Socher's Recursive Raises $4.65B to Automate AI Research
Natural language processing pioneer Richard Socher has launched Recursive, a startup valued at $5 billion that aims to build self-improving AI systems capable of automating scientific discovery.
Artificial intelligence pioneer Richard Socher has spun out a new startup called Recursive, securing a massive $4.65 billion seed round that values the company at $5 billion. Moving away from his previous frontier-model work at You.com, Socher is focusing the new venture on recursive self-improvement. The startup aims to build what Socher calls a Eureka Machine, a superintelligence designed to automate the process of scientific invention and accelerate AI research. Socher's vision is detailed in his upcoming book, The Eureka Machine, scheduled for release this September.
Recursive has already demonstrated early capabilities in automating complex engineering tasks. In less than two days, the startup's AI research system outperformed both human engineers and existing agents on optimization benchmarks. Additionally, the system successfully discovered performance improvements for Nvidia GPU kernels without requiring assistance from a dedicated team of CUDA experts. Socher suggests that these automated workflows could eventually compress AI research timelines from years of human labor down to mere weeks.
The startup's broader roadmap includes applying self-improving agents to scientific domains like biology, chemistry, and energy. Recursive is leveraging early results from projects like NanoChat, NanoGPT, and the AI Economist, which simulates entire economies to test policy. For practitioners, this shift toward automated research systems could redefine how hardware and software are optimized, shifting the developer's role from manual coding to reward engineering and sandbox design.
Socher, who also invests via AIX Ventures in startups like WhisperFlow, remains highly critical of current safety paradigms like Anthropic's constitutional AI, calling such frameworks "mostly marketing" because they fail to prevent vulnerabilities. Instead, Recursive is exploring evolutionary approaches and open-ended simulations to build resilient, aligned models. Socher argues that open-source AI is vital for maintaining geopolitical soft power and preventing a monoculture of alignment.
This is our own summary of reporting by Latent Space
