AlphaGo Zero learns from nothing

The original AlphaGo had started from thirty million moves by human experts. The version published in Nature on 18 October 2017 started from nothing. A single network, given only the rules, played itself, and at every move used tree search to improve on its own raw guess; the improved move became the training target, and the game's outcome trained the value. After three days and 4.9 million games it beat the Lee Sedol version 100 games to nil, and after forty days it beat the version that had defeated the world number one, Ke Jie, earlier that year.

It also rediscovered Go. Its self-play games passed through the standard human openings and beyond them, and some joseki that players had used for centuries it abandoned as inferior. Its successor AlphaZero, two months later, learned chess and shogi the same way in hours and beat the strongest engines.

AlphaGo Zero is the purest demonstration on this timeline of Sutton's bitter lesson before he wrote it: remove the human knowledge, add compute and self-play, and the result is stronger. It is also TD-Gammon's method vindicated twenty-five years on, and the origin of the search-plus-learning loop that reasoning models revived for language in 2024.