Back to Blog
May 14, 2026 • Russell Brandom

What Happens When AI Starts Building Itself?

What Happens When AI Starts Building Itself? What Happens When AI Starts Building Itself? What Happens When AI Starts Building Itself?

What Happens When AI Starts Building Itself?

Overview

Richard Socher, founder of You.com and key figure in early AI research (ImageNet), has launched Recursive Superintelligence with $650 million in funding. The San Francisco-based startup aims to create a recursively self-improving AI model—one that can autonomously identify its own weaknesses and redesign itself without human intervention.

Key Personnel

  • Richard Socher (Founder, formerly You.com and ImageNet)
  • Peter Norvig (Co-founder)
  • Tim Shi (Co-founder, Cresta)
  • Tim Rocktäschel (Led open-endedness & self-improvement teams at Google DeepMind)
  • Josh Tobin (Early OpenAI, led Codex and deep research teams)

Core Technical Approach: Open-Endedness

What Makes This Different?

  • Not just "improvement": Most AI labs use models to improve other systems, but that's not recursive self-improvement
  • True recursion: Automating the entire process of ideation, implementation, and validation of research ideas
  • Self-awareness of shortcomings: The AI develops awareness of its own weaknesses and fixes them

Technical Concepts Explained

Open-Endedness:

  • Inspired by biological evolution (billions of years of adaptation and counter-adaptation)
  • Example: Google DeepMind's Genie 3 world model—can create any concept, world, or agent interactively

Rainbow Teaming:

  • Evolution of red teaming for LLM safety
  • Two AIs co-evolve: one attacks, the other defends
  • They iterate millions of times across multiple attack angles
  • Result: safer, more robust AI systems
  • Now used across major labs

Key Insights from Socher

On Recursive Self-Improvement (RSI)

"Our main focus is to build truly recursive, self-improving superintelligence at scale, which means that the entire process of ideation, implementation, and validation of research ideas would be automatic."

On Product vs. Research

  • Not just a "neolab": Socher resists the research-only label
  • Products coming soon: Timeline measured in quarters, not years
  • Team has track record of shipping real products (Tim Shi built Cresta into a unicorn)

On Compute and the Future

"In the future, a really important question will be: how much compute does humanity want to spend to solve which problems? Here's this cancer and here's that virus—which one do you want to solve first?"

  • Once RSI is achieved, compute becomes the critical resource
  • The race becomes: how much processing power to allocate to which problems
  • Resource allocation becomes one of humanity's biggest questions

Key Takeaways

  • Recursive self-improvement remains an "elusive goal" across AI labs
  • Recursive's unique approach centers on open-endedness inspired by biological evolution
  • The team has deep research credentials (DeepMind, OpenAI) plus product shipping experience
  • Products expected in quarters, not years—despite ambitious research goals
  • Intelligence has bounds, but they're "astronomical"—we're very far from limits
  • Once achieved, RSI transforms the AI race into a compute allocation problem

What This Means

Recursive Superintelligence represents a new approach to the holy grail of AI: systems that can improve themselves indefinitely. Unlike pure research labs, they're committed to shipping products while pursuing this ambitious technical vision. The open-endedness framework—allowing AI systems to co-evolve and counter-adapt like biological organisms—offers a novel path toward recursive self-improvement that differs from current scaling-focused approaches.