Lilian Weng

Co-founder and Chief Research Officer — Thinking Machines Lab Distinguished Fellow — Fellows Fund VP of Research and Safety — OpenAI

Listen — profile

0:00 / 3:00

Profile

If you learned how transformers, diffusion models, or LLM agents actually work by reading a blog rather than a textbook, there’s a good chance the blog was Lil’Log. Lilian Weng has spent nearly a decade doing something rare in frontier AI: shipping serious research and then explaining the surrounding field so clearly that her posts became de facto course material. Her write-ups on autonomous agents, prompt engineering, and hallucination are cited in reading lists, onboarding docs, and lecture slides across the industry. For developers learning AI, she is one of the few practitioners whose explanations you can trust to be both current and technically honest.

Her research career is just as substantive. After a bachelor’s from Peking University and a PhD from Indiana University Bloomington — with a stint as a research scientist at Snapchat in between — Weng joined OpenAI in 2017. She started in robotics, technical-leading the famous project that taught a single robotic hand to solve a Rubik’s Cube via reinforcement learning. As OpenAI pivoted to large language models, she founded and led the Applied AI Research team, delivering the plumbing that a generation of developers now takes for granted: the fine-tuning API, the embeddings API, and the moderation endpoints. She eventually became VP of Research and Safety, running the Safety Systems team responsible for keeping production models from going off the rails.

In November 2024 she left OpenAI, part of a broader exodus of senior safety researchers from the company. Three months later she resurfaced as a co-founder — and Chief Research Officer — of Thinking Machines Lab, the startup assembled by former OpenAI CTO Mira Murati alongside other OpenAI alumni like John Schulman and Barrett Zoph. The lab raised one of the largest seed/Series A rounds in AI history at a multibillion-dollar valuation, backed by compute partnerships with NVIDIA (a gigawatt of next-gen Vera Rubin systems) and Google Cloud.

What makes Weng worth following now is that she is building, not just narrating. In May 2026, Thinking Machines shipped TML-Interaction-Small, a 276B-parameter Mixture-of-Experts model (12B active) that processes audio, video, and text in continuous 200ms micro-turns — a full-duplex “interaction model” that listens and speaks simultaneously instead of waiting for you to finish. And she still blogs: her 2026 posts on scaling laws and self-improvement show she hasn’t traded the teacher’s instinct for the founder’s. For anyone learning to build with AI, Weng is a rare two-for-one — read her to understand the field, and watch her to see where it’s going.