The clearest way to read the AI talent war in 2026 is to watch where a handful of senior researchers choose to work, and what they are hired to do. This week the signal was sharp. OpenAI rehired Lilian Weng, a co-founder of Thinking Machines and a longtime OpenAI research leader, to head a new team focused on recursive self-improvement. The role names the prize the labs are actually competing for.
OpenAI
Anthropic
Google
MetaWhy one hire is news
Lilian Weng stepped down from OpenAI after more than six years, in roles including vice president of research and safety, to help start Thinking Machines. Days after leaving that startup, citing the pace it demanded ("I don't feel I'm able to continue at the pace a startup requires"), she returned to OpenAI. What makes the move more than a personnel note is the mandate: leading internal research on recursive self-improvement, the idea that AI systems can help build their own successors.
I don't feel I'm able to continue at the pace a startup requires.
Lilian Weng, on leaving Thinking Machines
Her departure also underlines how hard even well-funded startups find it to hold founding talent. Thinking Machines, started by former OpenAI chief technology officer Mira Murati, has now seen multiple co-founders leave within a year of its launch, with Murati publicly wishing Weng well.
What recursive self-improvement means
Recursive self-improvement, often shortened to RSI, describes a loop rather than a single capability. Today's models already assist with research: writing code, proposing experiments, and evaluating results. The hypothesis is that this assistance helps produce a stronger successor model, which is then a better research assistant, which helps build the model after it. The quantity labs care about is not any one model but the rate at which each generation arrives.
What recursive self-improvement actually describes
The idea is a loop in which AI systems help build their own successors, so each generation arrives faster than the last. It is a research direction and a hypothesis, not a demonstrated result.
Skeptics note the loop can also stall: gains may shrink each round, or hit limits in data, compute, or verifiable feedback. Whether it compounds is precisely the open question.
It is worth being precise about status. RSI is a research direction and a hypothesis, not a demonstrated result. The loop could compound, or it could stall as gains shrink each round or run into limits in data, compute, or verifiable feedback. That uncertainty is exactly why a lab would staff a dedicated team against it rather than assume it happens on its own.
The war is really about a few people
The Weng hire sits inside a broader reshuffling of senior researchers across the frontier labs. Earlier in 2026, Noam Shazeer left Google for OpenAI, and Nobel laureate John Jumper left Google DeepMind for Anthropic. Meta spent heavily to recruit for its superintelligence group, only to see several prized hires leave soon after, some of them for OpenAI.
Senior researchers changed sides in 2026
A sample of the reported moves among the frontier labs. The through-line is that a small number of people are treated as decisive to a lab's trajectory.
Meta also spent heavily to recruit for its superintelligence group, and saw several prized hires leave soon after, some for OpenAI. Reported moves; roles and timing per cited coverage.
The scale of the hiring is its own data point. OpenAI has said it plans to roughly double its workforce, from about 4,500 to 8,000, by the end of 2026. When a field moves this fast, the scarce input is not capital or compute alone but the small number of people who can direct how those resources are spent.
Reading the signal without overreading it
Two cautions keep this in proportion. First, individual moves are noisy. People change jobs for health, family, money, and fit, and Weng named health and pace, not a grand thesis, as her reason for leaving the startup. Second, naming a team after recursive self-improvement is a statement of intent, not evidence that the loop works. The honest position is that the labs are betting on RSI, and the bet is unresolved.
What the reshuffling does show is a market in which the frontier is set by a concentrated pool of talent, and where the most valuable models change hands quickly. For everyone building on top of these systems, that argues for flexibility rather than allegiance. A model-agnostic approach, which is how we think about model choice at Metir, treats today's leader as provisional, because the people most likely to change the leaderboard just changed desks. In a year when the researchers move this often, the models will too.
Sources:
- OpenAI, Anthropic, Meta, Thinking Machines fight for AI talent | Axios
- OpenAI Rehires Thinking Machines Co-Founder Days After Startup Exit | Yahoo Finance
- OpenAI brings back Lilian Weng after Thinking Machines exit | TechBriefly
- Former Thinking Machines Co-founder Lilian Weng Returns to OpenAI to Research Recursive Self-Improvement | ABAB News
Image credits
- Hero: The Pioneer Building, San Francisco, OpenAI's headquarters. Photo by HaeB, licensed CC BY-SA 4.0, via Wikimedia Commons.