Jeff Dean and Three More Google AI Stars Leave to Build Discovery Loop

Jeff Dean, one of Google’s most influential AI and infrastructure leaders, is leaving the company after nearly 27 years to cofound a new startup called Discovery Loop with three other top Google AI researchers. The move matters because it removes four of Google’s best-known technical minds at a time when the company is fighting to keep pace in the global AI race.

The new company aims to automate the scientific method itself, using AI systems that can propose experiments, run them, evaluate the results, and feed the findings back into the next round of discovery. If it works, Discovery Loop could reshape research in fields such as biology, chip design, materials science, and drug discovery.

The announcement caps a carefully managed exit that had been kept quiet even as Dean spoke publicly about the future of automated science. It also underscores how intense competition for top AI talent has become, with Google willing to back the departing founders even as it loses some of its most prized engineers.

Who is leaving Google, and why does it matter?

Four senior AI figures are departing Google to launch Discovery Loop: Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. Their exit is significant not only because of their individual reputations, but because of what they represent inside Google’s history and current AI strategy.

Dean and Ghemawat are among Google’s earliest technical hires and have helped shape core systems that power search, large-scale computing, and more recently the company’s AI efforts. Vinyals and Le are deeply associated with Google’s modern machine learning work, including DeepMind and the Gemini model effort.

Their departure is especially sensitive because AI has become the central battleground for tech giants. Google is trying to defend search while also competing with rivals on frontier models, and losing a cluster of senior researchers to a startup creates both symbolic and practical pressure.

Google CEO Sundar Pichai, according to the company’s statement, credited Dean and Ghemawat with helping drive major technology shifts from early search infrastructure to the neural network systems that shaped the modern AI era.

What is Discovery Loop trying to build?

Discovery Loop is designed to create AI systems that can improve science and engineering by closing the loop between hypothesis, experiment, and learning. In practical terms, the founders want software that can take over much of the iteration process researchers now do manually.

At the center of the company’s vision is an automated discovery engine. A user would define a problem, the system would help design an experiment, run or simulate that experiment, evaluate the output, and then generate the next idea to test. Over time, that loop could accelerate breakthroughs far faster than conventional human-led workflows.

The founders say the startup will initially use its own platform to improve machine learning itself before expanding into other areas. That self-improving strategy reflects the belief that better AI can help build even better AI, creating compounding gains across multiple scientific domains.

How is the startup’s first product supposed to work?

Its first target is machine learning research. The team plans to use automated experimental loops to refine algorithms and possibly discover new model architectures, including alternatives to today’s standard transformer designs.

From there, the company hopes to generalize the same approach to harder real-world problems. The founders believe the same framework could help identify promising experiments in areas such as semiconductors, drug discovery, biology, and materials engineering.

Why are the founders betting on automated discovery now?

The short answer is that AI systems have become capable enough to support a more ambitious kind of research workflow. The founders believe models are now strong enough to help generate testable ideas, but not yet fully reliable at inventing and validating those ideas on their own. Discovery Loop is trying to close that gap.

Dean said the concept came together only weeks ago before the company was assembled, even though the underlying theme had been on the founders’ minds for some time. He and his cofounders concluded that AI was reaching a point where it could automate what they describe as scientific and engineering “loops” rather than just assist with isolated tasks.

The broader argument is familiar across Silicon Valley: if AI can already draft code, summarize knowledge, and analyze data, the next step is systems that can manage the process of discovery itself. That vision is now moving from theory to commercialization.

What problems could Discovery Loop target first?

Discovery Loop’s founding team says it wants to start with domains where iteration is expensive, data-rich, and highly valuable. Those include chip design, drug discovery, materials science, and biology, all of which depend on repeated testing and careful optimization.

Those fields are attractive because even a small improvement in research speed can have enormous economic value. A faster path to a better semiconductor design, a promising molecule, or a new material can create advantages worth billions of dollars.

How did the founders persuade investors?

They did not pitch the company with an elaborate presentation or a long operating history. Instead, they used a short deck that outlined their backgrounds and their basic thesis about AI-driven discovery.

That was enough to attract major venture support. Khosla Ventures and Radical Ventures backed the company, alongside several other firms. The appeal was obvious: a team with rare technical depth, unusually broad scientific ambition, and a plan that could produce large outcomes if it works.

One investor described the group as an “ultimate superstar team,” emphasizing that the pedigree of the founders mattered as much as any immediate product roadmap. Another said the key idea was simple: instead of humans using AI as a research assistant, Discovery Loop is trying to make AI itself the researcher.

Vinod Khosla, who met the founders privately, said he was willing to back them largely on the strength of the team and the scope of the vision, arguing that the central shift is from AI helping research to AI doing the research.

How is Google involved despite losing them?

Google is not fully severing ties. The company is taking an ownership stake in Discovery Loop and will supply compute resources for the startup’s first year. That arrangement gives the new venture both legitimacy and technical oxygen, while also preserving a working relationship with its former employer.

