Google’s AI brain trust is walking out the door to build a science accelerator

Google’s AI brain trust is walking out the door to build a science accelerator

Jeff Dean, one of Google’s earliest engineers and a driving force behind its AI rise, is leaving after 27 years. He is taking a handful of Google’s top AI minds with him to start Discovery Loop, a new company that wants to use AI to speed up scientific breakthroughs.

The team includes Sanjay Ghemawat, a senior fellow who helped design Google’s core systems, Quoc Le, a founder of Google Brain, and Oriol Vinyals from DeepMind. Their plan is to build AI that can run thousands of experiments at once, automating much of the trial-and-error that slows down research today. They even talk about AI that improves itself with less human input, pushing the limits of what machines can discover on their own.

Discovery Loop is set up as a public benefit corporation, meaning profit is not the only goal. Alphabet, Google’s parent, is among the backers, alongside venture firms like Radical Ventures and Khosla Ventures. The company says its mission is to move beyond AI that answers questions to AI that makes new discoveries.

Dean joined Google in 1999 as its 30th employee and shaped its search infrastructure and AI models like Gemini. His departure with other key figures signals a shift in how top AI talent is choosing to pursue ambitious ideas, outside the constraints of a big tech giant.

This matters because faster, AI-driven research could lead to breakthroughs in medicine, energy, or materials science years sooner than the traditional human-led process. But it also raises questions about whether AI should be making decisions about what to study next, and who gets to control that direction.

Next, watch for Discovery Loop’s first partnerships with research institutions and the kinds of experiments it tackles early on. The team has not yet shared details about its first projects or when its tools will be available to others.

What kind of scientific problems should AI be trusted to solve on its own

How much of a setback is this for Google’s AI ambitions


Filed under: AIDiscovery, JeffDean, GoogleAI, DeepMind, ScienceTech

Comments