When Deep Blue defeated world chess champion Garry Kasparov in 1997 it was treated as a milestone for computer intelligence. But Deep Blue was essentially a giant calculator: a brute-force engine evaluating 200 million positions per second, running on rules written by human masters. Go, an East Asian game 2,500 years old, resisted that approach. The board is 19 by 19 — 361 points, around 250 plausible moves per position, and in a typical 150-move game roughly 10^170 possible game trees. That is more than the number of atoms in the universe (10^80). Classical search was impossible. Most experts thought it would take another ten to twenty years for a computer to beat a human professional.

DeepMind — founded in London in 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman, and bought by Google for 400 million pounds in 2014 — tried a different path. AlphaGo combined three techniques: (1) a "policy" neural network trained on millions of human games — answering "what might be a good move here?" quickly; (2) a "value" neural network — predicting "who is winning from this position?"; (3) Monte Carlo tree search — an algorithm guided by these two networks to scan millions of possible futures fast. On top of all this, reinforcement learning: AlphaGo improved by playing millions of games against itself.

From 9 to 15 March 2016, at the Four Seasons Hotel in Seoul, Lee Sedol of South Korea — one of the strongest players of the twenty-first century, holder of eighteen international titles — played AlphaGo over five games. The prize was a million dollars, the global audience over 200 million, much of it in East Asia. AlphaGo won 4-1. The 37th move of the second game ("Move 37") went into history: professional commentators called it "not in any human style," "a move that looks wrong but turns out deep." In the fourth game Lee Sedol responded with his own historic move 78 ("the Hand of God") and won the only victory for a human — remembered as the last human win against AlphaGo. In retiring in 2019 Lee said he had been playing "an entity that is not unbeatable but cannot be overtaken."

What came next came faster. In October 2017 AlphaGo Zero, given only the rules and no human games at all, played itself for 40 days and beat the original AlphaGo 100-0. In 2018 the same technique was applied to AlphaFold: protein folding, one of biology's grand challenges for fifty years, was solved at near-human accuracy by neural networks; Hassabis shared the 2024 Nobel Prize in Chemistry. The 2017 "Transformer" architecture and AlphaGo's recipe of "big model plus big data plus big compute" then converged; ChatGPT in 2022, GPT-4 in 2023 and the 2024 generation of reasoning models all sit on that line. The five games against Lee Sedol have since been taken by AI historians as the practical opening point of the era of general-purpose AI.