Shin Jin-seo beat Go AI KataGo: how it differs from AlphaGo

Shin Jin-seo, the world’s No. 1 Go player, beat KataGo, the strongest Go AI in existence. At a July 2026 Kishin tournament in South Korea, Shin played a best-of-three against KataGo with a two-stone handicap and came from behind to win 2-1. He was thrashed in Game 1 on July 17, resigning after 254 moves, but took Game 2 on the 19th by 4.5 points after 290 moves and Game 3 on the 21st by 11.5 points after 221 moves (Hankyung, Ilgan Sports). He is the first professional to win two straight two-stone handicap games against KataGo in an official match since the AI’s big jump in strength.

2-1

Best-of-three resultcame from behind

2 stones

Handicapnot an even game

2019

KataGo first releasedMIT open source

This was not an even-game win. Shin started by placing two black stones on the board, a handicap match, which makes it a different thing from Lee Sedol’s single even-game win over AlphaGo in 2016. So this isn’t a human surpassing the AI on even terms; given a two-stone start, the world’s best found enough room to work. Shin drew the same line himself, calling it “a meaningful result in the sense that I beat the AI in a fight” (Hankyung).

What exactly is KataGo?

KataGo is a Go AI that David Wu built and first released on February 27, 2019. It’s written in C++ and distributed under the MIT license (Wikipedia). That one line is the whole point. AlphaGo was a closed system Google DeepMind only ran in-house; KataGo put its source code and trained neural networks out on the open internet. Anyone can download it and run it on their own laptop, and when pros review games with an engine, it’s usually KataGo. What Go circles now call “AI analysis” is, in practice, this.

Its first training run borrowed compute from Jane Street Capital: 28 Nvidia V100 GPUs for 19 days. After that it moved to distributed training, with volunteers donating compute, and kept getting stronger — reaching a 28-block, 512-channel (b28c512) network by August 2024 (Wikipedia).

How it differs from AlphaGo and AlphaGo Zero

AlphaGo (2016) first learned from professional game records (supervised learning) and layered reinforcement learning on top. AlphaGo Zero, the next year, threw out human games entirely and reached the top from a blank slate using only self-play. KataGo descends from that AlphaGo Zero / AlphaZero line. The backbone is the same: learning from self-play, with no human games.

	AlphaGo Zero (2017)	KataGo (2019–)
Open?	Private to DeepMind	MIT-licensed open source
Training	Self-play (no human games)	Self-play (same)
Design goal	General: Go, chess, shogi	Go-specific (liberties, ladders, komi inputs)
Practical use	Win-rate only	Score estimate, handicaps, many rulesets
Efficiency	Thousands of TPUs, days	Under 30 GPUs, 19 days

Two things set them apart. First, KataGo packs in a lot of Go-specific design. Where AlphaGo Zero aimed for a general architecture usable across Go, chess and shogi, KataGo hand-feeds Go-native information — liberties, ladders, komi, pass-alive status — as input features, and adds global-pooling layers that read the whole board (arXiv, “Accelerating Self-Play Learning in Go”). That pushed training efficiency way up. Per the paper, while AlphaZero ran thousands of TPUs for days and ELF ran thousands of GPUs for two weeks, KataGo surpassed ELF’s final model in 19 days on fewer than 30 GPUs — at least a single-digit multiple more efficient, by the developer’s conservative estimate (arXiv).

Second, KataGo is loaded with practical features. Unlike AlphaGo, which only spits out a win rate, it gives a score estimate, supports arbitrary komi and handicap scoring, small and rectangular boards, and multiple rulesets (Wikipedia). This Kishin match could even be framed as a two-stone handicap precisely because KataGo was built to handle handicap situations in the first place.

What does a human beating a top AI actually mean?

From where I sit, the number that matters is the handicap size. On even terms, even the world No. 1 rarely beats KataGo, and that has been the reality for the past few years. The real news in this 2-1 is that, converted into two stones, the human took more than half. The shape says the same: Shin lost Game 1 badly, then won the last two, a sign that a human’s practical feel can claw back the AI’s early handicap-game edge.

There’s one caveat worth keeping in mind. Even an AI that looks unbeatable has gaps. In 2022, researchers demonstrated an adversarial attack that lured KataGo into a specific pattern and toppled it absurdly, and adversarial training didn’t fully close the hole (arXiv 2211.00241). Superhuman strength and a local blind spot live in the same body. Shin’s two handicap wins show human feel finding that seam. They are not evidence that KataGo has gotten weaker.

FAQ

Did Shin Jin-seo beat KataGo on even terms?

No. He played with a two-stone handicap, placing two black stones before the game began. On even terms, KataGo is still stronger.

What is the biggest difference between KataGo and AlphaGo?

Openness. The AlphaGo family was private to DeepMind; KataGo is MIT-licensed open source that anyone can download and run.

Did KataGo learn from human games?

No. Like AlphaGo Zero, it learned purely through self-play. Its Go-specific design just makes that training far more efficient.

How is this different from Lee Sedol's 2016 win?

Lee won one game on even terms; Shin won two with a two-stone handicap. The conditions differ, so a direct comparison is hard.