Skip to content
AI Digested
History2014 to 20173 min read

AlphaGo and Move 37: when AI mastered the hardest board game

In March 2016, DeepMind's AlphaGo beat Go legend Lee Sedol 4 to 1 in Seoul. The story of the match, the move that stunned the world, and how AlphaGo learned to play.

First published · AI Digested

After Deep Blue beat Kasparov at chess in 1997, one great board game remained a fortress for human intuition: Go. Experts predicted it would take until the mid-2020s for a computer to beat a top professional. It happened in March 2016.

The game

Go was invented in China more than 2,500 years ago. Two players take turns placing black and white stones on a 19-by-19 grid, trying to surround territory. The rules take minutes to learn. Mastery takes a lifetime.

The number of possible board positions is around 10 to the power of 170, vastly more than the number of atoms in the observable universe. Brute-force search of the kind Deep Blue used couldn't cope. Top players often explain their moves by feel: a shape that "looks right".

DeepMind

DeepMind was founded in London in 2010 by Demis Hassabis, a former chess prodigy and video game designer turned neuroscientist, with Shane Legg and Mustafa Suleyman. Its goal was general-purpose AI. Google bought it in 2014.

DeepMind had already built an AI that learned to play dozens of classic Atari video games from the pixels alone. Go was the next great challenge.

How AlphaGo learned

AlphaGo combined several ideas:

  1. Learning from people. A neural network studied about 30 million moves from games between strong human players, learning to predict what an expert would play.
  2. Learning from itself. Using reinforcement learning, AlphaGo then played millions of games against versions of itself, gradually discovering which moves actually won.
  3. Judging positions. A second network learned to estimate who was winning from any position, something nobody had managed to program by hand.
  4. Looking ahead, selectively. A search method explored the most promising moves, guided by the networks, instead of every possibility.

In October 2015, AlphaGo beat the European champion Fan Hui 5 to 0, the first time a computer had beaten a professional at full size without a handicap. The result was published in Nature in January 2016, and a bigger challenge was set.

The match: Seoul, March 2016

Lee Sedol, from South Korea, had won 18 international titles and was one of the finest players of his generation. Before the match he predicted he would win 5 to 0, or perhaps 4 to 1. Around 200 million people watched online.

Game 1: AlphaGo won. Shock.

Game 2: Move 37. On its 37th move, AlphaGo placed a stone on the fifth line from the edge, a "shoulder hit" that professional players would almost never consider at that point in a game. Commentators thought it was a mistake. Lee Sedol left the room for several minutes. AlphaGo's own estimate was that a human would play the move about 1 in 10,000 times. As the game unfolded, it became clear the move was brilliant. AlphaGo won.

Game 3: AlphaGo won, and the match.

Game 4: Move 78. Lee Sedol found an extraordinary move of his own, a wedge in the centre of the board that AlphaGo had not anticipated. The machine stumbled, and Lee won. Fans called it "God's touch". It was celebrated as a victory for humanity.

Game 5: AlphaGo won. Final score: 4 to 1.

After the match

  • 2017: An improved version, AlphaGo Master, beat the world number one, Ke Jie, 3 to 0.
  • 2017: AlphaGo Zero learned Go from scratch, using only the rules and self-play, and beat the version that had defeated Lee Sedol 100 to 0. A general version, AlphaZero, mastered chess and shogi the same way.
  • 2019: Lee Sedol retired from professional Go, saying that even if he became number one, AI was "an entity that cannot be defeated".

Human Go changed too. Professionals now study AI games, and moves once considered bad are now standard.

Why it mattered

AlphaGo showed that AI could master a task requiring something like intuition, and could discover strategies that thousands of years of human play had missed. Move 37 became a symbol of machine creativity.

DeepMind then turned the same approach to science. Its AlphaFold system predicted the 3D shapes of proteins, a 50-year-old problem in biology, and in 2024 Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for it. Read about Google DeepMind.

Read on

Common questions

Why was Go harder than chess for computers?

Go has far more possible moves each turn, around 250 compared with about 35 in chess, so searching ahead is much harder. And it's very difficult to write rules that judge who is winning in a Go position. Strong players rely on intuition.

Did Lee Sedol win a game?

Yes, game four, thanks to a brilliant move of his own, move 78, which commentators called 'God's touch'. It remains the only game a human won against that version of AlphaGo in an official match.

What is AlphaGo Zero?

A 2017 version that learned Go entirely by playing against itself, with no human games at all. It beat the version that defeated Lee Sedol by 100 games to 0.

Sources and further reading

  1. Silver et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature
  2. Silver et al. (2017). Mastering the game of Go without human knowledge. Nature
  3. Google DeepMind: AlphaGo

Spotted a mistake or something out of date? Tell us and we will fix it. Read how we research and write our articles.