The archive · AI & Models · Strategic decision · 2025–2026
Recursive Superintelligence's self-improving AI bet: $650M at $4.65B before shipping
Ex-Meta, OpenAI and DeepMind researchers bet AI that rewrites its own code is the fastest safe path to superintelligence; GV and Greycroft led.
Recursive Superintelligence
What the business is
Recursive Superintelligence is an AI research lab building systems that autonomously experiment, identify their own limitations, write their own benchmarks, and rewrite their own code to become more capable, starting with the science of AI itself.
Starting capital：$650M early round at a $4.65B valuation, co-led by GV and Greycroft with participation from Nvidia and AMD (announced May 13, 2026).
How it started
In late 2025, as Meta cut around 600 AI roles on Oct 22, former Meta FAIR research director Yuandong Tian joined ex-Salesforce chief scientist Richard Socher and researchers with roots at OpenAI, Google DeepMind and Meta — including Tim Rocktäschel, Alexey Dosovitskiy, Jeff Clune, Tim Shi and Caiming Xiong — to found Recursive Superintelligence with offices in San Francisco and London and a team under 30.
What happened
After the FT reported a $500M round at a $4B valuation in April 2026, Recursive emerged from stealth on May 13, 2026 with $650M at a $4.65B valuation, co-led by GV and Greycroft with participation from Nvidia and AMD. The NYT reported that notable researchers, including Peter Norvig, had joined the effort.
No ending yet — it is still running.
Background
Recursive Superintelligence is a London-and-San Francisco AI lab founded in late 2025 by seven researchers from the top of the field: CEO Richard Socher, ex-chief scientist at Salesforce; Yuandong Tian, the former Meta FAIR research director who was cut in Meta's October 22, 2025 AI layoffs; Tim Rocktäschel, a UCL professor and ex-Google DeepMind scientist; Alexey Dosovitskiy, co-author of the Vision Transformer; and Jeff Clune, Tim Shi and Caiming Xiong.
Its thesis is the one GV summarized as 'AI is code, and AI can now code': instead of relying on human engineers to hand-design optimizations, Recursive builds systems that conduct their own experiments, learn to identify their own limitations, write their own benchmarks, and rewrite their own codebase — an open-ended loop the company argues is the fastest safe path to superintelligence.
The first milestone is training a system with the capability of '50,000 PhDs,' focused on the science of AI itself; once that engine runs, the company plans to point the same 'Eureka machine' at drug discovery, new battery chemistries and fusion physics.
After an April 2026 FT report of a $500M round at $4B, Recursive emerged from stealth on May 13, 2026 with $650M at a $4.65B valuation, co-led by GV and Greycroft with participation from Nvidia and AMD; the NYT reported Peter Norvig among the notable researchers joining the effort.
What has to be true
- The bet is explicit and falsifiable: recursive self-improvement, not scaling or alignment-by-debate, is the fastest safe route to superintelligence.
- The valuation is a pure bet on people: $4.65B for a startup with no product, no revenue and fewer than 30 staff.
- The founding team is unusually deep and verifiable — Meta FAIR, DeepMind, OpenAI, Salesforce and ViT co-authors among them.
- It is a distinct company and event from every library entry, including Safe Superintelligence and Reflection AI, which bet on different routes.
- The stated target — a system with the capability of 50,000 PhDs — makes the company's scope and ambition concrete.
What can be applied
When the product is the lab itself, capital prices the founders' research pedigree: $650M at $4.65B went to a pre-product startup whose first goal is a system with the capability of 50,000 PhDs.
Aftermath
As of September 2, 2026, Recursive operates from San Francisco and London with a team under 30 and has not shipped a product. It describes its first milestone as training a system with the capability of 50,000 PhDs, focused first on the science of AI itself before being pointed at drug discovery, battery chemistry and fusion physics. The company hired through 2026 and sits alongside Safe Superintelligence and Reflection AI as one of the highest-valued labs racing on recursive self-improvement.
Sources
- Recursive Superintelligence emerges from stealth with $650M raise
- Recursive Superintelligence: Why Self-Improving AI is the Next Frontier
- 华人科学家田渊栋创业公司获6.5亿美元融资，此前任Meta FAIR团队研究总监
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card