World No. 1 Go player Shin Jin-seo lost the opening game of a human-versus-AI match held by the Korea Baduk Association on July 17, falling to the open-source model KataGo despite receiving a two-stone handicap. Shin resigned after 245 moves.
Opening game goes to KataGo, with two more games still to play
The event is being played as a best-of-three series. After the first game on July 17, the next two are scheduled for July 19 and July 21, and the overall result has not yet been decided.
According to Yonhap, Shin played black with a two-stone handicap advantage and had ample main time. KataGo had no overall time limit and was given only 20 seconds of byo-yomi per move. In professional Go, that combination of handicap and time settings would usually be treated as a substantial edge for the human player.
An unusual opening and a turn after move 103
KataGo departed from standard opening patterns and disrupted Shin’s preparation from the start. Through the middle game, Shin at one point managed to preserve the handicap advantage and the position remained close.
The turning point came on move 103. The report said KataGo’s counterattack struck accurately at weaknesses in black’s formation, and the gap widened from there until Shin resigned on move 245.
South Korean media reviews after the game said Shin’s win rate began to decline after moves 70 and 76, with a critical mistake appearing on move 90. By the time the game moved past 100 moves, the position had fully swung.
Shin says the game left prepared lines behind almost immediately
After the match, Shin said the main reason for the loss was that the opening developed outside his expectations. He also said that, at this point, human players would find it difficult to defeat top Go AIs in a direct game.
According to the report, Shin had set a pre-match goal of winning two games. After the loss, he said, “White’s second move already made me panic. The plan I had prepared was completely disrupted, and my mindset started to shake as well.”
The report added that he remained seated in front of the board for a long time after the game, looking regretful.
KataGo was first released in early 2019
KataGo’s official GitHub page says the Go AI was first released in early 2019 by developer David J. Wu, known online as lightvector. The report identified him as a worker at quantitative trading firm Jane Street rather than a member of a full-time laboratory team.
That, the report argued, shows how far the barrier has fallen for building a Go AI capable of challenging the world’s best human players, without the need for enterprise-scale resources.
Built on the AlphaZero framework with training efficiency improvements
KataGo uses the same self-play reinforcement learning framework as AlphaZero, but with extensive optimization around training efficiency. The report said it can reach professional-player level within days on ordinary hardware, and exceed human level within months on a single high-end GPU.
It is now widely regarded as the strongest publicly available Go AI, stronger than other superhuman systems such as ELF OpenGo and Leela Zero. The report also said that without search enabled, KataGo’s strength is roughly comparable to a top-100 European player, while at 2,048 visits per move it is already far stronger than any human player.
Ten years after AlphaGo
The timing of the match also coincides with the 10th anniversary of the Lee Sedol-AlphaGo contest. In March 2016, AlphaGo defeated Lee Sedol by a 4-1 score. Lee’s “divine move” in Game 4, move 78, became the last human win against a top Go AI. When he retired in 2019, Lee described AI as “unbeatable.”
Ten years later, the system defeating the world’s top-ranked human player is no longer backed by an entire major research organization, but by an open-source project that can run on a consumer-grade graphics card. The scoreline in this series remains unfinished, with two games still scheduled for July 19 and July 21.