but ultimately it feels like simply an exploit. I expect even a relatively simple modern LLM/model working with the custom transformer would have been able to address these exploits after a game.
I don't think exploit is the right term?
Anyway. Yes if you throw examples into training it will be able to handle the situation - but handling unseen things for me is a key goal.