Recursive Superintelligence Emerges From Stealth With $650 Million to Build Self-Improving AI

A four-month-old startup with a provocative name and an even more provocative mission has emerged from stealth with $650 million in funding. Recursive Superintelligence, a London-based AI research lab, is building systems designed to improve themselves autonomously -- and some of the most respected names in AI research are backing the effort.

The round, which values the company at $4.65 billion, was led by GV and Greycroft with additional backing from AMD Ventures and NVIDIA. The funding will primarily go toward securing large-scale compute infrastructure and running what the company calls its first "Level 1" autonomous training system.

"We are building software that continuously generates and refines new capabilities in an open-ended cycle," said co-founder Richard Socher, the former chief scientist at Salesforce who previously created the widely used GloVe word embedding system. "The goal is AI that does not just learn tasks but independently discovers better ways to learn."

A Superstar Founding Team

The founding roster reads like a who's who of AI research. Alongside Socher, the team includes Tim Rocktaschel, formerly of DeepMind, where he led research on open-ended learning; Jeff Clune, a pioneer in AI-generating algorithms who previously held positions at OpenAI and Uber AI; Josh Tobin, co-founder of Gantry and a former OpenAI researcher; and Tim Shi, who led foundation model training at Meta AI.

The company currently employs more than 25 researchers and engineers across offices in San Francisco and London, with plans to scale significantly using the new capital.

"The next major leap in AI will not come from scaling existing architectures," Clune said in an interview with TechCrunch. "It will come from systems that can redesign their own learning processes. That is what we are building."

The Technical Vision

Recursive Superintelligence's approach differs fundamentally from the scaling paradigm that has dominated AI development over the past several years. Rather than simply training larger models on more data, the company aims to create systems that autonomously generate new training approaches, architectures, and evaluation methods.

The "Level 1" system the company plans to deploy represents the first tier of autonomy in this framework -- a system capable of proposing and testing modifications to its own training pipeline without human intervention. Higher levels would involve increasingly autonomous research capabilities, though the company has been deliberately vague about specific timelines.

The approach draws on decades of research into meta-learning, neural architecture search, and evolutionary computation, areas where several of the founders have published seminal work. What is new is the scale of compute and the ambition to run these processes continuously rather than as one-off experiments.

Controversy and Caution

The company's name alone has sparked debate within the AI safety community. Critics argue that explicitly pursuing recursive self-improvement -- the theoretical mechanism behind many AI risk scenarios -- is irresponsible, regardless of the technical safeguards in place.

Supporters counter that understanding and controlling self-improving systems requires building them, and that having safety-conscious researchers at the helm is preferable to the alternative.

The company has committed to publishing safety evaluations of its Level 1 system and has engaged with external auditors, though specific details of its safety framework have not been disclosed.

What to Watch

A public launch is targeted for mid-2026, with ambitions extending beyond AI research into broader scientific discovery. Whether Recursive Superintelligence can deliver on its audacious technical vision -- and do so safely -- will be one of the defining questions in AI research this year. The company's next milestone will be demonstrating that its Level 1 system can produce measurable improvements to its own training efficiency.