Jeff Dean, Google's 30th employee and one of the architects of its core search infrastructure, is leaving the company after 27 years to launch Discovery Loop, an AI startup aimed at automating scientific research. He is taking three senior researchers with him: Sanjay Ghemawat, a Google senior fellow; Quoc Le, a founding member of Google Brain; and Oriol Vinyals, a senior research scientist at Google DeepMind. Dean will serve as CEO.
The initial funding round, announced Wednesday, is co-led by Radical Ventures and Khosla Ventures, with participation from Kleiner Perkins, Lightspeed, Doerr Capital, and — notably — Alphabet itself. The round size was not disclosed. Discovery Loop is incorporated as a public benefit corporation and pitches itself as an infrastructure play for AI-driven science, running thousands of simulated and physical experiments in parallel with minimal human input in the loop.
The core thesis is recursive: use AI systems to run experiments, learn from the results, and iterate faster than a human research team can. Taken to its logical end, that includes using AI to help design more powerful AI — recursive self-improvement, in the field's language — which would remove the human iteration bottleneck entirely.
Key facts
- 01Jeff Dean is leaving Google after joining in 1999 as the company's 30th employee, taking three senior researchers with him.
- 02Co-founders include Sanjay Ghemawat, Quoc Le (a founding member of Google Brain), and Oriol Vinyals from Google DeepMind.
- 03The startup, Discovery Loop, is a public benefit corporation aimed at automating scientific experimentation using AI.
- 04The initial round is co-led by Radical Ventures and Khosla Ventures, with Kleiner Perkins, Lightspeed, Doerr Capital, and Alphabet participating.
- 05Dean will serve as CEO and plans to pursue recursive self-improvement, using AI to build more powerful AI.
Dean's departure is the highest-profile Google research exit in years. He joined the company in 1999 and helped build the crawling and indexing systems that made Google search work at scale, then MapReduce, then TensorFlow, then Google Brain. Most recently he shaped the multimodal architecture behind Gemini and served as chief scientist across Google's AI organization. Losing him alongside Ghemawat, Le, and Vinyals removes a meaningful slice of the institutional knowledge that produced modern Google AI.
Alphabet writing a check into the startup that just poached its top researchers is unusual but not without precedent — it preserves a strategic option on the technology while letting the founders operate outside Google's constraints. The parallel to how Anthropic split from OpenAI is imperfect but instructive: senior researchers leave to pursue a specific technical bet, and the mothership hedges by investing rather than fighting.
AI-for-science has been a stated ambition of every frontier lab for years, but commercial traction has lagged the marketing. Google DeepMind's AlphaFold remains the clearest success case. The pitch that a specialist company can move faster on automated experimentation than a research arm inside a $2T advertising business is plausible on its face, and the caliber of the founding team makes it fundable on day one.
Discovery Loop framed the problem in its launch materials as a bottleneck in how science gets done. "While science and engineering have tremendously advanced society over past centuries, progress has traditionally relied on slow, sequential human iterations, creating a significant bottleneck," the company said, describing its systems as tools to "automate complete experimental loops."
The specifics of what Discovery Loop will build first — which scientific domain, which experimental modality, what the go-to-market looks like — were not disclosed. The company's positioning suggests a horizontal platform rather than a vertical bet on drug discovery, materials science, or fusion, though any of those could be a first landing spot. The Alphabet backing implies at least some coordination with Google's existing science efforts.
The counterweight to the pitch is the same one that has dogged AI-for-science for a decade: automating the experimental loop requires the physical experiment to be cheap, fast, and instrumented. Most of biology, chemistry, and materials science does not clear that bar today. Software-only domains — including AI research itself — are the exception, which is why recursive self-improvement is on the roadmap. Whether Discovery Loop can pick a wedge where the loop actually closes is the open question.
Google's Gemini roadmap, its post-training research, and its multimodal work will continue without Dean, and DeepMind still has Demis Hassabis, who was elevated to Alphabet chief scientist earlier this month. But the departure lands during a stretch where frontier talent has become the scarcest input in AI, and where every senior researcher who leaves a large lab now commands a nine-figure round on their next move. The market for founding teams has repriced.
Dean's move is a bet that a small, focused team with unlimited compute and a specific mission can out-execute a research org of thousands on a defined problem. That thesis has worked before — it is essentially the OpenAI story circa 2019 and the Anthropic story circa 2021. If Discovery Loop can pick a scientific domain where the experimental loop is tractable and demonstrate real automation within 18 months, it becomes one of the most important AI companies to watch. If it cannot, it joins the long list of well-funded AI-for-science efforts that produced papers rather than products.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
