The story
In March 2016 in Seoul, South Korea, a computer program called AlphaGo, made by the London company DeepMind (owned by Google), beat Lee Sedol, one of the greatest Go players in the world, by four games to one. DeepMind says the match was watched by over 200 million people worldwide. Go, an ancient board game of strategy from China, has more possible positions than there are atoms in the observable universe, and most experts had expected a computer to beat a top human only in about a decade.
Chess was cracked by brute force, when IBM's Deep Blue beat Garry Kasparov in 1997, but Go is too large for that, and good moves depend on intuition about patterns rather than calculation. AlphaGo combined several ideas. Two deep neural networks (layered programs that learn from data), the first suggesting promising moves and the second estimating who is winning, were trained on about 30 million positions from expert games and then improved by playing millions of games against themselves. These guided a Monte Carlo tree search, a method that tests lines of play by running many random continuations. In October 2015 it had already beaten the European champion, Fan Hui, five games to none.
The match produced moments that stunned watchers. In game 2, AlphaGo played move 37, an unusual play on the fifth line that commentators first thought was a mistake, which a human would have been very unlikely to try, and it won. In game 4, Lee found a brilliant move 78 that the program had not expected, and he won the only game humans took. DeepMind then built AlphaGo Zero, which learned with no human games at all and beat the original version, and AlphaZero, which mastered chess and shogi too.
Why it mattered
- AlphaGo Zero and AlphaZero showed that self-play can surpass human knowledge without human examples, shifting how researchers thought about training AI.
- The success drew a surge of money and talent into deep learning, in the West and in China, where it was seen as a 'Sputnik moment'.
- DeepMind turned the same methods to science, leading to AlphaFold's protein structure predictions.
- Professional Go players began studying AI moves, changing opening theory in a game played for 2,500 years.
Sources
Drafted by AI (Claude) from the sources above and checked against them by AI; no historian has reviewed this page. Spot a mistake? Suggest a correction. How pages are made
Community notes
Loading notes…