- Rl envs + synthetic data + human annotated
- Usage data from codex/claude code/cursor
Most of the model abilities in coding come from post-training, not pretraining
unfortunately all the incentives right now are for repos to be private