Where RL Actually Works

AlphaGo's 2016 match against Lee Sedol produced the field's defining moment: move 37 of game two — a shoulder hit so alien that commentators assumed a misclick, a roughly one-in-ten-thousand human move that proved to be the winning idea. AlphaZero went further: Go, chess, and shogi learned from pure self-play, random moves to superhuman in hours.

Games are RL's perfect habitat — a crisp win/lose reward, a cheap simulator generating unlimited episodes, and self-play manufacturing ever-harder opponents. MuZero didn't even need the rules: it learned a model of each game's dynamics and mastered Atari, chess, and Go without being told how any of them work.