Recursive Superintelligence raises $650M at $4.65B valuation to build self-improving AI

A new artificial intelligence startup has emerged from stealth mode with an ambitious and potentially transformative vision: building AI systems that can improve themselves without human intervention. Recursive Superintelligence announced this week that it has raised $650 million in funding at a $4.65 billion valuation, positioning itself at the forefront of what could represent a fundamental shift in how artificial intelligence develops.

The company, founded by prominent researchers including former Salesforce Chief Scientist Richard Socher and former Google DeepMind scientist Tim Rocktäschel, brings together talent from some of the world's leading AI labs including OpenAI, Meta, and Uber. The funding round was led by Google Ventures and investment firm Greycroft, with additional backing from chip manufacturers NVIDIA and AMD Ventures. According to reports, the round was heavily oversubscribed, reflecting strong investor confidence in the company's approach.

Recursive's core mission centers on developing what it calls recursive self-improving superintelligence — AI systems capable of conducting their own experiments to optimize and improve themselves through automated scientific discovery. Unlike current artificial neural networks that rely on human-designed algorithms and manual improvements, Recursive aims to create AI that can autonomously discover ways to enhance its own code, architecture, and training methods. The company believes this approach mirrors how human intelligence evolved through open-ended evolutionary and cultural processes that continuously built upon previous discoveries. Rather than simply scaling up model size, Recursive's thesis suggests the next breakthrough in AI will come from automating the research process itself.

The company plans to begin by focusing on AI systems that improve AI itself before eventually expanding these capabilities into broader scientific disciplines. Recursive's first priority is building an AI model capable of improving its own code base, with the system developing experiment ideas, testing them, and validating results in what the company describes as an open-ended process of automated scientific discovery. The startup plans to run what it calls its first "Level 1" autonomous training system using the newly raised capital to secure large-scale computing infrastructure. A public launch is targeted for mid-2026, though the company currently has no publicly available products.

Safety considerations remain central to Recursive's approach. The company has emphasized that it will develop guardrails to prevent its AI models from producing risky output, stating that maximizing benefits for humanity while reducing associated risks will be core priorities as the technology develops. The startup currently operates from offices in San Francisco and London with a team of more than 25 researchers and engineers.

The timing of Recursive's emergence coincides with ongoing discussions about whether artificial intelligence is approaching a critical inflection point. Some observers have drawn parallels to what economist Robert Solow identified in 1987 as the Solow Paradox — the observation that despite computers becoming ubiquitous, productivity gains remained mysteriously absent from economic statistics for years. The analogy suggests that AI's transformative impact may have been largely invisible in traditional metrics, but could be poised for a dramatic visibility shift. Whether Recursive's approach to self-improving AI represents that tipping point remains to be seen, but the company's substantial funding and experienced leadership team suggest investors believe the potential is real.