Earn up to 50% APR
Boost earnings with NFTs
Play, HODL & earn more
A $650 Million AI Startup Just Came Out of Stealth With a Goal Nobody Has Achieved Yet
-
Recursive Superintelligence, a San Francisco-based AI startup founded by Richard Socher alongside Peter Norvig, Cresta co-founder Tim Shi, and Google DeepMind open-endedness researcher Tim Rocktäschel, has emerged from stealth with $650 million in funding and a research goal that the major labs have circled but not reached: truly recursive self-improvement. The distinction Socher draws between recursive self-improvement and what most AI systems already do is important. Asking an AI to improve another system — or even to improve itself through automated research — is just improvement. Recursive self-improvement means the entire process of ideating, implementing, and validating research ideas is automatic, with the AI developing what Socher describes as a genuine awareness of its own shortcomings and redesigning itself to address them without human involvement. "Our main focus is to build truly recursive, self-improving superintelligence at scale," Socher told TechCrunch, noting that the team has been researching this specific problem space for a decade and has a track record of shipping real products alongside academic contributions.The technical approach centers on open-endedness — a concept that Rocktäschel developed extensively at DeepMind, most visibly through the world model Genie 3. In biological terms, open-endedness is the property that allows evolution to keep producing interesting adaptations indefinitely without a fixed endpoint, as organisms and environments co-evolve against each other.
Applied to AI, it means building systems that can keep generating novel improvements without converging on a local optimum. Rocktäschel's rainbow teaming work — now used across major AI labs — illustrates the principle in a safety context: two AIs co-evolve against each other, one trying to elicit harmful outputs and the other hardening against them, producing far more comprehensive safety coverage than human red teamers could generate alone. Socher said the team has made enough progress that product timelines may be pulled forward from initial estimates, with the first products arriving in quarters rather than years.
