Self-play pretraining: LM learns from random init with zero real data, predictable scaling
burny_tech · x · 2026-09-27
Self-Play Pretraining with Zero Data: two models start from random initialization—a generator proposes programs for a universal Turing machine and a learner trains on their outputs, never touching real data. Zero-shot val loss on images, text, audio, and melodies decreases predictably with self-play compute, and the learner develops in-context learning. Commenters highlight the clean AdamW-preconditioned gradient inner product metric for data value, and ask about similar work in online data selection.
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