Richard Socher Launches Recursive Superintelligence With $650 Million To Build Self-Improving AI Models

Richard Socher, a former founder of the chatbot startup You.com and a contributor to the ImageNet project, launched Recursive Superintelligence out of stealth on Wednesday with $650 million in funding to pursue recursively self-improving AI models.

“Half a year ago, if you had clicked into the comment section of the official announcement tweet of Tian Yuandong, a former Meta AI scientist, you would have thought you had accidentally stumbled into the scene of a high - end "Boss Zhipin" recruitment event”

36Kr

The San Francisco-based startup says its goal is to build a system that can autonomously identify its own weaknesses, design fixes, and implement them without human intervention, framing it as “recursive self-improvement” rather than “just improvement.”

36Kr

In a Zoom interview after the launch, Socher said, “Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved,” and he added, “A lot of people already assume it happens when you just do auto-research.”

TechCrunch described Recursive Superintelligence as a San Francisco-based startup that came out of stealth on Wednesday with $650 million in funding, and it said Socher is joined by researchers including Peter Norvig and Cresta co-founder Tim Shi.

Socher tied Recursive Superintelligence’s approach to “open-endedness,” describing biological evolution as a model where animals adapt and others counter-adapt over time.

In the same interview, he said, “That’s how we developed eyes in our [heads],” and he pointed to Tim Rocktäschel’s experience leading open-endedness and self-improvement teams at Google DeepMind.

Bitcoin World

TechCrunch also reported that Rocktäschel worked on the world model Genie 3 as an example of open-endedness, and it described how “rainbow teaming” works by letting two AI systems co-evolve through iterative competition.

The CNBC report on Nvidia’s partnership with Ineffable Intelligence similarly framed the broader push toward systems that “learn by trial and error,” contrasting it with models trained on human data, while noting that Nvidia CEO Jensen Huang called the next frontier “superlearners.”

Multiple outlets tied Recursive Superintelligence’s launch to the same $650 million round and a $4.65 billion valuation, with SiliconANGLE saying Alphabet Inc.’s GV and Greycroft led the round and that Nvidia and Advanced Micro Devices’ venture arm also participated.

“Nvidia has announced a partnership with an AI startup, just months after it was founded by a former top scientist at Google DeepMind”

CNBC

SiliconANGLE reported that Recursive says the investment values it at $4.65 billion and that the company was founded earlier this year by former Salesforce Chief Scientist Richard Socher, while also stating the team is working to build an AI model that can improve its own code base.

In the Tech Funding News account, Recursive is described as operating from offices in San Francisco and London and expanding to more than 25 researchers and engineers, with plans to use the funding to secure large-scale compute infrastructure and run its first “Level 1” autonomous training system.

Tech Funding News also said a public launch is targeted for mid-2026 and that ambitions extend beyond AI research into broader scientific discovery over time, while SiliconANGLE added that Recursive did not disclose what machine learning methods will power its self-improving AI.