In March 2016, DeepMind’s AlphaGo program played a five-game match against Lee Sedol, one of the strongest professional Go players in the world. AlphaGo won four of the five games. The result was a milestone that had been considered far away as recently as a year earlier. Most experts in both AI and Go had estimated that defeating a world-class human player was at least a decade away. AlphaGo did it in 2016.
The technical achievement was substantial. Go had resisted previous AI approaches in ways that chess had not. The branching factor of possible moves was too large for brute-force search to be practical. The evaluation of board positions, the question of whether one player was winning or losing in a given configuration, was harder to encode in the heuristic functions that had worked for chess. AlphaGo combined deep neural networks for board evaluation with Monte Carlo tree search for move selection, using self-play to refine the networks. The combination worked in ways that previous Go programs had not approached.
The match produced a moment that became its own reference point in the history of AI. In the second game, AlphaGo played move 37, a stone placed in a position that human Go commentators initially described as a mistake. They went quiet on the broadcast for a long stretch as they tried to understand what AlphaGo was doing. Lee Sedol took fifteen minutes to respond. The move turned out to be central to AlphaGo’s win in that game.
What made move 37 significant beyond the immediate game was what it demonstrated about the kind of knowledge AlphaGo had developed. The move was not played by any human professional in that situation, not because human players could not play it, but because human intuition said it was wrong. AlphaGo, having trained on millions of games and refined its strategies through self-play, had developed a model of Go that included strategies humans had not discovered in two thousand years of playing the game.
The reaction in the Go community was complicated. There was loss in the recognition that a domain humans had considered to require uniquely human intuition could be played better by a machine. There was also genuine appreciation for what AlphaGo was demonstrating about the game itself. The strategies AlphaGo employed, including move 37 and similar ideas, became study material for human professionals in the months and years that followed. The game changed because the way it could be played had been visibly expanded.
The match was a milestone in AI capability. It was also a milestone in what artificial intelligence could discover about old problems by approaching them differently.