Go Kifu

Lee Sedol's Move 78: The "Divine Move" Against AlphaGo

Published October 17, 2026

Move 78 is White's 78th move in Game 4 of AlphaGo vs Lee Sedol, played on 13 March 2016. Lee placed a wedge (a stone pushed between two enemy stones) at L11, in the middle of AlphaGo's central framework. AlphaGo answered badly, its play fell apart over the next twenty moves, and it resigned at move 180. This was the only game Lee Sedol won in the match. The Chinese professional Gu Li called move 78 a "divine move", and the name has stuck.

The usual story goes like this: a human found a move the machine had not seen. A modern engine adds a few complications to that story.

Quick facts

The situation before Move 78

By Game 4 the match was already decided. AlphaGo had won the first three games, including Game 2 with its famous Move 37. After Game 3, Lee said he had misjudged AlphaGo's strength and felt powerless. The open question was whether a human could win even one game.

In Game 4 Lee took territory on the sides and let AlphaGo build a large framework (moyo: a loose area of potential territory) in the centre. If that area turned into solid territory, the game would be over. Lee had to break into it, and on move 78 he chose the most severe way to do so.

78AABBCCDDEEFFGGHHJJKKLLMMNNOOPPQQRRSSTT1919181817171616151514141313121211111010998877665544332211
White 78 at L11 wedges between Black's central stones and threatens to cut them apart.

What the wedge does

A wedge goes between two of the opponent's stones that are not yet solidly connected. L11 sits in the gaps of Black's central group and creates several cutting points at once: if Black defends one, White cuts at another. The tactic works only if every follow-up has been read out correctly, and that means reading many variations deep on a professional's clock.

Gu Li 9p said he had not foreseen the move at all. David Ormerod of Go Game Guru called it a brilliant tesuji, a clever tactical move that exploits a weakness in the opponent's shape.

Where AlphaGo actually went wrong

AlphaGo answered with 79 at K10, its first clear mistake. During the game, DeepMind's Demis Hassabis wrote that the mistake was on move 79, but that AlphaGo only realised it around move 87. According to DeepMind, AlphaGo still put its winning chances at about 70% at move 79. By move 87 that estimate had collapsed, and a run of weak moves followed.

81827980AABBCCDDEEFFGGHHJJKKLLMMNNOOPPQQRRSSTT1919181817171616151514141313121211111010998877665544332211
Moves 79–82. KataGo prefers Black L10 to AlphaGo's 79 at K10; White keeps up the pressure with 80 and 82.

From move 87 to about move 101, AlphaGo played moves that commentators struggled to explain. They were later understood as the program reacting to a sudden drop in its own evaluation.

8789909192939495889697AABBCCDDEEFFGGHHJJKKLLMMNNOOPPQQRRSSTT1919181817171616151514141313121211111010998877665544332211
Moves 87–97. AlphaGo's moves on the right side (89, 93) and 97 in the lower left are the ones that cost it the game.

A modern engine check: was Move 78 really "divine"?

We analysed the game with KataGo (v1.18.2, network b18c384nbt). Key positions got 2,000 visits each (a visit is one position the engine explores in its search), and a faster 300-visit scan covered every move from 78 to 140. We also used KataGo's human-style model, set to imitate 2016 professionals. Here is what the numbers show.

1. Lee was losing badly before move 78. In the position before the wedge, KataGo gives Black (AlphaGo) about a 97% chance of winning and a lead of about 14 points. AlphaGo had played the first 77 moves very well.

2. L11 was the best move, but it did not save the game on its own. KataGo's top choice in that position is exactly L11, so Lee found the move a superhuman engine would play. Even after L11, though, KataGo still has Black at about 96% and 13 points ahead. The wedge was the most testing try, and Black had a calm answer at L10 that keeps a winning position. The human-style model, which estimates how likely a 2016 professional was to play each move, gives L11 only about 0.4%, roughly one in 250, and ranks it 27th. Very few professionals would have found it.

3. AlphaGo's 79 hurt, but the game was lost later. AlphaGo's 79 at K10 was only KataGo's 4th choice and cost about three points, yet Black was still well ahead: around 85% after move 88. The real collapse came on the right side. Black 89 (R9) dropped AlphaGo's chances to about 60%, Black 93 (S9) handed the lead to White, and 97 (C4) left Black clearly behind, about seven points down. From there Lee never let go.

In short, Lee Sedol found the best practical chance in a losing position, and AlphaGo misread it. Move 78 created a fight complex enough for AlphaGo's 2016 search to get lost in. That is why it worked, rather than because it was a combination Black could not stop. After the match, some professional commentators questioned whether the wedge works against the best defence. Modern engines say it does not.

This takes nothing away from the move. Lee found L11 under match pressure, after three straight losses, and reading like that is what made him one of the greatest players of his generation.

How the game ended

After move 97 White was clearly ahead, and Lee played the rest of the game precisely. AlphaGo resigned after White's 180th move.

Final position after 180 moves. AlphaGo resigned.

At the press conference Lee called it a priceless win that he would not exchange for anything. No human won another official game against AlphaGo. In 2017 a stronger version beat the world's top-rated player, Ke Jie, 3–0, and an online version won 60 straight games against top professionals. Lee Sedol retired in November 2019, saying that AI had become an entity that could not be defeated.

Keep exploring

Sources

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. Other networks or longer searches will shift the exact percentages, but not the overall picture.

← All articles