Recursive Superintelligence launched out of stealth on Wednesday with $650 million in funding and a roster that reads like an AI research all-star team. The San Francisco company is led by Richard Socher, the You.com founder and Imagenet contributor, and counts Peter Norvig, Cresta co-founder Tim Shi, and former Google DeepMind researcher Tim Rocktäschel among its founding group. The stated goal is the long-running holy grail of contemporary AI research: a model that autonomously identifies its own weaknesses and redesigns itself to fix them, with no human in the loop.
Greycroft and GV are among the backers of the $650 million round, putting Recursive Superintelligence in the upper tier of research-first AI startups by capitalization at launch. Socher told TechCrunch the team's technical bet rests on a concept called open-endedness, which Rocktäschel worked on directly while at DeepMind, including on the Genie 3 world model. The pitch is that open-endedness is the missing ingredient between today's auto-research demos and true recursive self-improvement.
Socher draws a sharp line between the two. "Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved," he said. "A lot of people already assume it happens when you just do auto-research. You know, you can take AI and ask it to make some other thing better... But that's not recursive self-improvement. That's just improvement."
Key facts
- 01Recursive Superintelligence launched out of stealth Wednesday with $650 million in funding from backers including Greycroft and GV.
- 02Founders include Richard Socher (You.com), Peter Norvig, Cresta co-founder Tim Shi, and ex-DeepMind researcher Tim Rocktäschel.
- 03Tim Rocktäschel led open-endedness and self-improvement teams at Google DeepMind, working on the Genie 3 world model.
- 04Josh Tobin, an early OpenAI hire who led the Codex and deep research teams, has joined the founding group.
- 05Socher says first products will ship in quarters, not years, despite the company's research-heavy framing.
The mechanism Socher describes is co-evolution between models. He pointed to rainbow teaming, a technique from Rocktäschel's research that pits one model against another across millions of iterations to surface jailbreaks and failure modes from many angles at once. The same loop, applied to research ideation, implementation, and validation, is what Recursive Superintelligence intends to scale. "That's how we developed eyes in our [heads]," Socher said, drawing an analogy to biological adaptation.
“Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved.”— Jaeden Schafer
The founding team has been publishing in the open-endedness space for the last decade, which Socher uses to distinguish Recursive from newer entrants chasing the same problem. Tim Shi previously built Cresta into a unicorn. Josh Tobin was an early OpenAI employee who eventually led the Codex and deep research teams there. Norvig's name carries weight from decades of AI textbook and Google research work.
Socher pushes back on the "neolab" label increasingly applied to research-heavy AI startups that defer product shipping in favor of capability work. "I actually sometimes struggle a little bit with this neolab category. I feel like we're not just a lab," he said. "I want us to be become a really viable company, to really have amazing products that people love to use, that have positive impact on humanity." Asked when the first product ships, he said the team is pulling timelines forward. "But yes, there will be products, and you'll have to wait quarters, not years."
The premise of recursive self-improvement carries an economic implication that Socher acknowledges directly: once the loop closes, compute becomes the binding constraint. The faster the system runs, the faster it gets better, and the human inputs that have historically gated AI progress matter less. "Compute is not to be underestimated," he said. "I think in the future, a really important question will be: how much compute does humanity want to spend to solve which problems?"
That framing aligns Recursive Superintelligence with the same capital-intensive arc driving every frontier lab — but with a sharper edge, because a working recursive system would, in theory, compound returns on each additional GPU. It also raises the same unresolved technical question every lab faces: whether scaling alone produces the kind of self-aware research agent Socher describes, or whether new training methods are required. Socher claims his team has the latter; that claim is now testable in public, on a clock measured in quarters.
Skeptics will note that recursive self-improvement has been a stated goal at OpenAI, DeepMind, and Anthropic for years without a public demonstration, and that open-endedness as a research direction has yet to produce a model that meaningfully outperforms standard post-training on real benchmarks. The $650 million buys Recursive Superintelligence time to make the case, but the bar is the same one every neolab faces: ship something a customer pays for, or ship a capability the big labs haven't.
The launch lands in a market where the gap between research labs and product companies is narrowing fast. Anthropic, OpenAI, and Google are already pushing their own automated-research agendas inside existing model lines, which means Recursive Superintelligence has to deliver on the open-endedness thesis specifically — not just on auto-research broadly — to justify its valuation. If Socher's quarters-not-years product timeline holds, the company will face a tougher test than most neolabs: convincing buyers that a self-improving system produces measurably better outputs than a frontier model fine-tuned by humans. That's a benchmark question, and benchmarks are unforgiving.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
