Self-Play Pretraining with Zero Data: Models Learn From Self-Generated Compute Alone

burny_tech · x · 2026-09-27

A paper from Tel Aviv University, Stanford and others, "Self-Play Pretraining with Zero Data", proposes pretraining without human-curated data: two models train in tandem from random initialization. A generator proposes programs interpreted by a universal Turing machine to emit byte sequences, while a learner predicts those sequences with standard cross-entropy. The generator is updated via RL to target sequences at the frontier of the learner's ability—neither solved nor random—forming an adaptive curriculum, effectively searching the space of all computable structure (inspired by Solomonoff induction).

Remarkably, despite never seeing natural data, the learner improves zero-shot on unseen text, images, audio and code, and even develops in-context learning. Zero-shot loss scales predictably with self-play compute, suggesting general predictive structure can be discovered from self-generated computation alone, making training data limited by compute rather than human knowledge.

Related event: Self-Play Pretraining with Zero Data: Learning from No Real-World Data(11 posts)→

Original post →

More from Research

Research channel →