Blog

Move 78, After Three Losses: The Day Lee Sedol Beat AlphaGo

Published

Contents
  1. Act 1. Zero to three
  2. Act 2. The fourth game of a man who had already lost
  3. Act 3. Move 78
  4. Act 4. The machine wavers
  5. Act 5. A win without a price
  6. Epilogue. Four to one
  7. If this makes you want to sit down at a board

Seoul, the Four Seasons Hotel. March 13, 2016.

When Lee Sedol walked into the playing room that day, nothing was at stake for him. He had already lost three of the five games. The match and the prize belonged to the other side.

Only one question was left. Could a human beat this machine even once?

Act 1. Zero to three

The Google DeepMind Challenge Match began on March 9: five games of Go between a human and an artificial intelligence.

The mood before the match was nothing like it is today. Go has so many possible positions that most people believed machines were still years away from beating a top professional. That belief lasted three games.

Game one, game two, game three. AlphaGo never wavered. It did not tire and it did not rush. Whenever it was ahead, it chose the path that won by a little, safely. After three games the score was 0–3, and the match was decided.

After the third loss, Lee said he had misjudged AlphaGo’s abilities and had felt powerless. This came from one of the strongest Go players in the world.

Act 2. The fourth game of a man who had already lost

In game four, Lee Sedol played White.

He changed his approach. AlphaGo was at its best piling up small advantages. So a game of small exchanges could not be won. Lee let his opponent build a large framework in the center while he took solid territory along the edges. Then he steered toward a game where everything would ride on a single fight. If it worked, he would win. If it did not, his position would simply fall apart.

For a while the game seemed to be going just the way AlphaGo wanted. Black’s center kept growing thicker. To those watching, White was being pushed back once again.

Act 3. Move 78

Then a single white stone landed in the middle of the board.

Move 78. A wedge, forced in between Black’s stones.

The Chinese professional Gu Li called it a “divine move.” It was a brilliant tesuji that turned the whole game around.

Act 4. The machine wavers

From move 79, AlphaGo was different.

Its own estimate of its winning chances dropped sharply. And the machine that had not slipped once in three games began to play moves no human professional would play. Black’s moves from 87 to 101 were later described as “very bad.” Moves that lost points. Moves that made no sense.

For the first time, people outside the playing room began to consider another possibility. He might actually win.

Lee did not waver to the end. After the 180th move, AlphaGo resigned.

White wins by resignation.

Act 5. A win without a price

The score became 1–3. Nothing had changed. The match still belonged to AlphaGo.

But it was not the score people celebrated that day. After the game, Lee called it a priceless win that he would not exchange for anything. Then he asked to play Black in the final game. Having won with White, he wanted to try it with Black.

Epilogue. Four to one

The final game, two days later, was close. AlphaGo made a serious mistake early but turned the game around in the end, and the match went into the record books at 4–1.

AlphaGo won every game of the match except the fourth. The human side took exactly one.

That one game is still talked about because the man who won it had already lost. With nothing left to play for, he changed his approach, staked everything on one fight, and put a stone where no one else had looked.

In an age when the machines have only grown stronger, move 78 is the first story people reach for when they explain why humans still sit down at the board.

If this makes you want to sit down at a board

Try a game against the AI in Gomoku (Omok), played with the same board and stones as Go. On Gamengager you can play right away with no sign-up and choose from three difficulty levels.

Sources