Weng Li Returns to OpenAI After Leaving Thinking Machines Lab

Weng Li announced her departure from Thinking Machines Lab on July 27, 2026, in a social media post, stating that July 28 would be her last day. She wrote that over the past seven months, she had been ill more frequently and severely than at any other point in her life, and that the sustained stress and workload had exceeded what her body could handle. She had considered taking a role with narrower responsibilities, but concluded that she could not do her job well if she could not devote full energy to what she was in charge of. She said she still loves AI and wants her next role to have a clearer scope of responsibilities, a more predictable environment, and not be a co-founder position.

On July 29, 2026, OpenAI confirmed that Weng Li will return to the company to lead a high-level team focused on internal research, supporting cross-cutting work including recursive self-improvement. Recursive self-improvement refers to an AI system participating in improving its own training methods, research processes, and even next-generation models, creating a closed loop for continuous capability iteration. Weng Li's early career at OpenAI began in 2018, when she joined the Dactyl robotics project, using reinforcement learning and large-scale domain randomization to enable a robotic hand to solve a Rubik's Cube. Starting in 2021, she led the applied AI research team, working on language model deployment and safety, building harmful content evaluation datasets, content classifiers, and methods to reduce unsafe model outputs. After the release of GPT-4, she integrated the Safety Systems team, covering model behavior, safety evaluations, red-teaming, content moderation, and deployment safety. In August 2024, she was promoted to vice president of research and safety at OpenAI, and left the company in November 2024.

In February 2025, Weng Li co-founded Thinking Machines Lab with former OpenAI chief technology officer Mira Murati and others. The lab's initial direction was to make frontier AI systems more understandable, customizable, and usable, with an emphasis on multimodality, human-machine collaboration, and co-design of research and products. In July 2025, Thinking Machines completed a $2 billion seed funding round led by a16z, with participation from Nvidia, AMD, Jane Street, and others, at a post-money valuation of $12 billion. The company's products include Tinker, a model training and fine-tuning platform; Interaction Models, a research preview supporting real-time input of audio, video, and text; and Inkling, its first open-weight model. Inkling uses a mixture-of-experts architecture with 975 billion total parameters and 41 billion active parameters, supporting text, image, and audio input.

Weng Li has continued to focus on reasoning and model self-improvement, publishing "Why We Think" in 2025 and "Harness Engineering for Self-Improvement" on July 4, 2026.

After Weng Li's departure, only two of Thinking Machines' original six co-founders remain at the company: CEO Murati and chief scientist Schulman. Earlier, co-founder Andrew Tulloch joined Meta, while Barrett Zoph and Luke Metz returned to OpenAI. According to a Business Insider tally, 13 of the initial 42 founding members have left. Murati responded to Weng Li's departure by saying she would miss her, was glad that she was prioritizing her health, and thanked her for her contributions. Weng Li earned her bachelor's degree in information management and information systems from Peking University, where she received a national scholarship, and later earned a PhD from Indiana University Bloomington, focusing on network science, complex systems, and information diffusion in social networks. She is also the author of the technical blog Lil'Log, which systematically covers frontier technologies such as reinforcement learning, Transformers, and agents.

In November 2025, Thinking Machines attempted to launch a new funding round at a valuation of $50 billion to $60 billion, but related negotiations were not completed by January 2026, and the market valuation returned to around $12 billion. The company has reached a multi-year strategic partnership with Nvidia to deploy at least one gigawatt of next-generation Vera Rubin systems starting in 2027, and has signed a multi-billion-dollar computing capacity agreement with Google Cloud.