Lilian Weng Returns to OpenAI to Lead Recursive Self-Improvement Research

Lilian Weng is heading back to OpenAI, just months after co-founding Mira Murati's startup Thinking Machines. OpenAI confirmed the move on Wednesday, closing a loop that began with Weng's own departure announcement on July 27.

Weng cited health reasons when she stepped away from Thinking Machines. Murati responded publicly to that announcement, voicing support for her co-founder's decision to prioritize her wellbeing over the demands of running a fast-moving AI lab.

At OpenAI, Weng will not be returning to her old post. She previously served as the company's VP of AI Safety Research, a role focused on aligning powerful models with human oversight. This time, she takes charge of a new top-level team dedicated to a far more ambitious mandate: accelerating internal research into recursive self-improvement.

Why Recursive Self-Improvement Matters

Recursive self-improvement describes a theoretical process in which an AI system modifies or retrains itself to become progressively more capable, without waiting on the usual human-driven research and engineering cycle. It's a concept that has long sat at the center of debates about how advanced AI development could accelerate, and how quickly capabilities might compound once a system can meaningfully contribute to its own improvement.

Recursive self-improvement would allow an AI system to iterate on itself to become more powerful or capable.

Placing a researcher with Weng's safety background at the head of this effort is a notable signal. Her tenure leading AI Safety Research at OpenAI gave her direct experience wrestling with the risks of increasingly autonomous systems, and now she's positioned at the center of a research direction that many consider one of the most consequential, and most fraught, in frontier AI.

A Short but Telling Detour

Weng's stint at Thinking Machines was brief. The startup, co-founded with former OpenAI CTO Mira Murati, has positioned itself as a serious contender in frontier model research since launching, drawing significant attention and talent from established labs.

Her return underscores how fluid the senior research talent pool remains across the top AI labs, even for figures who helped build a rival's founding team. It also suggests OpenAI is treating recursive self-improvement as enough of a priority to recruit a proven safety-focused leader to run it at the highest level, rather than folding the work into an existing division.

No specifics have emerged yet on team size, timeline, or how this research group will interface with OpenAI's existing safety and alignment functions. But the appointment itself signals where the company sees its next major research frontier.