Nathan Lambert's Cornell deck maps the biggest open problems in open LLM post-training
natolambert · x · 2026-11-21
Nathan Lambert (Executive Director, Trillium Labs) published his Cornell talk slides, "Pushing Frontiers in Open Language Model Post-Training" (natolambert.com/cornell-2026). Key points:
- Resource squeeze: full frontier-grade post-training is becoming as inaccessible as pretraining looked in early 2023 — solvable but demanding major funding and commitment.
- Literature gaps: open post-training research has years-wide empty space in ideas that will matter 5+ years out — an exciting opportunity.
- Structure: a history of the field and its three optimization pillars (SFT, RLHF/preference finetuning, RLVR), how he picks projects, where AI is heading, and how RSI is changing this research — framed as a statement of Trillium Labs' goals.
Cites o1 AIME accuracy scaling with train-time and test-time compute (OpenAI, Sep 2024).
Related event: Nathan Lambert shares 51-page slides on open post-training problems(2 posts)→
More from AGI Musings
- David Bessis: a 2M-line Lean Riemann proof would collapse the 'theorem economy' — beenwrekt · 2026-11-21
- Narayanan: labs chasing RSI may be racing toward 'recursive self extinction' of their business model — random_walker · 2026-11-21
- Professor reports 5x surge in PhD inquiries, suspects AI boom is the cause — Limor_Raviv · 2026-11-21
- AI commentator: human knowledge institutions were already rotten before AI arrived — teortaxesTex · 2026-11-21
- Nathan Lambert lays out six open problems in post-training in new Cornell slide deck — natolambert · 2026-11-21
- Fake AI breakthroughs debate: critics say hyped claims are freezing academic research — teortaxesTex · 2026-11-21