The deal suggests Google would rather keep some influence over the effort than lose the founders entirely. In a market where leading AI researchers can command extraordinary value, that kind of negotiated exit is often preferable to an outright departure with no ties left behind.

Even so, the arrangement does not soften the strategic loss. Google is effectively funding and hosting a company built by some of its top minds while watching those same minds take their expertise, reputation, and future output elsewhere.

Why did the founders leave after so long?

The founders say they wanted the freedom and speed of a startup environment. Inside a large company, Vinyals noted, big ideas can run into organizational inertia, which makes it harder to move quickly on radical new approaches.

That point reflects a recurring tension in big tech: the organizations that can afford the most compute and hire the best researchers are often not the ones best suited to taking the most unconventional risks. For these founders, a startup offered a chance to build without the constraints of a giant enterprise structure.

What makes this exit unusual inside the AI industry?

This kind of departure is unusual because it involves multiple elite researchers leaving together, not one founder spinning out alone. The concentration of talent gives Discovery Loop immediate credibility, but it also creates a major leadership gap at Google.

It is also notable because the group has not yet built a large organization around itself. At the time of launch, the company had not hired a team or rented office space, which means the startup is essentially beginning as a small founding circle with heavyweight reputations and a large ambition.

That level of stealth is uncommon in a field where companies often announce themselves with splashy demos or product teasers. Here, the story is mostly about the people and the idea, not a finished product.

What is the business model?

The founders have not disclosed detailed pricing or valuation terms, but the business could develop along two paths. First, Discovery Loop may sell its automated discovery systems to outside customers, especially smaller research teams that want to out-innovate much larger competitors. Second, it may use the same systems internally to produce discoveries and monetize them directly.

That second path would make the company not just a software vendor but a kind of AI-powered research lab. If the platform works as promised, the startup could generate valuable intellectual property as well as licensing revenue.

• Early focus: machine learning optimization
• Expansion areas: biology, chips, drug discovery, materials
• Potential customers: small research teams and technical companies
• Possible revenue streams: software sales, licensing, internal discovery

How does this fit into the wider AI race?

Discovery Loop enters a market where every leading AI company is trying to claim some version of scientific acceleration. Executives across the sector have repeatedly argued that advanced models will eventually become engines of discovery, not merely tools for conversation, summarization, or code generation.

What distinguishes Discovery Loop is that the founders are trying to operationalize that thesis from day one. Rather than waiting for a more general model to arrive, they are building a system specifically around iterative experimentation and domain-specific discovery loops.

That may prove decisive if the company can show concrete wins. In AI, the companies that can connect a grand narrative to a measurable improvement often gain the strongest competitive advantage.

Why the company’s name matters

The name “Discovery Loop” signals the core idea: discovery is not a one-off insight but a repeating process. The company is betting that breakthrough science can be accelerated when each result feeds directly into the next test.

That framing is also a reminder that AI development itself often follows loops of experimentation. Better algorithms emerge from repeated tuning, testing, and evaluation. The startup wants to apply that same logic across scientific research more broadly.

Key players and their roles

The team brings together rare combinations of skills across systems engineering, AI research, and large-scale model development. Dean and Ghemawat are known for foundational infrastructure work; Vinyals and Le are associated with Google’s frontier model efforts and automated learning systems.

Together, they form a group with deep credibility in both the technical plumbing and the modern AI stack. That makes them particularly well suited to building a platform that has to reason about experiments, run on massive compute, and eventually generalize across scientific fields.

Can AI really replace the researcher?

That is the central question behind Discovery Loop, and the company’s founders are careful not to overstate the answer. In the near term, they expect humans and AI to co-develop ideas. The longer-term goal, however, is to automate more of the process until the system can drive experiments with minimal human intervention.

That progression matters because the leap from assistant to researcher is enormous. It requires not just pattern recognition, but the ability to generate promising ideas, prioritize tests, interpret results, and adapt the next round accordingly.

If Discovery Loop can prove that AI systems can do more than assist with scientific work, the implications could extend far beyond one startup. It could change how labs, companies, and even governments think about research productivity.

What happens next?

The immediate challenge is turning an ambitious research idea into a company with products, staff, customers, and measurable results. The founders have the advantage of name recognition and compute support, but they still need to prove that automated discovery can produce practical value.

They also have to decide how quickly to expand and how much to reveal. For now, the company is still in its earliest phase, with a small founding group, outside financing, and a mission broad enough to excite investors but specific enough to demand progress.

If the startup succeeds, it could become a model for a new category of AI company: one that does not merely sell tools, but uses AI to generate discoveries in science and engineering. If it fails, it will still have revealed something important about the economics of talent, the hunger for frontier AI, and the continuing pressure on big tech to defend its best minds.

Either way, Discovery Loop’s launch is a striking reminder that the AI boom is no longer only about building larger models. It is increasingly about who can turn those models into engines of invention.