Self-play pretraining with zero data: LMs learn to generate all their own training data
inductionheads · x · 2026-09-26
Stanford's Noah Goodman group introduces Self-Play Pretraining with Zero Data: two models start from random init — a generator proposes programs for a universal Turing machine, a learner trains on its outputs, with no real data at all. Zero-shot val loss on images, text, audio and melodies falls predictably with self-play compute, and the learner develops in-context learning. Goodman adds a philosophy angle: distinguishing universal from contingent facts lets models self-generate universal facts for pretraining.
More from Research
- Explosion launches Ellf directory: 52 NLP templates plus multi-agent annotation workflow — _inesmontani · 2026-11-11
- STABLE Optimizer Accepted at NeurIPS 2026: Taming Continual Learning Forgetting at the Optimizer Level — v_lomonaco · 2026-11-11
- From Bahdanau to FlashAttention: a timeline of attention's 8-year evolution — cneuralnetwork · 2026-11-11
- Three years building CPUs and spiking nets: one engineer's path to AI accelerators — k7agar · 2026-11-11
- When AI becomes an oracle, what happens to mathematics? — stevenstrogatz · 2026-11-11
- 200 Years of Linear Algebra in One Chart: From Leibniz's Determinants to Peano's Vector Spaces — burny_tech · 2026-11-11