Richard Socher's Recursive Superintelligence raises $650M for self-improving AI
Richard Socher, the prominent AI researcher best known for founding the search engine You.com and his pioneering work on ImageNet, has emerged from stealth with a new startup called Recursive Superintelligence. The San Francisco-based company has raised $650 million in funding and assembled a team of prominent AI researchers, including Peter Norvig — one of the founding figures of modern AI — and Cresta co-founder Tim Shi.
The startup's ambitious goal is to build a recursively self-improving AI model: a system that can autonomously identify its own weaknesses, design experiments to test potential improvements, and implement those improvements without human intervention. This concept — recursive self-improvement — has been a long-held holy grail of AI research, theoretically enabling an AI system to improve at an accelerating rate.
Socher is careful to distinguish his approach from what he calls the «neolab» trend — the recent wave of AI startups that raise enormous sums for pure research without shipping products. In an interview with TechCrunch, he insisted that Recursive Superintelligence will focus on building practical products alongside its research efforts, though he declined to specify what those products might look like.
The technical approach, while still largely under wraps, reportedly involves a fundamentally different architecture than the transformer-based models that dominate current AI. Rather than simply scaling up existing approaches, Recursive's team is exploring new ways to build models that can modify their own architecture and training processes.
The $650 million raise reflects continued investor appetite for ambitious AI bets, even as questions mount about the ROI of massive AI infrastructure investments. Whether Recursive Superintelligence can deliver on its extraordinarily ambitious vision remains to be seen, but the caliber of its team and the scale of its funding suggest it will be a company to watch closely in the coming years.
Recursive self-improvement is the most ambitious goal in AI — and the most speculative. The $650M raise shows that investors are still willing to fund moonshots, but Socher's insistence on shipping products (unlike typical neolabs) is a smart signal. If they achieve even partial self-improvement, it could reshape how AI models are developed.
What does recursive self-improvement mean for AI safety?
It's both the promise and the danger. A system that improves itself could accelerate scientific discovery enormously. But without proper safeguards, it could also improve in unpredictable directions. Most AI safety researchers consider recursive self-improvement one of the key scenarios that requires careful oversight and control mechanisms.