Nathan Lambert maps post-training's open problems as RL compute surpasses pretraining
rickasaurus · x · 2026-11-22
Nathan Lambert released a slide deck on the biggest open problems in post-training and the field's evolution, outlining Trillium Labs' goals. Key research areas: harness specialization and environment design for context management, character training (model personality, safety, trust), and open-source RL environments. The deck ends with a debate on how RSI is changing research. Accompanying data shows RL compute share exploding: DeepSeek R1 used 5% as much RL as pretraining compute, while Beam (Oct 2026) spent >1.7× more on RL (7.1M vs <4.1M GPU-hours) and Inkling 2 Small 4×.
More from Models
- Gary Marcus clashes over neurosymbolic AI, claims Claude Code is neurosymbolic — mattyp · 2026-11-22
- Nemotron 3.6 Ac Academia Support Lands in llama.cpp: Looped Hybrid with Gated Delta Product — maternion · 2026-11-22
- Unverified: OpenAI morale 'down bad' after GPT-6.2 review flop while Google surges — SumitGup · 2026-11-22
- Gemini 4 Neon Matches Top Models but Costs ~30 Cents Per Task, Same as Fable 5.2 — ChrisGPT · 2026-11-22
- AIs often refuse open math problems, citing low odds of success — Jsevillamol · 2026-11-22
- Mathematician pushes back on Anthropic RH proof skepticism: using everything is a good sign — harris_edouard · 2026-11-22