Richard Socher, a veteran of artificial‑intelligence research best known for founding the early chatbot company You.com and for his contributions to the ImageNet project, has announced a new venture focused exclusively on recursive self‑improvement. The San Francisco‑based startup Recursive Superintelligence emerged from stealth on Wednesday with a $650 million funding round.
The company’s leadership includes a roster of established AI researchers, among them Peter Norvig, former Director of Research at Google, and Tim Shi, co‑founder of Cresta. Their goal is to build an AI system that can automatically identify its own shortcomings, redesign its architecture, and validate the changes without human input—a capability long described as the “holy grail” of AI development.
In a Zoom interview after the launch, Socher explained the team’s distinctive technical direction. Unlike approaches that merely automate particular research tasks, Recursive aims to create a fully open‑ended loop in which ideation, implementation and evaluation are all generated by the AI itself. “Our focus is on true recursive self‑improvement at scale,” he said. “The system should be able to propose new research ideas, develop them, test them, and then use the results to generate the next generation of ideas.”
The concept of “open‑endedness” has a precise meaning for the team. Co‑founder Tim Rocktäschel, who previously led DeepMind’s open‑endedness and self‑improvement groups, cites the world‑model project Genie 3 as an exemplar. Genie can be given any concept, environment or agent and will construct an interactive simulation without pre‑programmed constraints. The Recursive team intends to extend that flexibility, allowing the AI to explore a virtually unlimited space of hypotheses, including those that pertain to its own architecture.
One practical mechanism the founders highlighted is a variant of “rainbow teaming.” In cybersecurity, red‑team exercises probe a system for vulnerabilities by attempting to make a language model generate disallowed content, such as instructions for building a bomb. Recursive proposes pairing two AIs: one tasked with exposing the other’s unsafe outputs, the other tasked with defending against them. By iterating this adversarial co‑evolution millions of times, the system can discover and patch failure modes far more thoroughly than human testers. The method, which originated in Rocktäschel’s research, is now being adopted by many leading labs.
When asked whether the project marks a departure from the work of larger organisations, Socher noted that while he could not speak for others, Recursive’s commitment to open‑ended self‑improvement sets it apart. The team’s background includes building commercial products—Shi grew Cresta into a unicorn, and Josh Tobin, an early OpenAI employee, led the Codex effort—giving the founders experience in turning research breakthroughs into viable services. Socher added that the “neolab” label feels limiting; he envisions a company that both advances fundamental AI capabilities and delivers products that users can adopt within a few quarters rather than years.
The interview touched on the broader implication that, once truly recursive AI is achieved, compute power will become the primary bottleneck. Faster hardware would accelerate the self‑improvement cycle, shifting the competitive landscape from algorithmic ingenuity to the scale of processing resources. Socher warned that society will soon face decisions about how much compute to allocate to solving specific problems, such as disease eradication versus climate mitigation.
Recursive Superintelligence’s launch signals a new front in AI research, where the ambition is not merely to build smarter models but to create systems that can redesign themselves autonomously. With substantial capital and a team versed in both academic breakthroughs and product development, the venture is positioned to test whether open‑ended, self‑improving AI can move from theory to practice. The next months will reveal whether the company can translate its recursive approach into concrete applications and how its progress will influence the broader AI ecosystem.
