Percy Liang: open weights aren't enough — hidden data mixtures make CS336 impossible without open labs
stanfordnlp · x · 2026-09-09
In Stanford's CS336, Percy Liang unpacks what 'open' really means: weights are only the first layer; AI2, NVIDIA and Marin go further with paper, code and data. He notes Qwen-style papers still hide the data mixture, making runs irreproducible — better than GPT-4's report, worse than a lab notebook — and states flatly the course wouldn't be possible without the open ecosystem: syllabi are downstream of release policy.
Related event: Percy Liang: Open Weights Alone Don't Make Open AI(2 posts)→
More from Companies & People
- Poke's iMessage AI product faces huge platform risk, and Apple will never tolerate it — signulll · 2026-09-09
- Sentry now runs on 5% of 2.1M new sites scanned, founder says that feels low — zeeg · 2026-09-09
- OpenAI to host GPT-6 creator meetup in SF on Sep 16, applications due Sep 10 — gabrielchua · 2026-09-09
- Interpretability researcher Jatin joins Goodfire AI — burny_tech · 2026-09-09
- NUS argues universities should pursue 'X + AI', not 'AI first' — ProfChesterman · 2026-09-09
- NUS professor's new book asks what universities are for in the AI era — ProfChesterman · 2026-09-09