When Should Coding Models Switch from NTP to RL? NTP Overfits Fast, RL Scales
brianryhuang · x · 2026-10-15
A technical discussion on coding-model training: how to pick the tipping point from next-token prediction (NTP) to RL. Suggested recipe: use pass@k on a forward-looking holdout of PR/issue SWE tasks plus general evals as the switch metric; midtrain/SFT data (commit diffs, issue/PR threads, fork diffs, traces) can absorb 1-2 orders of magnitude more compute than snapshot NTP before overfitting, while RL is where compute can scale unboundedly — with task synthesis as the key lever.
Related event: When Should Code Models Switch from NTP Pretraining to RL?(2 posts)→
More from coding & agent
- Distill a cheap specialized model per game NPC instead of using frontier LLMs, argues researcher — QuintinPope5 · 2026-10-15
- Entire Minecraft Crammed Into Elden Ring as an AI-Built Mod — DeryaTR_ · 2026-10-15
- Indie dev builds a MOBA in one week with Opus 5.5: 24 heroes, 3 battlegrounds — TAbrodi · 2026-10-15
- AI-built MOBA hits 13.1K plays and 11.3K players in its first week — TAbrodi · 2026-10-15
- Paige Bailey: Ship an Eval People Care About and You Own What Labs Hill-Climb On — DynamicWebPaige · 2026-10-15
- Voice-braindump to PRs: one dev's four-step workflow routing tasks across ten cloud agent projects — vinvan · 2026-10-15