In self-play, the opponent is the system itself. Every game produces new data, and since the rival improves at the same pace, difficulty adjusts itself: an automatic curriculum. It is a variant of reinforcement learning in which the reward is winning.
Lineage. Arthur Samuel's checkers program (1959) already learned by playing against itself; M. Mitchell (2019) uses it to introduce three all-important concepts: the game tree, the evaluation function, and self-play (2019, ch. 9). TD-Gammon (Tesauro, 1995) reached close to the level of the best backgammon players this way. AlphaGo combined human games with a phase of self-play against earlier versions of itself (Silver et al., 2016, pp. 485–486). AlphaGo Zero dropped human data and all knowledge of Go beyond the rules: the program "becomes its own teacher" and beat the version that had defeated Lee Sedol 100–0 (Silver et al., 2017, p. 354).
What it shows and what it does not. For Buckner (2024), the fact that a system trained only by self-play outperforms one that used human game libraries suggests that its strength comes from extracting abstract patterns from experience, not from brute force or expert knowledge (2024, p. 53). DeepMind spoke of learning tabula rasa; Gary Marcus objected to the Monte Carlo search and to the hyperparameters set in advance for Go, and Buckner finds only the latter suspect (2024, pp. 4, 20). Inefficiency remains: some 30 million games against itself, compared with the fewer than 50,000 Lee could have played in his lifetime (2024, p. 144).
The method needs a closed world: known rules, a cheap simulator, and an unambiguous criterion of victory. That is why it is hard to carry over to language, where nobody "wins" a conversation; the closest thing is the verifiable rewards in mathematics and programming used by reasoning models (Guo et al., 2025). How much of the success is due to learning, and how much to the fact that the rules were given?