Jeff Dean, one of Google’s longest-serving technical leaders, is leaving the company after 27 years to co-found Discovery Loop, a public benefit corporation focused on automating scientific and engineering research. He is expected to serve as CEO.

Dean is joined by Google veterans Sanjay Ghemawat, Quoc Le, and Oriol Vinyals. The team says Discovery Loop will use large-scale computation and AI systems to run and refine experiments faster than conventional, human-led research workflows. The move comes as Google reorganizes its AI leadership and faces intense competition for researchers, models, and developer adoption.

Discovery Loop’s proposed research engine

According to reporting by TechCrunch AI, Discovery Loop plans to initiate and iterate thousands of experiments at the same time. Its initial work will focus on machine-learning experiments, with ambitions to apply the same approach across scientific and engineering disciplines.

The company describes its goal as automating what it calls the complete experimental loop: proposing work, running experiments, evaluating results, and using those results to determine the next step. That would move AI systems beyond producing answers or code suggestions and toward managing repeatable research processes.

The founders are also interested in using AI to improve the systems that build AI. That concept, often called recursive self-improvement, remains technically and operationally uncertain. The available evidence does not establish that Discovery Loop has achieved such a system; it shows that the founders identify it as a long-term area of interest.

Discovery Loop is structured as a public benefit corporation, indicating that its stated purpose extends beyond maximizing conventional commercial returns. The company has not publicly detailed its first research program, product roadmap, staffing plans, or expected timeline for deploying its systems.

A significant loss for Google’s AI organization

Dean’s departure removes a senior figure whose work spans Google’s early search infrastructure and later artificial-intelligence efforts. TechCrunch reported that he contributed to systems supporting Google Search and had a major role in the company’s AI research and Gemini multimodal models.

The other founders also bring unusually long histories at Google. Ghemawat is a senior fellow and prominent engineer who worked with Dean on foundational infrastructure. Quoc Le was a founding member of Google Brain and helped advance deep-learning and sequence-to-sequence research. Vinyals is a senior research scientist at Google DeepMind and has held leadership roles connected to Gemini and earlier systems such as AlphaStar.

The departures coincide with a broader change at Google DeepMind. The Decoder reported that Demis Hassabis is stepping back from day-to-day management to become Alphabet’s chief scientist, while former DeepMind chief technology officer Koray Kavukcuoglu is taking over operational leadership. Kavukcuoglu is expected to oversee Gemini development, frontier research, the Gemini app, and Google’s AI developer platforms.

The leadership moves do not by themselves demonstrate a weakening of Google’s research capabilities. However, they make retention and succession more visible issues as Google competes with OpenAI, Anthropic, and well-funded startups for senior technical talent. Other departures cited by The Decoder include DeepMind researcher David Silver and former Gemini co-lead Noam Shazeer, who left for OpenAI.

Evidence and claims behind the startup

The strongest details about Discovery Loop come from company statements reported by TechCrunch AI and The Decoder, rather than from independently verified technical demonstrations. The founders say that faster, more automated experimentation could increase both the number and quality of research iterations, but no benchmark, customer result, or production system was provided in the available evidence.

The initial financing is being led by Radical Ventures and Khosla Ventures, with participation from Kleiner Perkins, Lightspeed, and Doerr Capital, according to TechCrunch. Alphabet is also backing the company. The Decoder reported that the financing is expected to close in the coming weeks and that Alphabet plans to remain involved as a cloud partner while collaborating on a research framework for machine-learning systems and infrastructure.

Those arrangements could give Discovery Loop access to significant computing resources and a close relationship with its former parent. They also raise practical questions about independence, intellectual property, data access, and whether the startup will rely heavily on Google Cloud. The source material does not provide the round’s valuation, total size, ownership terms, or details of any formal technology-transfer agreement.

The broader premise is gaining attention across the AI sector. TechCrunch noted that AI-assisted scientific discovery has attracted interest for years but has only recently begun moving toward larger commercial applications. Anthropic and OpenAI have also discussed more automated research as part of their wider AI strategies, although the sources do not establish a direct competitive relationship between those efforts and Discovery Loop.

Why the move matters to builders and enterprises

For AI builders, Discovery Loop points to a shift from task-level assistance toward systems that coordinate entire workflows. A useful implementation would need more than a capable model: it would require experiment planning, tool use, simulation or laboratory access, data management, reproducibility checks, and reliable methods for rejecting weak results.

That makes infrastructure as important as model quality. Research automation depends on orchestration, evaluation, compute scheduling, and audit trails. In scientific settings, a system that generates many experiments but cannot distinguish robust findings from noise could increase costs rather than accelerate progress.

Enterprise buyers should therefore treat the announcement as an indication of direction, not as evidence that autonomous research is ready for general deployment. The near-term opportunities are more likely to involve bounded workflows, such as running machine-learning evaluations, optimizing engineering parameters, or proposing candidates for human review. High-consequence fields will still require domain experts, validation, and governance.

The startup may also intensify competition for researchers who can combine frontier-model development with scientific or engineering expertise. Google retains major advantages in infrastructure, data, and the Google DeepMind research organization, but Discovery Loop’s small founding group could move quickly if its financing and cloud arrangements are completed as reported.

What to watch next

The first signal will be whether Discovery Loop discloses a concrete research system, technical paper, or demonstration showing automated experiment selection and iteration. Claims about scientific acceleration will be more meaningful when accompanied by reproducible comparisons with conventional research workflows.

The financing close and the terms of Alphabet’s involvement will clarify how independent the startup is operationally. Developers should watch for information about its cloud stack, model-access strategy, hiring, and whether it plans to offer software to outside researchers or operate primarily as an internal research organization.

At Google, the key follow-up is how Kavukcuoglu’s leadership changes the priorities around Gemini, developer platforms, and frontier research. The market will also be watching for further senior departures, new appointments, and evidence that Google can maintain research continuity through the transition.

Creati.ai perspective

Discovery Loop is notable less because it promises another general-purpose AI model than because it targets the structure of research itself. If the founders can make automated experimentation reliable, the important product may be a research operating layer that connects models, tools, simulations, and evaluation systems.

That outcome is far from guaranteed. The central challenge is not generating more hypotheses; it is building closed-loop systems that produce trustworthy, reproducible results at a cost and speed that justify deployment. For now, the event is best read as a high-profile bet on that direction—and as a reminder that Google’s deepest AI advantage is also a source of talent competition.