Recursive Superintelligence Emerges from Stealth with $650M to Build Self-Improving AI
Richard Socher's new startup wants to create AI that redesigns itself autonomously — no humans required.
Last verified:
Recursive Superintelligence, a San Francisco-based AI startup, exited stealth on May 14, 2026, announcing $650 million in funding and a high-profile founding team. The company’s central ambition is technically ambitious even by today’s standards: building AI that can locate its own weaknesses, devise fixes, and rewrite itself — all without human involvement at any stage of the loop.
The Team Behind the $650M Bet
According to TechCrunch AI, the startup was co-founded by Richard Socher — known for establishing the early chatbot startup You.com — alongside AI researcher Peter Norvig, Cresta co-founder Tim Shi, and Tim Rocktäschel, who previously led open-endedness and self-improvement research efforts. The roster represents a convergence of industry veterans and specialists in machine learning architecture.
What “Recursive Self-Improvement” Actually Means
TechCrunch AI reports that Socher draws a sharp distinction between genuine recursive self-improvement and the simpler, more common practice of using one AI to improve another. In the latter case, a human-defined system is still doing the directing. Recursive Superintelligence wants the entire research cycle — concept generation, experimental execution, and outcome verification — to run autonomously, with the AI targeting itself as the primary subject of improvement.
The technical vehicle for this is “open-endedness,” a framework co-founder Rocktäschel worked on at prior research roles, including work on the world-building model Genie 3. The concept draws on continuous novelty generation rather than optimizing toward any fixed objective.
Why This Matters
Self-improving AI has been a theoretical goal since the earliest days of the field, but it has remained out of reach because each step of the improvement cycle — spotting the problem, designing the fix, confirming it worked — has required human judgment. If Recursive Superintelligence can close that loop reliably, it would represent a meaningful architectural shift rather than incremental capability gains. Researchers evaluating long-horizon autonomy and teams tracking AGI-adjacent milestones should watch how the startup defines and measures verifiable self-improvement, since those benchmarks will determine whether this $650M thesis holds up under scrutiny.
Frequently Asked Questions
What is Recursive Superintelligence trying to build?
The startup aims to create AI that can autonomously generate, test, and apply improvements to itself, covering the full research cycle from idea generation through validation — without human intervention.
Who founded Recursive Superintelligence?
Richard Socher, previously the founder of chatbot startup You.com, co-founded the company alongside Peter Norvig, Cresta co-founder Tim Shi, and Tim Rocktäschel, among other AI researchers.
What is 'open-endedness' in this context?
Open-endedness refers to a technical approach where AI systems can keep generating novel, useful outputs indefinitely — drawing inspiration from how biological evolution produces continuous adaptation without a fixed endpoint.