Richard Socher has been around the AI block more than once. The former Salesforce chief scientist, ImageNet competition winner, and founder of You.com is back with something that reads less like a startup pitch and more like a science fiction premise — a company called Recursive Superintelligence that just raised a staggering $650 million to build AI that can improve itself, automatically, without any humans in the loop.
The San Francisco-based startup emerged from stealth this week backed by a heavyweight lineup that includes Peter Norvig (widely known for his work on “Artificial Intelligence: A Modern Approach”), Cresta co-founder Tim Shi, and DeepMind open-endedness specialist Tim Rocktäschel. The vision: a recursively self-improving AI that can identify its own weaknesses, propose fixes, implement them, and validate the results — all autonomously.
What Makes Recursive Self-Improvement Different
Socher is quick to draw a distinction between basic AI-driven improvement and what his team is chasing. Asking an existing model to tweak a parameter or optimize a routine isn’t recursive improvement — it’s just automation. True recursive self-improvement means the system handles the full pipeline: ideation, implementation, and validation of research ideas, starting with AI research itself and eventually expanding into any domain.
“The entire process of ideation, implementation, and validation of research ideas would be automatic,” Socher explained in a post-launch interview. “It’s particularly powerful when it’s AI working on itself, developing a new kind of sense of self-awareness of its own shortcomings.”
The technical lever here is what the team calls “open-endedness” — a concept borrowed from evolutionary biology. Rather than optimizing toward a fixed goal, open-ended systems create environments where multiple AI agents can co-evolve, adapt, and counter-adapt over millions of iterations. Rocktäschel previously applied this at DeepMind with Genie 3, an interactive world model, and with “rainbow teaming,” a technique now used across major AI labs where one AI constantly probes another for vulnerabilities.
Not Just Another Neolab
Socher pushes back on the “neolab” label that’s been applied to a wave of research-focused AI startups. He insists Recursive Superintelligence is building toward real products, not just papers. “I want us to become a really viable company, to really have amazing products that people love to use, that have positive impact on humanity,” he said.
The team includes Josh Tobin, one of OpenAI’s earliest hires who later led Codex and deep research teams, alongside Shi, who scaled Cresta to unicorn status. With that pedigree, Socher hinted that product timelines are moving faster than expected — “quarters, not years” — for the first release.
The Compute Question
A recursive self-improvement system introduces a fascinating dynamic: once the AI can improve itself without human input, compute becomes the only meaningful bottleneck. The question shifts from “how do we make this smarter?” to “how much compute should we allocate to which problems?”
“In the future, a really important question will be: how much compute does humanity want to spend to solve which problems?” Socher noted. “Here’s this cancer and here’s that virus — which one do you want to solve first? It becomes a matter of resource allocation.”
That framing moves the conversation from technical capability to something far more strategic. For startup founders, it raises a pressing question: if the race for AI intelligence increasingly becomes a race for compute, where does that leave companies that don’t have $650 million rounds or access to data center clusters?
The Takeaway for Founders
Recursive self-improvement isn’t just a technical milestone — it’s a potential inflection point for how we think about building companies. If AI can eventually improve itself at machine speed, the competitive advantages that matter today (team quality, data moats, proprietary algorithms) could erode far faster than most business models account for.
The winning strategy might not be building the smartest AI. It might be figuring out which problems are worth solving once intelligence becomes a commodity. Socher’s team is betting that superintelligence arrives as an allocation problem, not a discovery problem. For founders watching from the sidelines, now is the time to start asking the same question: which cancer, which virus, which customer problem would you prioritize if you had infinite intelligence but finite resources?
The answer to that question might matter more than the technology itself.
This article was based on reporting from TechCrunch. Original reporting by Devin Coldewey.
