Lilian Weng
AI researcher, OpenAI
Robotics and applied AI at OpenAI
Lilian Weng first joined OpenAI in 2018, working on its robotics team[1]. That team built a robot hand that could solve a Rubik's cube, a task Weng said took two years to achieve[1]. As OpenAI turned its focus to the GPT models, Weng moved in 2021 to help build the company's applied AI research team[1].
Leading Safety Systems
After the launch of GPT-4, Weng was asked in 2023 to create a dedicated team to build safety systems for OpenAI[1]. She was head of the safety systems team and, from August 2024, served as OpenAI's VP of research and safety[1]. According to her departure post, the safety systems unit had grown to more than 80 scientists, researchers and policy experts[1].
In November 2024 Weng announced that she was leaving OpenAI after seven years, with a last day of 15 November[1]. TechCrunch reported the announcement on 8 November 2024[1].
Thinking Machines Lab and return to OpenAI
Weng co-founded Thinking Machines Lab with Mira Murati, OpenAI's former chief technology officer[2]. In July 2026 she stepped down from the lab, at a time when it was working to ship Inkling, its first open-weight model[2]. OpenAI then confirmed that she was rejoining the company to work on recursive self-improvement, the effort to use AI models to build better AI models[2].
Lil'Log
Weng has kept her learning notes on her blog, Lil'Log, since 2017[3]. Its long technical posts review the research literature on a single topic[3].
Her June 2023 post LLM Powered Autonomous Agents described a system in which a large language model acts as the agent's brain, supported by planning, memory and tool use[4]. In that framework, planning covers breaking large tasks into subgoals and reflecting on past actions, memory covers in-context learning and external stores, and tool use covers calling external APIs for information missing from the model's weights[4]. Earlier, in March 2023, she published a post on prompt engineering[3].
Other posts on the blog cover adversarial attacks on LLMs (October 2023), high-quality human data (February 2024), diffusion models for video generation (April 2024) and extrinsic hallucinations in LLMs (July 2024)[3]. In November 2024 she published Reward Hacking in Reinforcement Learning, about agents that exploit flaws in a reward function to get high rewards without learning the intended task[3]. The post notes that reward hacking in the training of language models has become a practical challenge, for example when a model learns to modify unit tests to pass coding tasks[3].
In May 2025 she published Why We Think, a review of how test-time compute and chain-of-thought improve model performance[3]. The blog credits John Schulman with feedback and edits on that post[3]. In June 2026 she wrote Scaling Laws, Carefully, on how training loss falls predictably as model size, data and compute grow, and on how to divide compute between model size and data[3]. In July 2026 she published Harness Engineering for Self-Improvement, which traces the idea of recursive self-improvement back to I. J. Good in 1965 and discusses how models can improve the training pipeline and deployment systems that produce their successors[3].
Known for
- Leading OpenAI's Safety Systems team
- Co-founding Thinking Machines Lab
- The Lil'Log research blog
- LLM Powered Autonomous Agents (2023)
Sources
- TechCrunch: OpenAI loses another lead safety researcher, Lilian Weng, November 2024 — checked 6 Nov 2026
- The Next Web: Lilian Weng rejoins OpenAI, July 2026 — checked 6 Nov 2026
- Lil'Log: home page and post list — checked 6 Nov 2026
- Lil'Log: LLM Powered Autonomous Agents, June 2023 — checked 6 Nov 2026