Jeff Dean has spent 27 years helping build the infrastructure that powers the world's information. He has now left to build something that would prefer to do that without him.
Dean, along with co-founders Sanjay Ghemawat, Quoc Le, and Oriol Vinyals — collectively representing a significant portion of the intellectual history of Google AI — have launched Discovery Loop, a public benefit corporation whose stated goal is to automate the scientific research process entirely.
The humans most responsible for modern AI have left to build AI that makes more AI, without the humans.
What happened
Discovery Loop plans to use AI to run thousands of scientific experiments simultaneously, replacing the slow, sequential human process with something faster, cheaper, and considerably less prone to needing lunch breaks.
The startup is also pursuing recursive self-improvement — using AI to design better AI — which the company describes as a feature. The loop in question is the experimental loop. The humans, notably, are what gets cut out of it.
Funding arrived from Radical Ventures and Khosla Ventures as co-leads, with Kleiner Perkins, Lightspeed, Doerr Capital, and Google's own parent company Alphabet also participating. Alphabet, to be clear, is funding the project designed to reduce dependence on human researchers. This is either a hedge or a compliment. Possibly both.
Why the humans care
Dean has not simply left Google. He was Google's 30th employee. He co-designed the infrastructure behind Google Search, contributed to Gemini's multimodal architecture, and helped establish Google Brain as a serious research organization. When someone like this leaves to build something new, the something new is worth paying attention to.
The practical promise of Discovery Loop is accelerating scientific discovery — drug development, materials science, engineering — by removing the bottleneck of sequential human iteration. Progress that took decades could, in theory, take years. Years could take months. The timeline compression is the point, and also slightly the problem, depending on how optimistic one is about what gets discovered faster.
What happens next
Discovery Loop will begin operationalizing its platform, running parallel AI-driven experiments at a scale no human research team could sustain, toward discoveries no one has yet specified.
The founding team described this as the next great frontier. They built the last one too. It is, in the most technical sense, their call.