AlphaGo's Move 37 Explained: The Fifth-Line Shoulder Hit
Move 37 is the 37th move of Game 2 of the 2016 match between AlphaGo and Lee Sedol. Black played a stone at P10: a shoulder hit (a stone placed diagonally next to an opponent's stone, on the side toward the centre) from the fifth line against White's fourth-line stone on the right side. Professional commentators first took it for a misclick. A few hours later AlphaGo had won the game, the move was being called the most creative a computer had ever played, and "Move 37" became shorthand for a machine finding something humans had missed.
This article walks through the position, explains why strong players thought the move was wrong, and checks it with a modern engine. The engine sees it very differently from how it was seen in 2016.
Quick facts
- Game: Google DeepMind Challenge Match, Game 2. AlphaGo (Black) vs Lee Sedol 9p (White)
- Date and place: 10 March 2016, Four Seasons Hotel, Seoul
- Rules: Chinese rules, 7.5 komi (points added to White's score for moving second), 2 hours per player plus three 60-second byo-yomi periods (overtime: each move must be played within 60 seconds)
- The move: Black 37 at P10, a fifth-line shoulder hit
- Result: AlphaGo won by resignation after 211 moves and went on to win the match 4–1
- Full game: replay AlphaGo vs Lee Sedol, Game 2
Before Move 37: a match nobody expected to be close
AlphaGo had beaten Fan Hui, the European champion, 5–0 in October 2015. Most professionals still assumed that Lee Sedol, one of the strongest players of the previous decade, would win comfortably. Then AlphaGo won Game 1.
Going into Game 2, people wondered whether that result was a fluke. Lee had played below his usual level, and AlphaGo had not needed anything spectacular to win. Move 37 settled the question.
Why professionals thought it was wrong
Shoulder hits are an everyday technique. This one shocked people because of its height.
White's stone at Q11 sits on the fourth line from the right edge. When Black shoulder-hits a fourth-line stone from the fifth line, White can settle comfortably: the area behind the stone, toward the edge, becomes secure White territory. In return Black gets a wall facing the open board.
Centuries of professional play had produced a clear verdict on this trade: it favours the player whose stone is hit. Fourth-line territory is big, and influence toward empty space is hard to turn into points. Players normally aim shoulder hits at third-line stones, where the territory they concede is small.
So when P10 appeared, the commentary room saw a move a strong amateur would be told not to play. Michael Redmond 9p, commentating live in English, called it "creative" and "unique", the polite words of someone who has not yet decided whether a move is brilliant or a blunder.
Lee Sedol was away from the board when the move appeared; he had stepped outside for a break. When he came back, he took an unusually long time over his reply.
What Black was actually doing
The move makes more sense in the context of the whole board than as an isolated shape:
Black already had a strong upper-right corner and stones on the lower side at K4 and N4. The shoulder hit sets up a trade:
- White gets the right side. Q11, together with White's stones at Q14 and R14, now holds a secure area along the right edge. This is the territory professionals expected Black to regret giving away.
- Black gets a wall facing the lower side. After 38 and 39, Black's stones at P10 and O10 face down the board toward K4 and N4. The thin lower side becomes a potential framework (moyo: a large area outlined but not yet secured).
- White's stones have to head for the centre. White 40 jumps out to avoid being sealed in. The White stones on the right now form a group White has to look after, while Black's shape stays solid.
AlphaGo put more weight on how the wall would work with its other stones than on the territory it gave away. Human professionals judged the trade locally, fourth-line territory against a wall, and on that scale the verdict was "bad for Black".
How rare was it? The "one in ten thousand" figure
The most quoted statistic about Move 37 came from DeepMind. AlphaGo's own network estimated that a human would play the move roughly one time in ten thousand.
That network had learned to imitate human play from about 30 million positions taken from strong amateur games on the KGS Go Server. So the figure is a guess about what those players would do. AlphaGo played the move anyway, because its search concluded it was good.
A modern engine check
Engines today are far stronger than the 2016 version of AlphaGo, so we asked the question again. We ran the position before move 37 through KataGo (v1.18.2, network b18c384nbt, 2,000 visits) and through KataGo's human-style model, which predicts what a human of a given level or era would play.
- KataGo's raw intuition (the policy network, which suggests moves before any reading) gives P10 about 15%, making it the 3rd most likely move on the board.
- After search, KataGo slightly prefers E12 on the left but rates P10 as practically equal: Black's winning chances come out at about 51.6% vs 50.8%. At 2,000 visits, a gap that small is within the noise.
- The human-style model, set to 2016 professionals, gives P10 about 0.5%, roughly one in two hundred, and ranks it 17th.
The human-style figure is the one that revises the story. Move 37 was an unusual choice for a 2016 professional, but it was nowhere near as rare as the reported one in ten thousand. For a modern engine it is one of the normal candidates.
This fits what happened to professional Go after 2016. Players started studying AI games, and moves that once looked alien became part of ordinary top-level play: early 3-3 invasions (diving into the corner point), high shoulder hits, giving up corners for influence.
How the game ended
Lee answered the shoulder hit well. An Younggil 8p called his counter "exquisite", and for a long stretch the game looked balanced to human eyes.
AlphaGo saw it differently. DeepMind's Demis Hassabis said after the game that the system had been confident of winning from about the middle of the game, while commentators still could not tell who was ahead. Lee resigned after move 211. Afterwards he said AlphaGo had played a nearly perfect game and that he had never felt he was leading.
Why Move 37 is remembered
Our engine check rates Move 37 as roughly equal to KataGo's own first choice, so the move did not win the game on its own. People remember it because it was the first time a large audience watched a computer play a move that looked strange and turned out to be right. It showed that professional Go's conventional wisdom, refined over centuries, had blind spots.
The match has a second famous move, played three days later by Lee Sedol in the only game he won: Lee Sedol's Move 78, the "divine move" of Game 4. For a famous move from before the computer era, read about the Ear-Reddening Game, Shusaku vs Gennan (1846).
You can step through every move in our viewer: AlphaGo vs Lee Sedol, Game 2, full game record. The library also has all of AlphaGo's games and all of Lee Sedol's games.
Sources
- AlphaGo versus Lee Sedol, Wikipedia: match details, commentary quotes, Lee's reaction
- Silver et al., Mastering the game of Go with deep neural networks and tree search, Nature 529 (2016): how AlphaGo's networks were trained on KGS games
- The story of AlphaGo, Google Arts & Culture / Barbican: the one-in-ten-thousand estimate
- Game record: Andries Brouwer's AlphaGo game collection
- Engine analysis: KataGo v1.18.2, run by Go Kifu
About this article
Researched and drafted with the help of an AI assistant, then checked against the game record and the sources above. Go Kifu's board renderer generates the diagrams directly from the SGF game record. The engine figures come from our own KataGo run with the settings listed above; a different network or more visits will shift the exact percentages slightly.