AI Builds AI: What Happens Next?

Richard Socher, a prominent figure in the field of artificial intelligence, known for founding the early chatbot startup You.com and his foundational work on Imagenet, has launched a new venture. He is now leading Recursive Superintelligence, a San Francisco-based AI startup that emerged from stealth with an impressive $650 million in funding this week. This new company positions itself among the current generation of research-focused AI enterprises.

Socher is joined in this ambitious endeavor by a distinguished group of AI researchers, including Peter Norvig and Cresta co-founder Tim Shi. Their collective mission is to develop a recursively self-improving AI model—a long-sought "holy grail" in contemporary AI research. This model would possess the unprecedented ability to autonomously identify its own deficiencies and redesign itself to rectify them, entirely without human intervention.

Following the launch, I connected with Socher via Zoom to delve into Recursive’s distinctive technical approach and understand why he views this project as distinct from a "neolab," an informal term often applied to new AI startups prioritizing pure research over product development. This interview has been edited for both length and clarity.

When questioned about the increasing prevalence of recursion as a goal across various AI labs, Socher articulated Recursive's unique strategy: "Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved. It’s an elusive goal for a lot of people." He clarified that many mistakenly equate "auto-research"—where an AI improves an external system or task—with true recursive self-improvement, emphasizing that the latter is a much deeper form of autonomous evolution.

He further elaborated on their core objective: "Our main focus is to build truly recursive, self-improving superintelligence at scale, which means that the entire process of ideation, implementation, and validation of research ideas would be automatic."

Initially, this automation would target AI research ideas, eventually extending to all forms of research, even those in physical domains. Socher highlighted the particular potency of "AI working on itself, and it's developing a new kind of sense of self-awareness of its own shortcomings."

Regarding the technical term "open-endedness," Socher confirmed its specific meaning, referencing co-founder Tim Rocktäschel, who previously led open-endedness and self-improvement teams at Google DeepMind. Rocktäschel's work on the world model Genie 3 serves as a prime example of open-endedness, capable of interactively creating "any concept, any world, any agent" upon instruction.

Socher drew an analogy to biological evolution, where "animals adapt to the environment, and then others counter-adapt to those adaptations. It's just a process that can evolve for billions of years, and interesting stuff keeps happening, right? That's how we developed eyes in our [heads]." This illustrates the continuous, emergent nature of open-ended systems.

Another concept from Tim Rocktäschel's work, "rainbow teaming," was introduced as a further illustration. Socher first explained "red teaming" in the context of Large Language Models (LLMs), where the aim is to prevent an LLM from generating harmful instructions, such as "how to build a bomb."

While human red teaming is resource-intensive, Socher proposed a more efficient approach: "What if you tested this first AI with a second AI, and that second AI now has the task of making the first AI [try to] say all the possible bad things. And then they can go back and forth for millions of iterations."

This method allows two AIs to "co-evolve," with one continuously challenging the other from "many different angles," hence the "rainbow analogy." This process "inoculates the first AI," leading to enhanced safety. Socher noted that this innovative idea from Rocktäschel is now widely adopted in major AI labs.

Addressing the question of when such a system would be "done," Socher conceded, "Some of these things will never be done. You can always get more intelligent. You can always get better at programming and math and so on." He acknowledged the existence of theoretical bounds on intelligence, which he is currently working to formalize, but described them as "astronomical," indicating humanity is "very far away from those limits."

When pressed on whether Recursive's distinct approach implies a belief that major labs won't achieve recursive self-improvement with their current methods, Socher refrained from commenting on competitors. Instead, he emphasized Recursive's unique strategy: "We really embrace the concept of open-endedness, and our team is entirely focused on that vision." He highlighted the team's decade-long research and publications in this specific area, along with their proven track record of significant advancements and successful product delivery, citing Tim Shi's role in building Cresta into a unicorn and Josh Tobin's early contributions to OpenAI, leading their Codex and deep research teams.

Socher admitted to struggling with the "neolab" categorization, stating, "I feel like we're not just a lab. 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."

Regarding product timelines, Socher indicated that the team's rapid progress might accelerate initial assumptions. "Yes, there will be products, and you’ll have to wait quarters, not years," he confirmed, signaling an imminent product launch.

Finally, reflecting on the idea that recursive self-improvement could make compute the sole critical resource, leading to a "processing power race," Socher affirmed the importance of compute. He posited that in the future, a key societal question will be: "how much compute does humanity want to spend to solve which problems?" He offered an example: "Here’s this cancer and here’s that virus — which one do you want to solve first? How much compute do you want to give it?" He concluded that this will ultimately become a crucial matter of global resource allocation, one of the most significant challenges humanity will face.