Lilian Weng Rejoins OpenAI for Recursive Self-Improvement

The interesting part of Lilian Weng returning to OpenAI is not the return itself. Researchers move; that is not news. The interesting part is what she has been assigned to work on, and how quickly the round trip happened.

According to The Information, Weng announced July 29 as her last day at Thinking Machines Lab and is heading back to OpenAI to work on recursive self-improvement, using AI to speed up how OpenAI designs, trains and evaluates its own future models. She left OpenAI in November 2024 after nearly seven years there, where she had been VP of AI Safety. A post she published on her personal blog, Lil'Log, only weeks before quitting Thinking Machines was on exactly this topic.

Why this matters if you are not tracking cofounder gossip: recursive self-improvement is the workflow the frontier labs quietly care about most, because it is the thing that starts to compound. If OpenAI can put a senior safety-trained researcher on automating parts of its own model development, and rivals cannot, the gap widens on something other than raw compute. This is a hire with a roadmap attached.

The other thread is Thinking Machines itself. Mira Murati, formerly OpenAI's chief technology officer, cofounded the lab in February 2025 and later closed a reported $2 billion seed round led by Andreessen Horowitz at a $12 billion valuation. Weng is now the fourth of the six original cofounders to leave. Barret Zoph and Luke Metz departed earlier in 2026, and by the reporting she is the third of the cohort to have returned to OpenAI this year. That is a pattern rather than a coincidence, and it is starting to be the story around the lab as much as any product it has shipped.

The honest caveat is that Weng's stated reason is her health. She wrote that she had been sick more in the past seven months than at any prior point in her life, and pointed to the unpredictable schedule and open-ended responsibilities of a startup cofounder. What the reporting does not give you is whether her new OpenAI role sits inside safety or the applied research group she used to run, who inherits her Thinking Machines work, or whether other cofounders are quietly weighing the door.

What to watch: whether OpenAI puts recursive self-improvement on a public roadmap now that it has a senior researcher pointed at it, and whether Andreessen Horowitz and the rest of Murati's cap table start asking harder questions about retention before the next round.