AlphaGo versus Lee Sedol

Google announced the match on 27 January 2016, six weeks before it began: "AlphaGo's next challenge will be to play the top Go player in the world over the last decade, Lee Sedol. The match will take place this March in Seoul, South Korea." The terms came the day before play — five games, no handicap, and a purse of one million US dollars that DeepMind said in advance would be "donated to UNICEF, STEM charities and Go organizations" if AlphaGo won. The money was never the stake.

What made the challenge plausible was in that same January post. It reported that "Using a single machine, AlphaGo won all but one of its 500 games" against the other strong Go programs, and revealed a match that had been played behind closed doors between 5 and 9 October 2015 against the three-time European champion Fan Hui, which AlphaGo won five games to nil. The post also noted that "experts predicted it would be at least another 10 years until a computer could beat one of the world's elite group of Go professionals." The paper is Silver et al., Nature 529 (7587), 484-489.

AlphaGo won the Seoul match four games to one. The game worth keeping is the fourth, on 13 March, and the move worth keeping is Lee's 78th, played with white. DeepMind's own account of the match singles it out: "Known as 'God's Touch', this move was just as unlikely and inventive as the one AlphaGo played two games earlier." Hassabis said that day that Lee "was too good for us today and pressured #AlphaGo into a mistake that it couldn't recover from." By DeepMind's account the match "earned AlphaGo a 9 dan professional ranking — the first time a computer Go player had received the highest possible certification."

Nineteen months later the version Lee played was itself obsolete. On 18 October 2017 DeepMind published AlphaGo Zero, which "learns to play simply by playing games against itself, starting from completely random play," and which after three days of self-play beat "the previously published version of AlphaGo ... by 100 games to 0." The supervised bootstrap from human games — the part that had made a machine strong enough to sit across from Lee — turned out to be a handicap, not a foundation.

Lee stopped competing in 2019. He told Yonhap that "with the debut of AI in Go games, I've realised that I'm not at the top even if I become the number one," and that "even if I become the number one, there is an entity that cannot be defeated."

That last sentence is now known to be false in a specific and strange way. In November 2022 a group including Tony Wang, Adam Gleave, Kellin Pelrine, Sergey Levine and Stuart Russell published adversarial policies against KataGo — which is, in the paper's words, "the strongest publicly available Go AI system at the time of writing." How badly they beat it is a number that moved. That first version reported "a >99% win-rate against KataGo without search, and a

50% win-rate when KataGo uses enough search to be near-superhuman"; the attack was strengthened across three revisions, and by July 2023 the paper reported "a >97% win rate against KataGo running at superhuman settings." "Our adversaries do not win by playing Go well." They exploit a blind spot around large cyclically connected groups: "first, set up an 'inside' group and let or lure the victim to surround it, creating a cyclic group. Second, surround the cyclic group. Third, guarantee the capture before the victim realizes it is in danger and defends." The engine holds a win probability above 99% for most of the game and typically sees the loss coming about one move before the group falls.

The attack is legible enough that a person can run it by hand: "one of our authors, a Go expert, was able to learn from our adversary's game records to implement this attack without any algorithmic assistance," and, playing under ordinary conditions on the KGS server, "obtained a greater than 90% win rate against a top ranked KataGo bot that is unaffiliated with the authors" — also beating "KataGo and Leela Zero playing with 100k visits each, which is normally far beyond human capabilities." Adversarial training did not close the hole: the core vulnerability "persists even in KataGo agents adversarially trained to defend against our attack."

Game four in Seoul was not the last time a human beat a top Go engine. It was the last time one did it by playing better Go.