AI Pioneer Jeff Dean Leaves Google for Discovery Loop Startup
For decades, machines have aided research. They compute, they simulate, they analyze. But the scientific method itself has always passed through a human mind. A scientist proposes an experiment. The scientist implements what is needed to run it. The scientist evaluates what happened. Then the scientist decides what to try next. That loop is the engine of discovery. It is also the bottleneck. Jeff Dean, chief scientist of Google DeepMind and Google Research, described the loop in public on July 25. [2] He stood before 6,000 would-be founders at Y Combinator’s Startup School at San Francisco’s Chase Center. [4] “You propose an experiment, you implement what you need to run the experiment, you evaluate the experiment, and then you get results from that,” he said. [2] His argument: run thousands of these loops automatically, and breakthroughs follow. The domains he named were science, biology, chip design, and even the improvement of AI models themselves. Through all those decades, the researcher has remained human. Vinod Khosla, the venture capitalist, later put his finger on exactly that gap. “Humans have been using AI to do research, not using AI to be a researcher,” he said. [7] The fundamental thing in the new venture, he added, is that AI is the researcher.
The crowd at the Chase Center did not know what they were hearing. Dean was describing a startup he had just organized in secret. It is called Discovery Loop. Even though Google took a stake in the new company, the departure is a devastating blow to the search giant as it attempts to keep pace in the frantic AI model competition. After almost 27 years, Dean has left Google. [5] He left with Sanjay Ghemawat, his frequent collaborator. Together they wrote significant parts of Google’s computing and search infrastructure. Both were among Google’s first hires. The departure has been compared to Mick Jagger and Keith Richards ditching the Rolling Stones to start a new band. The other two cofounders are just as celebrated. Oriol Vinyals was vice president of research at DeepMind and a technical lead for Gemini. [1] Quoc Le cofounded Google Brain and was the key scientist behind AutoML-Zero, a project that uses machine learning algorithms to autonomously build AI products. [6] A running joke in geekdom is a list of surreal “facts” about Dean’s technical prowess. One holds that Jeff Dean’s resume lists the things he has not done, because it is shorter that way. Dean is also cofounder of Google Brain and technical colead on the company’s flagship model, Gemini. This is the team that now says the researcher can be automated.
The idea itself came together only weeks before the launch. Dean says the four were working on different things. They were all starting to see the same possibility: AI automating scientific and engineering loops. They are not just long-time colleagues. They are friends who vacation together. “We were working on different things, but we were all starting to see the possibility of AI being able to automate scientific and engineering loops,” Dean said. [2] The quartet worked up a road map. Discovery Loop would be its own first customer. The pitch deck they showed investors was simple. A few slides with their backgrounds and a rough sketch of the approach. They did not use AI to create it.
From Its Own Algorithms to New Domains
The timeline starts inside the company. The first loops Discovery Loop builds will improve the company’s own advanced machine learning algorithms. Quoc Le is explicit about the prize. “I’m very excited about automating machine learning,” he says. [6] “It might be that we will discover a different transformer architecture.” [6] A transformer is the neural network design at the core of most modern AI models. A different transformer architecture would be a breakthrough in the very structure that powers today’s AI. This is the first measurable effect of the founding. The loop is aimed at the thing that runs the loop.
From machine learning, the road map widens. “As we create the core intelligence and the discovery loop fundamentals, we’ll be able to quickly generalize to other domains,” Vinyals says. [1] The domains the team names are chip design, biology, drug discovery, and material design. The ambition is superhuman advances in each of them. At first, the ideas will be codeveloped with humans. The goal, over time, is to deeply automate the process. Vinyals points to the hardest piece: how models come up with new ideas to try. “That’s not something that currently they’re super strong at,” he says. The loops must learn to propose, not merely evaluate. If it works out, the founders contend, a few dedicated individuals may use Discovery Loop to out-invent the world’s largest research organizations. Those small teams would presumably be customers.
The resources for this timeline were raised in secrecy. Vinod Khosla met the founders on a Saturday at his Sand Hill Road office. The weekend timing meant no one would get a hint of Google’s upcoming loss. Other meetings took place at Dean’s house. “With this team, I wouldn’t need to know what they were doing before I backed them,” Khosla says. [7] “It’s the ultimate superstar team.” [7] Khosla Ventures and an AI-oriented firm called Radical Ventures bought in, along with a handful of other firms. The founders are not sharing the funding amount or the valuation. Some companies pay tens or even hundreds of millions for a single AI superstar. These four are GOATs, the greatest of all time. Radical managing partner Jordan Jacobs will join the startup’s board. “Building something to solve all those problems is a powerful idea,” he says. The negotiation with Google was harder than the fundraising. Alphabet CEO Sundar Pichai tried over multiple meetings to keep the four at the company. [5] In the end, Google gave its blessing, took an investment stake, and arranged to provide computing power for the first year. That computing capacity is the runway for the opening phase.
The Question the Departure Leaves Behind
The mission itself is not novel. Just about every major figure in AI has claimed it at one point or another. Dario has said that AI will become an engine of scientific discovery. Sam has said the same. Jensen has said it as well. All of them predicted that AI would solve problems that humans could never crack on their own. Those claims are words. This team is betting that its knowledge and skills will prevail in building a system that actually makes such breakthroughs. Pichai, in his official statement, acknowledged the scale of what Google is losing. “Over 27 years, Jeff and Sanjay helped to drive some of the most significant technology transitions, from our early search infrastructure to the neural networks that helped create the modern AI era,” he said. [5] Google will continue to work with Discovery Loop as a founding investor and Cloud partner, he added, and will collaborate on a research framework for machine-learning systems and related infrastructure advances.
The company is still in its quiet phase. The founders have not yet begun to hire a team. They have not rented office space. That silence may mean even more defections from Google. The stealth has been deliberate, in part because of the impact this mini-exodus will have on the search giant. Vinyals explained why they left. “In a large organization there is always a lot of inertia you have to overcome to make any radical changes,” he said. “We want to build something different.” Asked who the CEO is, there was a brief pause. “I think I’m the CEO,” says Dean, sheepishly. “Everyone pointed at me.” The company has a name, a road map, funding, and four of the most accomplished engineers in the field.
One question hangs over the entire departure. Will the work Discovery Loop does as a startup be as significant as the talent gap it has created by leaving Google? That is the narrower mark for success the team faces. The startup is set up as a public benefit corporation, a structure that allows a company to pursue public benefit alongside profit. The founders are betting that their knowledge and skills will prevail. A few dedicated individuals, they contend, may use Discovery Loop to out-invent the world’s largest research organizations. Those small teams would presumably be customers. The startup also may make its own loops to seek and monetize breakthroughs. The talent gap is the part that is already real. Whether the work will match it is the part that is not.
Sources
3. Y Combinator
5. Google
6. Google Brain