AlphaGo defeats Lee Sedol (Mar 2016)

What it was

AlphaGo was a Go program built at DeepMind, with David Silver as lead researcher and Demis Hassabis as the company's chief executive. It combined deep neural networks, trained first on human expert games and then by playing against itself with reinforcement learning, with Monte Carlo tree search. Its method was published in Nature on January 27, 2016, with the news that it had beaten the European champion Fan Hui 5-0 the previous October.

The five-game match against Lee Sedol of South Korea ran from March 9 to March 15, 2016 at the Four Seasons Hotel in Seoul. AlphaGo won the first three games, Lee won the fourth, and AlphaGo won the fifth, for a 4-1 result that DeepMind says more than 200 million people watched. In game two AlphaGo played Move 37, which DeepMind says had "a 1 in 10,000 chance of being used" by a human player; in game four Lee answered with Move 78, which DeepMind puts at the same odds.

Read the original

What it changed

The result came sooner than people who built Go programs had predicted. In May 2014 Rémi Coulom, whose program had just beaten a professional with a four-stone handicap, told Wired that a machine would win without a handicap in "maybe ten years." DeepMind now describes the Seoul win as "a decade ahead of its time."

DeepMind then removed the human games. AlphaGo Zero, announced on October 18, 2017, learned only from self-play and beat the version that played Lee "by 100 games to 0." DeepMind said it was "no longer constrained by the limits of human knowledge." Lee retired in November 2019 and told Yonhap, "Even if I become the number one, there is an entity that cannot be defeated."

The arguments it moved

What can AI learn through experience?

AlphaGo, and AlphaGo Zero after it, are the evidence Silver cites for the claim that reinforcement learning through self-play can discover things people do not know. The morning after game two, Silver told Wired's Cade Metz that AlphaGo had estimated a one-in-ten-thousand chance that a human would play it and, drawing on its self-play, played it anyway. Silver had described the system in January 2016 as having "learned to discover new strategies for itself, by playing millions of games between its neural networks, against themselves, and gradually improving." In the 2025 paper "Welcome to the Era of Experience," written with Richard Sutton, he argued that agents learning from their own experience rather than human data will reach superhuman capability.

Compare every leader's position on this

How close is human-level AI?

The match gave a concrete case of a forecast by specialists proving too slow. Coulom's 2014 estimate of about a decade was off by eight years. The win did not settle how close general intelligence is, since Go is a closed game with fixed rules and a clear score. Wired's Cade Metz, writing from Seoul, noted that the same machine "can't carry on a conversation" and "can't pass an eighth grade science test."

Compare every leader's position on this

Positions it bears on

David Silver, Self-play and creativity

Argues that a system learning by trial and error can discover things no human knew, and that self-play is the essence of machine creativity.

Silver led the AlphaGo project, and Move 37 is his evidence that a machine can discover something no human knew.

Critics and counterpoints
David Silver, Limits of learning from human data

Argues that imitating human data can reproduce human competence but not exceed it, and that in mathematics, coding and science the useful human data has largely been consumed.

His argument that human data caps performance traces to AlphaGo Zero, which grew stronger when the human games were removed.

Critics and counterpoints
David Silver, Reinforcement learning as the path to general intelligence

Holds that reward maximisation by an agent acting in its environment is sufficient to account for intelligence, and that reinforcement learning will sit at the core of any general system.

The AlphaGo line of programs is the main empirical support for his view that reward-maximizing agents can reach general intelligence.

Critics and counterpoints
Gary Marcus, Innate structure and reinforcement learning

Argues that systems presented as learning from scratch, such as AlphaGo Zero, depend on structure their designers built in, and that AI should study which innate machinery to include rather than minimize it.

Marcus's January 2018 paper argued that the "tabula rasa" framing of AlphaGo Zero understated the search procedure, architecture and rules its designers built in.

Critics and counterpoints

Sources

Last verified 2026-09-22.