AlphaGo Defeats World Champion

In March 2016, a computer program called AlphaGo, developed by the London-based AI company DeepMind, defeated Lee Sedol of South Korea, one of the greatest Go players of his generation, in a five-game match in Seoul. The program won four games to one. Because Go had long been regarded as far harder for computers than chess, the result was widely seen as a landmark for artificial intelligence and a vivid demonstration of what deep learning could achieve.

Go, an ancient board game that originated in China more than two thousand years ago, is played on a 19-by-19 grid with black and white stones. Its rules are simple, but the number of possible positions is astronomically large, and judging who is ahead often depends on intuition about patterns and shape rather than calculation alone. The brute-force search methods that allowed IBM's Deep Blue to beat Gary Kasparov at chess in 1997 did not scale well to Go. For years, the best Go programs played at strong amateur level, and many experts expected a decade or more to pass before a machine could challenge top professionals.

DeepMind, acquired by Google in 2014, combined several techniques. AlphaGo used deep neural networks to suggest promising moves and to evaluate board positions, and paired them with Monte Carlo tree search, a method that explores possible futures by sampling. The networks were first trained on records of games played by strong human players, then improved further through self-play and reinforcement learning. In October 2015, AlphaGo had already beaten the European champion Fan Hui in a private match, which was announced in early 2016.

The match against Lee Sedol was broadcast online and followed by an enormous audience, especially in East Asia, where Go is a major cultural pastime. AlphaGo won the first three games, securing the series. Lee won the fourth with an inventive move that is often called a "divine move" by commentators, which the program appeared to misjudge. AlphaGo took the fifth game. Observers noted that the machine made moves that seemed strange or even beautiful to professionals, prompting fresh thinking about how the game could be played.

The legacy extended well beyond Go. In 2017 DeepMind introduced AlphaGo Zero, which learned entirely from self-play without human game records, and later AlphaZero, which handled chess and shogi as well. Lee Sedol retired from professional play in 2019, citing the strength of AI. The match is often credited with boosting public and commercial interest in machine learning, and it foreshadowed later DeepMind work such as AlphaFold, which tackled protein structure prediction.