Lilian Weng returns to OpenAI for recursive self-improvement research

Why it matters

OpenAI is reclaiming a third Thinking Machines cofounder and assigning Weng to research that could shorten the cycle for designing, training and evaluating successor models.

Lilian Weng (@lilianweng) is returning to OpenAI to work on using AI systems to develop new models, according to reporting by The Information on Wednesday.

The move takes Weng back to the lab where she spent nearly seven years working across robotics, applied AI and model safety. She left OpenAI in November 2024, months after becoming vice president of research and safety, and went on to cofound Thinking Machines Lab with former OpenAI chief technology officer Mira Murati.

Weng's return follows her decision this week to leave Thinking Machines Lab. She cited the health toll of startup work, including sustained stress and repeated illness, and said July 29th would be her final day. Thinking Machines Lab had demanded a founder's broad responsibilities and uncertain schedule; Weng said she wanted clearer boundaries and a more predictable working environment.

OpenAI is giving her a technically ambitious assignment. Stephanie Palazzolo (@steph_palazzolo), who reported the move with Amir Efrati, described the work as recursive self-improvement: using AI to help create better AI models.

Weng had already laid out a detailed version of that research agenda. In a July 4th essay on her technical blog, she wrote that recursive self-improvement can extend beyond a model directly rewriting its own weights. An AI system could instead improve the training pipeline, tools, evaluation methods and deployment software used to produce a successor model.

That framing makes the surrounding engineering system, often called the harness, central to the work. The harness controls how a model plans, uses tools, manages context, records results and evaluates its own attempts. Automating parts of that loop could let researchers run more experiments and refine training methods faster, while still leaving people responsible for objectives, evaluation and compute allocation.

Weng's background spans several of the disciplines needed for that work. She contributed to OpenAI's Dactyl research, including the project that trained a robotic hand to solve a Rubik's Cube, before moving deeper into model safety. She later built and led OpenAI's Safety Systems organization, which worked on evaluations and safeguards for deployed models. Her Lil'Log technical essays have also covered reinforcement learning, reward hacking, autonomous agents, scaling laws and test-time reasoning.

The hire continues OpenAI's recruitment of people who left to help build Thinking Machines Lab. In January, Thinking Machines Lab cofounders Barret Zoph and Luke Metz returned to OpenAI, alongside researcher Sam Schoenholz. Weng is the third Thinking Machines Lab cofounder to make that trip back in 2026.

That talent flow matters because Thinking Machines Lab was financed largely on the record of Murati and the former OpenAI researchers around her. Thinking Machines Lab raised a $2 billion seed round at a $12 billion valuation in 2025, before releasing a model or product. The funding was among the largest early-stage bets placed on an AI research group whose primary asset was its staff.

Thinking Machines Lab began putting products behind that bet with Tinker, its model-customization platform, and Inkling, an open-weights model released on July 15th. Thinking Machines Lab has positioned its research around customizable models and human-AI collaboration rather than competing solely for the top general-purpose benchmark score.

Weng's exit removes a founder with experience spanning model development, safety systems and research management shortly after that first model release. Her destination compounds the loss: OpenAI is reclaiming researchers from a lab founded partly as an alternative home for OpenAI alumni.

For OpenAI, the hire adds an experienced research leader to a field that frontier labs increasingly view as a way to accelerate model development. OpenAI recently published work using self-play to train automated safety red-teamers, another example of AI systems generating and evaluating work that feeds into subsequent training.

The Information's report does not specify Weng's title, reporting structure or start date. The stated mandate, however, closely matches the technical direction she has been developing publicly: closing more of the model-development loop with AI while confronting the evaluation, memory and reward-design problems that can make automated improvement unreliable.