Jeff Dean has spent the better part of three months building a company around the idea that AI can run the loop of scientific discovery for itself, faster and at a scale no team of humans could match. But at a Stanford event on October 28, the man who left Google’s chief scientist job to build exactly that company pushed back hard on the idea that the loop has no speed limit.
Speaking with Stanford professor Fei-Fei Li, Dean addressed the gap between what he’s actually building at Discovery Loop and the more breathless version of the recursive self-improvement story that has been circulating through Silicon Valley.
“Some people have this illusion that you’re gonna be able to run this loop infinitely fast, and I just find that kind of implausible,” Dean said. “And it’s good to be grounded in reality. Having automated systems that can try to improve things is definitely going to improve the rate at which we can improve those systems, but it’s not gonna make them go infinitely fast.”
The RSI hype problem
The comment lands in the middle of a year when recursive self-improvement, or RSI, has gone from a fringe theoretical concern to a term thrown around on earnings calls. Google DeepMind researchers have said they’re seeing “first signs” of self-improvement in how newer models help train their successors, and Sam Altman has written that OpenAI believes it has crossed into a “larval version” of the same phenomenon. Not everyone at the frontier labs agrees on the pace, though — an OpenAI researcher said as recently as earlier this year that the lab doesn’t have self-improving AI yet and won’t be seeing a fast takeoff either, arguing progress will be “extremely gradual across many years,” which is roughly the camp Dean’s Stanford remarks place him in.
Dean’s framing at the event drew a sharper line than most. He told Li that Silicon Valley tends to conflate two different things: self-improvement, where a system is handed a measurable goal and tries different approaches until it hits that target, and recursive self-improvement, where the system doing the improving also gets better at the business of improving itself. He pointed to his co-founder Quoc Le’s decade-old work searching for neural architectures as an early, narrow version of the first kind, and described how Gemini’s generations actually came together — researchers running small-scale experiments, some ideas failing, the ones that worked feeding into the next model. Automating that cycle speeds it up, he said, but it doesn’t remove the physical steps in between: computation has to run, results have to be checked, and another model has to be used to decide what comes next. Those steps, in his telling, are exactly why a loop can’t run arbitrarily fast no matter how much of it gets automated.
What Dean has actually built since leaving Google
Dean announced his exit from Google on August 5, after 27 years at the company and nearly 15 years running its AI research efforts — a departure significant enough that Google’s stock dropped roughly 4% the day it became public. He didn’t go to a rival lab. Instead, he, Sanjay Ghemawat, Oriol Vinyals and Quoc Le — a group whose combined Google tenure runs to something like a century, and whose former collaborators populate the founding teams of half the AI industry — started Discovery Loop, a public benefit corporation with Google itself on board as a founding investor and cloud partner. The framing wasn’t entirely new: Dean had described the same loop of proposing, running and evaluating experiments at a YC event roughly two weeks before the announcement, without telling anyone he was about to build a company around it.
At Stanford, Dean gave his fullest account yet of what the company has built in its first three months. The team has been assembling what he called an “agent harness” — shared infrastructure built for AI-first collaboration rather than adapted from tools made for human teams — and currently has roughly 20 automated improvement loops running against its own machine learning research and engineering work. The stated goal is to lift the productivity of a small number of domain experts by 20 to 50 times, not to replace the humans directing the work.
The roadmap starts narrow and expands outward: machine learning research and engineering and hardware design first, since all three can be run end-to-end in digital form, before the company moves into biology, drug discovery, materials design and eventually healthcare, which Dean described as a long-standing personal interest he’s deliberately holding off on until the core loop is working.
Grounded optimism, not a ceiling
None of this reads as Dean downplaying what automation can do — Discovery Loop’s entire premise is that shortening and parallelizing the experimental loop is worth leaving Google over. The Stanford comments are better read as him drawing a line between accelerating a process that still has to run through real computation, real evaluation and real infrastructure, and the “illusion,” as he put it, that the same loop could somehow be freed from those constraints altogether. For a founder whose company is valued in the tens of billions on the promise of automating science, that’s a notably specific thing to be “grounded” about.