Lilian Weng

Lilian Weng joined OpenAI in 2018 on its robotics team and was one of the authors of the October 2019 paper in which a robot hand, trained entirely in simulation with reinforcement learning, solved a Rubik's Cube. When OpenAI turned to the GPT models, Weng helped build its applied AI research team in 2021 and was listed as applied research lead on GPT-4. After GPT-4's release in 2023 Weng was asked to set up a team to build the safety systems for OpenAI's products; by late 2024 Safety Systems had more than 80 scientists, engineers and policy experts. Weng became vice president of research and safety in August 2024 and left on 15 November 2024, writing that "after 7 years at OpenAI, I feel ready to reset and explore something new."

Outside the lab, Weng has written Lil'Log since 2017, survey posts that often run past 30 minutes of reading, each with a long reference list. "LLM Powered Autonomous Agents" (June 2023) described an agent as a language model wrapped with planning, memory and tool use, and "Reward Hacking in Reinforcement Learning" (November 2024) catalogued how models game their training signals.

In February 2025 Weng co-founded Thinking Machines Lab with Mira Murati, John Schulman and others from OpenAI. Weng stayed through the release of the lab's open-weight Inkling model in July 2026, then announced in late July that 29 July would be the last day, writing, "I don't feel I'm able to continue at the pace a startup requires." Days later OpenAI told TechCrunch that Weng was rejoining to lead a top-level team to speed up OpenAI's internal research, including work on recursive self-improvement. Weng had published a long post on the subject, "Harness Engineering for Self-Improvement," on 4 July 2026.