Recursive Superintelligence Launches with $650 Million
Richard Socher, a key player in AI known for starting the chatbot company You.com and his earlier contributions to ImageNet, has launched a new venture. Recursive Superintelligence, based in San Francisco, revealed itself on Wednesday after operating in stealth mode. The startup secured $650 million in funding to pursue its goals.
Socher teams up with notable AI experts like Peter Norvig and Tim Shi, who co-founded Cresta. Their mission centers on developing an AI system that improves itself recursively. This means the model would spot its own flaws and rebuild itself to address them, all on its own. Researchers have chased this concept for years.
Team Background and Expertise
The group includes researchers with strong track records. Tim Rocktäschel, a co-founder, headed teams on open-endedness and self-improvement at Google DeepMind. He contributed to the world model Genie 3, which generates interactive worlds, agents, or concepts from any input. Josh Tobin joined early at OpenAI, later leading Codex and deep research efforts. Tim Shi grew Cresta to unicorn status.
Peter Norvig brings decades of AI leadership, including his role as director of research at Google. Socher's past work laid groundwork in natural language processing and vision models. Together, they focus on pushing AI boundaries through proven methods and real-world deployments.
Unique Path to Self-Improvement
In a Zoom interview after the announcement, Socher explained their strategy. Many labs chase recursion, but Recursive stands out by using open-endedness to reach true recursive self-improvement. No one has fully cracked it yet. Common efforts involve AI tweaking other systems, like improving code or text, but that falls short of full recursion.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Their core aim involves automating the full cycle of research: coming up with ideas, coding them, and testing them. This starts with AI refining itself, building awareness of its limits. Over time, it could extend to all research, even physical areas. Open-endedness draws from biology, where species adapt and counter-adapt over eons, leading to complex traits like eyes.
Rocktäschel's rainbow teaming builds on red teaming. In AI safety, red teaming tests models against harmful prompts, such as bomb-making instructions. Humans craft examples slowly. Instead, one AI attacks another, iterating millions of times across angles. This co-evolution strengthens defenses, a method now adopted widely.
No Finish Line and Lab Differences
Improvement never ends fully. Intelligence can grow in programming, math, and more, with vast upper limits far off. Socher works on formalizing those bounds. As a research outfit, often called a neolab, they differ from big labs by centering on open-endedness. Their team has published papers and shipped products for a decade.
Socher resists the pure lab label. He envisions a company with products users love, benefiting society. Timelines have accelerated due to progress. First products arrive in quarters, not years.
Once achieved, self-improving AI shifts priorities to compute power. Speed determines improvement rate, sidelining human input. Compute allocation becomes key: how much for cancer versus viruses? It poses major global questions.
Backers include Greycroft and GV, supporting this push toward superintelligence.