AI fixed itself overnight. Found the bug, patched it, verified the fix — no humans in the loop. OpenAI's GPT-5.3 Codex is already described as debugging its own training and managing its own deployments, and Anthropic says most of Claude's code is now written by AI itself. And now, a team has come out of stealth trying to close this loop at the company level — with $650M in hand.
What kind of loop is this, exactly?
Recursive Superintelligence was founded in late 2025 and came out of stealth in May 2026 — a San Francisco AI startup. Leading it is Richard Socher — former Chief Scientist at Salesforce, founder of You.com, the researcher who first mainstreamed deep learning for NLP. Stanford PhD, 215,000 citations, #4 in NLP citations worldwide.
But honestly, it's the co-founder lineup that's even more impressive.
- Tim Rocktäschel
UCL professor and former Director of the Open-Endedness group at Google DeepMind. Co-author of the Rainbow Teaming paper, lead on the Genie 3 world model. - Alexey Dosovitskiy
Co-author of the Vision Transformer (ViT) paper — the person who first applied transformers to computer vision. - Jeff Clune
Pioneer in evolutionary algorithms and open-endedness. Involved in Darwin Gödel Machine research (AI autonomously rewriting its own code). - Josh Tobin
Early OpenAI hire, led the Codex and Deep Research teams. - Yuandong Tian
Research Director at Meta FAIR, developer of DarkForest Go and ELF OpenGo.
And Peter Norvig as an advisor — the author of the AI textbook Artificial Intelligence: A Modern Approach and former Director of Google Research (25-year tenure). All of these names are in one company, right?
Their mission is singular: build a loop where AI autonomously handles every stage of AI research — idea generation, implementation, and validation — fully automated. Socher put it this way: "Our core focus is building a truly recursively self-improving superintelligence, meaning that the ideation, implementation, and validation of research ideas is fully automated."
How is this different from current AI research?
Today's AI research is a linear process: human scientists form hypotheses, design experiments, write code, and analyze results. AI assists at each step, but closing the loop still requires humans.
What Recursive Superintelligence is going after is having the AI itself close that loop. GV's (Google Ventures) investment thesis explains it well: "AI is code, and now AI can code. When those two realities connect, the self-improvement loop closes."
| Current AI Research | Recursive Self-Improvement | |
|---|---|---|
| Hypothesis Generation | Human Researcher | AI autonomous generation |
| Experiment Implementation | Human Coding | AI autonomous implementation |
| Results Validation | Human Analysis | AI autonomous validation |
| Improvement Direction | Passed to next researcher | Same AI repeats recursively |
| Parallel Experiments | Limited by team size | "50,000 PhDs" level parallel |
The signs were already there. OpenAI's GPT-5.3 Codex was described as "the first model to debug its own training, manage its own deployment, and analyze its own evaluations," and Anthropic says most of Claude's code is now written by AI. Google DeepMind's AlphaEvolve is being used to evolve solutions in neural network optimization and chip design.
Recursive is the first team to explicitly package this trend as a company-level mission. A 25-person startup is more directly declaring "closing the loop" as its business goal than any of the big labs.
The core tech: What Recursive is actually building
- Open-Endedness
Instead of optimizing for a fixed goal, the system generates new goals on its own. Inspired by biological evolution — just as an ever-changing environment drives adaptation, the AI continuously creates new challenges to push itself. This is what Rocktäschel spent years researching at DeepMind. - Rainbow Teaming
Two AIs co-evolve — one attacking, one defending — engaging simultaneously from multiple angles. Presented at NeurIPS 2024, it achieved over 90% attack success rate across tested models and an average 50% transfer rate against closed models like GPT-4o. Major AI labs are now using this for safety testing. - Darwin Gödel Machine
A system where AI directly rewrites its own code. Published by Sakana AI and UBC (with Jeff Clune), it autonomously improved performance on SWE-bench from 20% to 50%. Recursive is essentially trying to scale this line of research to the company level. - World Models
Rocktäschel's Genie 3 is a general-purpose world model that generates interactive environments at 24fps 720p from text prompts. It's also being used to generate training environments for Waymo's autonomous driving. This technology feeds directly into the AI self-improvement environment. - Goal: 'Level 1 Autonomous Training System'
The first milestone is building an autonomous training system with the capabilities of 50,000 PhDs and doctors in AI science. The roadmap then expands to physics, chemistry, preclinical biology, battery chemistry, and fusion physics.
Why now?
Rocktäschel cites Stanisław Lem's concept of an "information barrier" — the point where knowledge accumulates faster than humans can process it. Recursive's mission is to automate AI research methodology itself to break through that barrier. Jeff Clune put it this way: "We've just turned the corner where recursively self-improving systems will transform science and technology."
