Recursive Superintelligence, founded by You.com founder Richard Socher and joined by Peter Norvig and Cresta co-founder Tim Shi, emerged from stealth with $650 million in funding. The startup aims to create a recursively self-improving AI model that can autonomously identify its weaknesses and redesign itself to fix them — without human involvement.
What Is Recursive Superintelligence?
The San Francisco-based startup is working to build the first truly recursive self-improving AI system. Unlike simple AI improvement loops where an AI helps optimize a specific narrow task, recursive self-improvement means the entire process of ideation, implementation, and validation of research ideas would be automatic. The system would start with AI research and eventually extend to any kind of research.
Who Is Behind the Startup?
The founding team reads like an AI hall of fame:
- Richard Socher — former CEO of You.com, known for pioneering work on ImageNet
- Peter Norvig — former Director of Research at Google, co-author of “Artificial Intelligence: A Modern Approach”
- Tim Shi — co-founder of Cresta
- Tim Rocktäschel — former lead of open-endedness and self-improvement teams at Google DeepMind
How Does “Open-Endedness” Drive Self-Improvement?
The approach is inspired by biological evolution. Just as animals adapt to their environments and counter-adapt to each other, the system would use “rainbow teaming” — where two AIs co-evolve, one attacking and the other defending — to continuously improve. This open-ended process can theoretically continue indefinitely, with intelligence increasing without a fixed endpoint.
Key Takeaways
- $650 million in funding for recursive self-improving AI
- Founded by Richard Socher, Peter Norvig, Tim Shi, and Tim Rocktäschel
- Goal: AI that autonomously identifies and fixes its own weaknesses
- Uses open-endedness and rainbow teaming approaches
- Aims to automate AI research entirely
- Part of a wave of new AI labs attracting billions in investment
Frequently Asked Questions
How is this different from what existing AI labs are doing? Current AI labs use human researchers to identify model weaknesses and design improvements. Recursive Superintelligence aims to automate that entire cycle.
Is recursive self-improvement safe? The company acknowledges safety concerns and incorporates red-teaming techniques (including rainbow teaming) as part of the improvement process.
When will the first recursive improvements happen? The company just emerged from stealth and has not announced a timeline for specific capabilities.