Recursive Superintelligence Secures $650M to Build Self-Improving AI

Richard Socher, a prominent figure in AI known for founding You.com and his work on ImageNet, has unveiled his latest venture: Recursive Superintelligence. The San Francisco-based startup emerged from stealth mode on Wednesday, announcing a staggering $650 million in funding. Socher is joined by a formidable team of AI researchers, including Peter Norvig and Cresta co-founder Tim Shi, all dedicated to achieving what many consider the holy grail of contemporary AI research: recursively self-improving AI. The core mission of Recursive Superintelligence is to develop an AI model that can autonomously identify its own weaknesses and redesign itself to rectify them, entirely without human intervention. Socher emphasizes that their approach goes beyond mere "auto-research" or simple improvement. Instead, they are focused on building truly recursive, self-improving superintelligence at scale, where the entire process of ideation, implementation, and validation of research ideas – initially for AI and eventually for broader scientific and even physical domains – becomes fully automatic. This unique capability is particularly potent when AI works on itself, fostering a new kind of self-awareness regarding its own shortcomings. Central to their distinctive methodology is the concept of "open-endedness." This technical term, championed by co-founder Tim Rocktäschel (who previously led open-endedness teams at Google DeepMind and worked on the world model Genie 3), refers to a continuous evolutionary process. Much like biological evolution where organisms adapt and counter-adapt over billions of years, leading to novel developments, open-endedness allows AI to continuously evolve and explore new possibilities. Socher illustrates this with "rainbow teaming," a concept where two AIs co-evolve: one acts as an attacker, constantly finding new vulnerabilities in the other, which then adapts to become safer, iterating millions of times to achieve robust security. When questioned about the completion point for such an endeavor, Socher acknowledges that some aspects will never truly be "done," as intelligence can always be enhanced. He notes that while there are theoretical bounds to intelligence, they are astronomically distant. Furthermore, Socher challenges the informal "neolab" designation, asserting that Recursive Superintelligence aims to be a viable company with amazing products that positively impact humanity, not just a research lab. The team's strong track record, including Tim Shi building Cresta into a unicorn and Josh Tobin's foundational work at OpenAI, underscores their capacity to deliver. Despite the profound research focus, Recursive Superintelligence is not solely a long-term academic pursuit. Socher revealed that products are on the horizon, with timelines measured in quarters, not years. This rapid development underscores their ambition to translate cutting-edge research into tangible applications. Looking further ahead, a significant implication of achieving recursive self-improvement is that compute power will become the singular most important resource. Once such a system is in place, the speed of its improvement will directly correlate with the processing power thrown at it, setting the stage for an unprecedented race in computational capabilities.