Richard Socher just came out of stealth with a new AI startup and a striking pitch: build an AI that rewrites itself. According to TechCrunch AI, his San Francisco-based venture, Recursive Superintelligence, launched Wednesday with $650 million in funding and a team that reads like a who’s-who of the field. Peter Norvig, Cresta co-founder Tim Shi, former OpenAI Codex and deep research lead Josh Tobin, and DeepMind’s Tim Rocktäschel are all on board.
The goal is the one researchers have been chasing for years: recursive self-improvement, or RSI. That means an AI system that can spot its own weaknesses, redesign itself, and validate the fix without a human in the loop. Socher told TechCrunch AI that most of what people call recursion today is really just “improvement”: using AI to make something else better. He’s drawing a sharper line. Recursive’s bet is on automating the full research cycle: ideation, implementation, and validation.
The technical wedge: open-endedness
What sets Recursive apart, in Socher’s framing, is a focus on “open-endedness.” Rocktäschel led open-endedness and self-improvement work at Google DeepMind, including on the Genie 3 world model. The analogy Socher uses is biological evolution: animals adapt, predators counter-adapt, and the process keeps producing novelty for billions of years.
He pointed to “rainbow teaming” as a working example. It’s like red teaming, but two AIs co-evolve against each other. One tries to make the other produce harmful outputs. The defender hardens. They iterate millions of times, surfacing many attack angles, not just one. Rocktäschel’s idea is now used across major labs, according to TechCrunch AI.
Why this matters
RSI is the unspoken endgame behind most frontier AI research. If a system can genuinely improve itself, the bottleneck stops being human researchers and becomes raw compute. Socher acknowledged this directly. The faster you run the loop, the faster the system gets smarter, and the more the race collapses into a question of how much processing power you’re willing to spend.
That’s a meaningful shift from the current paradigm, where progress depends on talent density, dataset quality, and human-driven research bets. If Recursive’s thesis works, the moat moves from research org to data center.
The neolab debate
Recursive joins a wave of research-first AI startups that observers have started calling “neolabs.” Socher pushed back on the label. He told TechCrunch AI he wants Recursive to ship real products people use, not just publish papers. First product timelines, he said, are measured in quarters, not years, and the team may even pull those forward.
That’s a notable position. Most neolab peers have leaned into the “pure research” framing. Socher’s signaling something closer to a hybrid: deep research with a commercial spine. His track record at You.com and Salesforce’s Einstein gives that pitch some weight.
What to watch
A few things will tell us whether Recursive is the real deal or another well-funded science project:
- First product shipping window. Socher’s “quarters not years” claim is testable. Watch for an initial release inside 2026.
- Open-endedness benchmarks. Whether the team publishes measurable progress on self-improvement loops, not just rainbow teaming variants.
- Compute footprint. If RSI is compute-bound, Recursive’s infrastructure deals will signal how serious the scaling ambition is.
- Talent gravity. $650 million buys runway, but the real test is whether more senior researchers keep joining over the next six months.
The bigger question Socher raised at the end of his interview is one the whole industry will face: how much compute does humanity want to spend solving its hardest problems? If self-improving systems work, that decision becomes the defining policy question of the decade.
Full interview at the original TechCrunch AI piece.