Recursive Superintelligence emerges with $650 million funding

On Wednesday, San Francisco-based startup Recursive Superintelligence emerged from stealth with $650 million in funding. The company is working to create a recursively self-improving AI model that can autonomously identify weaknesses and redesign itself. The founding team features prominent AI researchers, including founder Richard Socher, who previously founded the startup You.com, and co-founders Peter Norvig, Tim Shi, and Tim Rocktäschel.

The startup’s team includes veterans from Google DeepMind and OpenAI, such as Rocktäschel, who previously led open-endedness and self-improvement teams at Google DeepMind, and team member Josh Tobin, who previously worked at OpenAI. Tim Shi also brings entrepreneurial experience, having co-founded the startup Cresta. Rather than operating strictly as a “neolab”—an informal term for a new generation of AI startups that prioritize research over building products—Recursive Superintelligence is focusing on “open-endedness.” This technical approach to AI evolution, where systems adapt to environments and counter-adapt, is intended to achieve recursive self-improvement. According to Socher, “Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved.” The ultimate goal is to automate the entire research lifecycle, including ideation, implementation, and validation.

To ensure safety during this automated process, the company utilizes “rainbow teaming,” a method of co-evolving two AIs to attack and inoculate each other for safety. This moves beyond standard “red teaming”—the process of testing AI for vulnerabilities or harmful outputs, such as trying to get a model to “tell it how to build a bomb.” Instead of relying solely on human testers, a second AI is tasked with finding vulnerabilities in the first. The two systems co-evolve back and forth over millions of iterations, allowing the primary model to become safer against various angles of attack.

While the company is research-focused, it plans to ship products in timelines that may span quarters, not years. Over the long term, the startup operates with the view that computing power will become the primary resource for scaling intelligence. As Socher noted, “Compute is not to be underestimated.” In this framework, the speed of AI self-improvement will increasingly depend on the allocation of processing power to solve specific problems.

Why it matters

The emergence of Recursive Superintelligence signals a shift toward research-heavy AI startups aiming to automate the research process itself, moving beyond standard LLM development. By focusing on recursive self-improvement, the company represents an ambitious attempt to bypass human-driven model training in favor of autonomous, machine-led optimization.