2016-2026 · A decade of AI joseki

AlphaGo turned joseki from answers into choices

Ten years ago, professionals still asked whether a joseki had been completed. Today the sharper question is which exchange must happen now, and which can wait ten moves. A 96,058-SGF archive, two tournament records, and node-by-node KataGo analysis trace that change.

2016

Completing the joseki ceased to be an obligation

In Lee Sedol-AlphaGo game two, Black 13 tenukied before the lower-right corner was fully settled. Black 15 then played a peep once dismissed as crude, forcing professionals to recheck whether local completion was a false requirement.

2017

The 3-3 point moved into the opening

Master's 60 wins, DeepMind's explanation of the new 3-3, and Ke Jie's move-three 3-3 at Wuzhen turned an alien coordinate into a professional research program.

2018-2020

Adoption jumped from single digits to forty percent

Pros copied more than the first move; they absorbed the option to leave local sequences unfinished. In this public sample, early 3-3 invasions under a bare 4-4 rose from 9.72% in 2017 to 40.06% in 2018-2020.

2021-2026

The revolution became ordinary grammar

An early 3-3 is no longer news. Together with attachments, probes, hanes, and tenuki, it forms a timing network whose answer changes with nearby strength, ladders, komi, and sente.

How one pattern moved from 1.66% to nearly half of games

2013-2015: 1.66% (85/5121)
2016: 1.16% (26/2248)
2017: 9.72% (235/2417)
2018-2020: 40.06% (2588/6461)
2021-2025: 46.92% (5165/11007)
2026*: 45.59% (444/974)

The denominator is eligible deduplicated games in each period; the numerator is games with at least one defined event by move 40.

* 2026 is an incomplete snapshot through August 24. CWI notes that tournament coverage is uneven, so the figures show a time trend inside one consistent archive, not a census of all professional Go.

Experiment one: a joseki can be left unfinished

After move 12 of game two, the lower right still looked as if Black owed another local move. AlphaGo played Black 13 at J17 instead. Fan Hui's notes preserve Gu Li and Zhou Ruiyang's surprise: could even this fundamental joseki be left incomplete?

Modern KataGo gives a calm answer. Before Black 13 it evaluates Black at about 37.47% and -0.57 points; after J17, 37.48% and -0.57. To one decimal place, the cost vanishes. The machine was not breaking joseki. It judged that further local exchanges were not urgent.

Experiment two: the 3-3 buys future options, not only corner territory

DeepMind's 2017 explanation identified the point. The traditional early 3-3 was disliked not merely for taking territory, but because finishing every exchange handed over outside influence at once. AlphaGo omitted closing exchanges, left the corner unsettled, and preserved miai to escape either way or finish later.

A decade later the idea is normal in a world event. In Shin Jinseo-Liao Yuanhe at the 2026 Chunlan Cup, the first four moves occupied four corners and Black 5 immediately invaded the upper-left 3-3. At 1,000 visits KataGo's first choice is C17. Its PV plays only five local moves, tenukies to P3, and returns later at G18.

Experiment three: modern joseki is a timing network

Continue through Shin's game and the upper-left joseki pauses repeatedly. KataGo's main lines move between the corner and whole-board points. A sequence is no longer joseki because its coordinates stay contiguous, but because each exchange is worth cashing now.

Every node below uses the same game, fast model, and 1,000-visit cap. Win-rate and score-lead ordering do not always match, so small differences are not absolute grades. Visits and PVs show where the search spends attention.

Joseki did not disappear. It finally returned to the whole board.

Professionals knew before AlphaGo that joseki must serve the whole board. Human research costs still allowed many local conclusions to harden into maxims. Strong AI put each settled exchange back into a search tree and let it bid against the other three corners.

A PNAS study analyzed more than 5.8 million professional decisions from 1950-2021 and generated about 58 billion counterfactual game patterns with AI. It found that human decision quality rose significantly after superhuman AI, while novel decisions became more frequent and more strongly associated with quality. This article's 3-3 curve gives that population result one concrete board shape.

So learning joseki today is not replacing a 2015 book with a 2026 book. Learn three questions: if I omit this exchange, how can the opponent use it; where is the largest point if I tenuki now; and after one round elsewhere, will this corner's answer change?

Method and reproducibility boundary

The professional-game statistic is generated by scripts/analyze-joseki-adoption.mjs from the CWI games.tgz downloaded on 2026-08-24. It parses 96,058 main lines and deduplicates them by full-move-sequence SHA-1 to 88,268 games before applying year, 19x19, even-game, and move-40 filters.

The 2016 game uses Read19's published deep analysis, kata1-zhizi-b40c768nbt-s11272M-d5935M, at roughly 2,000 visits per ordinary node. The 2026 game was analyzed on 2026-08-24 with fast model b10c384h6nbttflrs at 1,000 visits for each of eight nodes, returning 8,056 visits. Because the model sets differ, comparisons stay within each game.