Startupspublished

Jeff Dean Leaves Google to Build Discovery Loop, a Science AI Startup Backed by Radical Ventures

The new public-benefit company has an unusually experienced founding team and support from both Radical Ventures and Alphabet. Its central technical question is still unanswered: what system will turn the ambition to automate research into a usable product?

By 4 min read
Jeff Dean Leaves Google to Build Discovery Loop, a Science AI Startup Backed by Radical Ventures
Jeff Dean Leaves Google to Build Discovery Loop, a Science AI Startup Backed by Radical Ventures

Listen to this story

The audio brief

About 1:35
0:001:35
Read transcript
Jeff Dean has left Google to build Discovery Loop, a new public-benefit corporation aimed at automating machine learning, science, and engineering. The company is backed by Radical Ventures, with Alphabet as a founding investor and Google Cloud as a partner. But the most important detail is what Discovery Loop has not revealed: there is still no public product, research workflow, or technical approach. Dean is joined by Sanjay Ghemawat, Quoc Le, and Oriol Vinyals—four researchers with deep Google experience. Dean spent twenty-seven years at the company, served as its chief scientist, and worked on TensorFlow and large-scale distributed computing. Vinyals announced on August fifth that he was leaving Google to join the venture. That gives Discovery Loop an unusually credible founding team, but not yet a way to judge the business. Automating research could mean software for scientists, autonomous research systems, engineering tools, or something else entirely. The company has described the destination, not the operating mechanism: how it will generate ideas, test them, validate results, and deliver useful output. There is also a major financial claim to treat carefully. Business Insider reported that Discovery Loop was seeking one billion dollars at roughly a ten-billion-dollar valuation. Neither figure has been confirmed, so they are reported fundraising terms, not evidence of demand. The next meaningful fact is narrow and concrete: what product will Discovery Loop release, and how will it prove that automation improves scientific or engineering work?

Story brief

3 key points

Discovery Loop is being formed as a public-benefit corporation to automate machine learning, science, and engineering, but it has not yet disclosed a product or technical approach. Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals bring deep Google research experience, while Radical Ventures is backing the company. Alphabet is a founding investor and Google Cloud a partner, preserving ties to Google’s...

  1. 01

    Discovery Loop’s mission is public, but its research workflow, product category, and evaluation method are not.

  2. 02

    Jeff Dean spent 27 years at Google and formerly served as its chief scientist.

  3. 03

    Vinyals announced his Google DeepMind departure on August 5 to join Dean, Ghemawat, and Le.

Jeff Dean has left Google to co-found Discovery Loop, a new AI company with Sanjay Ghemawat, Quoc Le and Oriol Vinyals. The startup is focused on accelerating scientific research, has backing from Radical Ventures, and is entering the market before it has publicly described a specific product. That makes its founding team and mission unusually clear, while the operating mechanism remains unresolved.

The mandate comes before the mechanism

Discovery Loop is structured as a public-benefit corporation. Its stated mission is to automate machine learning, science and engineering. Those terms describe a broad target: work that produces and applies new knowledge, including the process of developing machine-learning systems themselves.

That is a mission statement, not yet a disclosed technical design. Discovery Loop has not made specific products public, so there is no public basis yet to judge whether it will supply research software, autonomous research systems, engineering tools, or another form of scientific AI. The distinction matters: a company can aim to shorten research cycles without having shown how it will generate, test, validate or deliver results.

What is known at launch

  • Dean’s co-founders are Ghemawat, Le and Vinyals, all longtime Google researchers.
  • Dean spent 27 years at Google and was formerly its chief scientist; his work included TensorFlow and large-scale distributed computing.
  • Radical Ventures is backing Discovery Loop. Alphabet is identified as a founding investor, and Google Cloud as a partner.

A startup still connected to its former home

The Alphabet investment and Google Cloud partnership make Discovery Loop neither a conventional internal Google project nor a clean break from Google’s infrastructure orbit. The arrangement gives the new company a declared connection to its founders’ former employer while its founders pursue a separate corporate structure and research mandate.

Vinyals publicly said on August 5 that he was leaving Google DeepMind to build Discovery Loop with Dean, Le and Ghemawat. His departure follows Dean’s exit after a long career at Google, concentrating senior technical experience inside a company that remains at an early stage.

The Google backdrop is changing too

Discovery Loop’s formation coincides with a reshaped leadership map at Google. Demis Hassabis has become chair of Google DeepMind and Alphabet chief scientist, giving up day-to-day operating responsibilities. Koray Kavukcuoglu now holds responsibility for Gemini models, frontier research, the Gemini app and developer teams, reporting to Sundar Pichai.

The product sequence adds context without proving a cause for the founders’ departures. Alphabet had expected Gemini 3.5 Pro in June; in July, Pichai said it was in testing. Google then released Gemini 3.7 Flash in August, positioning it around coding and agentic workflows and pricing it initially below the original cost of 3.6 Flash.

Capital is a signal, not an operating result

A reported fundraising target has added to the attention around the venture. Business Insider reported that Discovery Loop was seeking $1 billion at an approximate $10 billion valuation, according to an account that emphasized the discussions were not a completed or publicly confirmed transaction. Those figures should therefore be treated as reported fundraising terms, not a settled valuation or proof of commercial demand.

For now, the company’s proposition rests on a defined mission, a high-profile founding group and unusual ties to both a venture investor and Alphabet. The meaningful next evidence will be narrower: a public product, the research task it is designed to handle, and a way to evaluate whether automation actually improves scientific or engineering work.

Editorial analysis

Our Read

Our read: Discovery Loop is a sharper test of the small-team thesis than of scientific AI itself. The founders have moved from Google DeepMind into a company whose stated remit spans machine learning, science and engineering, while Alphabet remains a founding investor and Google Cloud partner. That combination may give the venture room to move independently without starting from zero on infrastructure. The next material signal is a product description that identifies the research workflow it will automate and how users will assess its results.

Sources

  1. tekedia.comJeff Dean’s New AI Startup Targets $10 Billion Valuation - Tekedia
  2. marketscale.comJeff Dean and top Google researchers quit Alphabet to launch Discovery Loop, backed by Radical Ventures