Pretraining progress is mostly from data: 12x compute gains from data vs 3.7x from models
sebkrier · x · 2026-09-09
- Dwarkesh Patel and Jerry Han pretrained combinations of year-representative open model recipes and data corpora from 2019-2025 at up to 1e19 FLOPs, evaluating end capabilities with OLMES rather than fixed-dataset loss.
- Data improvements delivered 12.0x in compute multipliers versus 3.7x from model improvements — data contributed 3.24x more to pretraining progress.
- Gains stack independently: better datasets help every architecture roughly equally and vice versa, with implications for frontier lab economics and the pace of future progress.
Related event: Dwarkesh Experiments: Data Drives Most Pretraining Progress(3 posts)→
More from Research
- HarnessMM cuts agent tokens per task by up to 88% while raising success rates — omarsar0 · 2026-10-26
- Ben Recht: The world needs a truly open language model, and it's easier to build than you think — beenwrekt · 2026-10-26
- New Blog Breaks Black-Box Distillation Into Five Practical Steps — chrisrohlf · 2026-10-26
- DetectorScope turns black-box AI text detectors into auditable ones with sparse autoencoders — _akhaliq · 2026-10-26
- OPST paper shows LLMs self-improve with zero human data, guided only by topic descriptions — _akhaliq · 2026-10-26
- Tencent's HiSTEVE 1.0: hierarchical VLA agent plays Minecraft with native keyboard and mouse, hits 85% SOTA — _akhaliq · 2026-10-26