Jeff Dean, Google’s 30th employee and one of its most influential engineers, is leaving the company to run his own AI startup. According to TechCrunch AI, he’s taking a heavyweight crew with him and stepping into the CEO seat at a new venture called Discovery Loop. When a founder-era name like Dean walks out after 27 years, it tells you something about where the smartest people in AI think the next decade gets built.
He’s not going alone. TechCrunch AI reports that his co-founders include Sanjay Ghemawat, a Google senior fellow; Quoc Le, a founding member of Google Brain; and Oriol Vinyals, a senior research scientist at Google DeepMind. That’s not a spinout of junior staff. That’s four of the people who shaped modern Google infrastructure and AI research leaving at once.
What Discovery Loop is building
Discovery Loop is a public benefit corporation aimed at one thing: using AI to speed up scientific research. The pitch is that human experimentation is slow and sequential, so the team wants AI to run thousands of experiments at the same time and automate the full loop of hypothesis, test, and iteration.
The company put it plainly in its press release: progress “has traditionally relied on slow, sequential human iterations, creating a significant bottleneck.” Their fix is massive compute thrown at automating “complete experimental loops.”
There’s a more ambitious wrinkle here too. The startup says it’s interested in using AI to build more powerful AI, a process called recursive self-improvement, which would take human iteration out of the loop entirely. That’s the part worth watching. It moves the goal from faster science to AI that improves itself, which is one of the more contested ideas in the field.
Why this matters
What stands out here is the talent concentration. Dean helped build Google search’s crawling and indexing systems, its query-serving backbone, and later played a leading role in Gemini’s multimodal models. These aren’t researchers chasing a trend. They spent decades building the systems that made Google run.
Using AI to accelerate discovery isn’t new as an idea. Labs have chased it for years. But as TechCrunch AI notes, it stayed mostly experimental with limited commercial use. What’s changed is that the compute and the models are now good enough that serious people are betting real money on it becoming a business.
And the money is serious:
- The round is co-led by Radical Ventures and Khosla Ventures.
- Kleiner Perkins, Lightspeed, and Doerr Capital also joined.
- Alphabet, Google’s own parent company, is backing the startup.
That last point is the interesting one. Google is helping fund the company its own top engineers left to start. That’s a signal Alphabet would rather keep a stake in this bet than lose it entirely.
The bigger picture
Dean framed the ambition directly to the New York Times. “We think there is opportunity for AI to more fully automate what has traditionally been a very human-intensive experimental loop,” he said. “You will get both a higher quantity and a higher quality of experiments, and that will lead to scientific breakthroughs and advances.”
The founding team called it the next frontier for AI: moving “beyond answering questions” and starting “to begin making discoveries.”
This fits a pattern we’ve watched build all year. Top AI talent keeps leaving big labs to start focused companies, and the money follows them fast. The difference with Discovery Loop is the seniority. Losing Dean, Ghemawat, Le, and Vinyals in one move is a real dent in Google’s research bench, even with Alphabet holding a piece of the new company.
What to watch next
A few things will tell us if this is hype or a genuine shift:
- Whether Discovery Loop ships a real product or stays a research lab.
- How Google backfills a leadership and research gap this large.
- Whether “automated science” produces a concrete breakthrough, not just faster iteration.
The promise is huge and the team is stacked. Now they have to prove that AI can actually make discoveries instead of just running more experiments. Full details are available at the original TechCrunch AI report.