Four of Google's Most Celebrated AI Scientists Are Leaving to Build a Machine That Does Science Itself

Key points

• Jeff Dean, co-founder of Google Brain and technical co-lead on Google's Gemini AI model, is leaving Google after almost 27 years to start a new company called Discovery Loop.
• Three other senior Google AI scientists are joining him: Sanjay Ghemawat, Oriol Vinyals and Quoc Le.
• Discovery Loop wants to build AI that runs the scientific method on its own, proposing experiments, running them, and learning from the results without human involvement at each step.
• Google will take a stake in the new company and supply computing power for the first year, but the founders are leaving full-time employment there.
• Investors Khosla Ventures and Radical Ventures have backed the company; the funding total and valuation have not been disclosed.

Who is leaving, and why does it matter?

Jeff Dean is to Google what a head chef is to a three-Michelin-star kitchen. He helped build the search infrastructure millions of people use every day and co-led the team behind Gemini, Google's flagship AI model. Leaving with him are Sanjay Ghemawat, his long-time collaborator and fellow architect of Google's core computing systems; Oriol Vinyals, vice president of research at DeepMind (Google's AI research division); and Quoc Le, the scientist behind AutoML-Zero, a project that lets machine learning, meaning software that improves itself by spotting patterns in data, design new AI tools automatically.

Alphabet CEO Sundar Pichai met with the group multiple times trying to keep them. He couldn't. "In a large organisation there is always a lot of inertia you have to overcome to make any radical changes," Vinyals told Wired AI. "We want to build something different."

Google gets a consolation prize of sorts: an equity stake, cloud computing credits for the startup's first year, and a research collaboration. It is still a significant loss.

What will Discovery Loop actually do?

The core idea is to automate the scientific method. A scientist normally forms a hypothesis, designs an experiment, runs it, checks the results, and starts over. Discovery Loop wants software to handle every step of that loop, thousands of times in parallel, without waiting for a human to sign off between rounds.

The first target is the company's own backyard: improving the machine learning algorithms it will use to build its AI system. Better algorithms feed back into the loop, making the next round of experiments faster and smarter. "It might be that we will discover a different transformer architecture," Le said, referring to the underlying design that powers most modern AI models including ChatGPT.

After that, the team plans to point the same loops at chip design, biology, drug discovery and materials science. Small teams using Discovery Loop, the founders argue, could eventually match or beat the output of entire corporate research departments.

Should ordinary people care about this?

For now, this is a story about a startup that has not yet hired staff or rented office space. Its ambitions are enormous; its current head count is four.

But the direction of travel matters. If Discovery Loop delivers even a fraction of what it promises, it changes who gets to make scientific breakthroughs. Today that requires thousands of researchers and billions in funding. A tool that automates the experiment cycle could shrink that dramatically.

Whether four brilliant engineers can build that tool is, of course, the whole question. "Building something to solve all those problems is a powerful idea," said Radical Ventures managing partner Jordan Jacobs, who is joining the board. "These people have been doing this kind of work in the past so they know what they're doing."

Venture capitalist Vinod Khosla, whose firm Khosla Ventures is also backing the company, put it plainly: "Humans have been using AI to do research, not using AI to be a researcher."

That single sentence is the entire bet.