The world’s #1 Korean Go champion faces off against the open-source AI KataGo, and in the first match they still lose despite playing with two stones advantage

On July 17, the Korea Baduk Association held a Go human-vs-AI match. With a two-stone handicap, world No. 1 Shin Jin-seo battled for 245 moves but still lost to the open-source AI model KataGo. The event follows a best-of-three format (two wins). On July 19 and 21, there are two more games, and the results have not been announced yet.
(Background: The LLM chess tournament ends: OpenAI o3 wins the title; xAI Grok 4 was shut out without winning a single game)
(Additional context: Google fires a shot: AGI is already dead, and the ASI threshold is actually 100 million ordinary people)

Table of contents

Toggle

• The 103 moves that couldn’t hold despite the two-stone handicap
• Who wrote this program
• A decade, from a team to a single graphics card

A decade ago, behind AlphaGo’s victory over Lee Sedol was an entire DeepMind team. But ten years later, it was not the resources of an organization that caused the world No. 1 Korean player Shin Jin-seo—despite losing only by taking advantage of the two-stone handicap—to still fall short. Instead, he lost to an open-source program that anyone can download for free and run on a consumer-grade graphics card.

The 103 moves that couldn’t hold despite the two-stone handicap

The match was hosted by the Korea Baduk Association and took place on July 17, coinciding with the 10th anniversary showdown between Lee Sedol and AlphaGo. According to Yonhap News Agency, in the first game Shin Jin-seo received a two-stone handicap, started as Black, and also had ample extra time to use; his opponent KataGo had no time limit, with only a 20-second per-move countdown. In professional Go circles, such a handicap and time-control setup was originally regarded as a huge advantage sufficient to swing the result.

In the opening stage, KataGo broke away from conventional openings with special moves that disrupted Shin Jin-seo’s preparation rhythm. In the middle game, Shin Jin-seo managed to hold onto the handicap advantage for a time and the game became tense and evenly matched—until move 103, when a turning point appeared. KataGo’s counterattack precisely hit weaknesses in Black’s formation. From then on, the gap gradually widened; after 245 moves, Shin Jin-seo resigned.

After the match, Shin Jin-seo said that diverging from expectations in the opening was the main reason for the loss, and he also stated directly that it is difficult for human Go players to defeat the top Go AI in direct head-to-head games. Reports say that his pre-match goal was to “take two wins.” In post-match interviews, he added: “My nerves were already thrown off when White’s second move came. My prepared plan was completely disrupted, and my mindset also started to wobble.”

The report describes how he sat in front of the board for a long time afterward, looking regretful. Post-match analysis by Korean media pointed out that his win probability started to drop after moves 70 and 76. A critical mistake appeared on move 90, and after a hundred-plus moves the situation had been completely reversed. This event uses a best-of-three (two wins) format, with two more games on July 19 and 21, and the results are not yet known.

Who wrote this program

According to KataGo’s official GitHub page, this Go AI was first released in early 2019 by its developer, the handle lightvector, David J. Wu. His day job is at quantitative trading firm Jane Street—not a full-time in-house team from a lab. In other words, training a Go AI capable of challenging human top players has already been reduced to a level that does not require enterprise-grade resources.

KataGo uses the same self-play reinforcement learning framework as AlphaZero, but it has made major optimizations in training efficiency. With general hardware, it can be trained to the level of professional Go players within days, and with a single high-end graphics card it can surpass the superhuman level within months. It is currently widely considered the strongest among publicly available, usable Go AI systems, outperforming ELF OpenGo and Leela Zero—both of which also reach superhuman strength. Without enabling search, its strength is roughly equivalent to among the top 100 European Go players; with 2,048 computations per move, it is far stronger than any human Go player.

A decade, from a team to a single graphics card

In March 2016, AlphaGo beat Lee Sedol 4–1. That year, Lee Sedol’s final human win against top Go AI came via “the move of the gods” in Game 4—the 78th move. His comment left when he retired in 2019 was: “AI cannot be defeated.”

A decade later, defeating the world No. 1 was no longer about propping up a whole research institution. It was instead an open-source project plus a consumer-grade graphics card. There are two more games on July 19 and 21, but perhaps the real suspense is already no longer contained in the scoreline.