Stanford
academicStanford's AI surface spans SAIL, the HAI institute, and the Center for Research on Foundation Models (CRFM) — the university lab that comes closest to running a frontier-style training operation. CRFM's Marin "open lab" pretrains fully-open models from scratch in JAX on TPUs (Marin 8B, and Marin 32B — the largest fully-open-process university model, beating OLMo 2 32B), with every experiment landing as a public PR on the Levanter training stack, plus the Delphi open scaling suite.
The systems and methods lines are equally load-bearing: Chris Ré's Hazy Research (FlashAttention lineage) now ships the ThunderKittens/HipKittens kernel DSLs; Hashimoto and Liang's groups produced s1 test-time scaling, Alpaca before it, and the pretraining-science pair on optimizers and data efficiency; Stanford NLP maintains DSPy (36K★); OVAL's STORM agent has 30K★; and the MIMI group builds medical imaging FMs (Merlin, Nature 2026). HELM remains the reference holistic eval.
The spin-out boundary is the entry's defining attribution problem: Liang and Ré co-founded Together AI, Fei-Fei Li runs World Labs, Finn co-founded Physical Intelligence, and Chatbot Arena (joint with Berkeley) incorporated as LMArena — company output stays with the companies; Stanford keeps the original dated releases.
People
- Percy Liang Google Scholar — Associate Professor, CS; Director, CRFM; leads the Marin open lab (formerly Co-Founder, Together AI)
- Christopher Ré Google Scholar — Professor, CS; Hazy Research (formerly Co-Founder, Together AI, SambaNova, Snorkel AI)
- Tatsunori Hashimoto Google Scholar — Assistant Professor, CS
- Fei-Fei Li Google Scholar — Professor; Co-Director, HAI; Co-Founder & CEO, World Labs
- Chelsea Finn Google Scholar — Associate Professor, CS; IRIS Lab; Co-Founder, Physical Intelligence