Recursive Intelligence emerges from stealth with $650M and a team of AI luminaries to build recursive self-improving AI

Ricursive Intelligence emerges from stealth with $650M raised at a $4B pre-money valuation, led by GV (Google Ventures) with participation from Nvidia, AMD, and Greycroft. Richard Socher, formerly Chief Scientist at Salesforce and founder of You.com, is leading the company alongside seven co-founders drawn from some of the most cited research groups in AI.

The co-founder list is the most immediately legible part of the pitch. Tim Rocktäschel co-invented Retrieval-Augmented Generation and led Genie 3, DeepMind's world model. Jeff Clune published the Darwin Gödel Machine and HyperAgents papers. Josh Tobin was among OpenAI's earliest hires and led work on ChatGPT agents, Codex, and Deep Research. Alexey Dosovitskiy co-invented the Vision Transformer. Samin Shiung co-invented prompt engineering. Tim Shih built Cresta to unicorn status. Yuan Dong Chen led reinforcement learning at Meta.

“We raised north of $650,000,000 at a $4,000,000,000 pre, led by Google Ventures, with participation from Nvidia, AMD, and many other incredible funds. At Recursive, we're looking to build recursive self-improving superintelligence to automate knowledge discovery. The team has already made some incredible meta inventions that led the AI to make its own inventions.”

The core thesis

Ricursive's stated goal is recursive self-improving superintelligence — AI that can automate the scientific method applied to itself, covering ideation, implementation, and validation of its own research. Socher is explicit that today's coding agents are useful for what he calls "auto research," but fall short of true RSI because they don't control the full stack: model weights, architecture, infrastructure, and the entire training pipeline. Owning all of that is the design requirement.

The near-term commercial roadmap is deliberately vague. Socher says the team originally planned to keep research closely held for at least a year, but early breakthroughs may accelerate productization. No dates are committed.

Fundraise and scale

Socher frames $650M as "step one" at this competitive tier. He's direct that competing at the frontier will eventually require more, but the current raise gives the company meaningful runway unless early progress justifies accelerating spend. He describes a potential new scaling law where more compute yields more AI-generated inventions, not just better benchmark performance.

On open source, Socher says Ricursive will use whatever is available — open and closed — but expects to eventually own and build the full stack itself.

Competitive positioning

The four-part right-to-win Socher describes is focus, team, the RSI architecture itself, and early internal results he characterizes as "meta inventions" — the AI making its own inventions. He argues that adjacent research threads pursued by other labs (continual learning, world models, memory) are all special cases that a successful RSI system would produce as outputs rather than independent workstreams.

Socher explicitly wants Ricursive positioned as a company, not just a lab. The commercial analogy he reaches for is Replit and Lovable — platforms that let others build new product categories on top of AI-generated code — applied at a more foundational level.