Jeff Dean Launches Discovery Loop as Last Transformer Co-Author Exits Google

Jeff Dean, for more than a quarter century the face of artificial intelligence research at Google, has announced his departure from the company. Together with three other longtime engineers, he is founding Discovery Loop, a startup that aims to automate the process of scientific research using AI. Curiously, Google itself is among the investors in the new company.

Discovery Loop is a non-profit public benefit corporation that aims to automate the entire cycle of the scientific method, from forming a hypothesis through running an experiment to evaluating the result. The team is starting with the automation of machine learning research, with plans to eventually expand into hardware design, drug discovery and clean energy technology.

A Loop That Can Be Computerized

Dean explained the company's premise directly, arguing that in many fields of science the entire research process can be translated into algorithms and run at a scale no human team could match. Instead of individual experiments carried out over months, Discovery Loop is meant to run thousands of experiments simultaneously, iterating on results in real time.

In many fields you can fully computerize this entire process, and that will yield both a greater number and a higher quality of experiments, which will lead to breakthroughs and scientific progress - Jeff Dean, co-founder of Discovery Loop

Sanjay Ghemawat, one of the co-founders and a longtime Google senior fellow, noted that the new team needs infrastructure built for entirely different requirements than those governing consumer products like Search or Gemini. Dean, Ghemawat, Vinyals and Le were behind foundational Google technologies including MapReduce, Bigtable, Spanner, TensorFlow, sequence-to-sequence translation systems, and much of the work done at Google Brain.

Google Invests in Its Own Competition

Despite losing four key researchers, Google chose to become a founding investor in Discovery Loop and the new company's cloud partner. It's an unusual arrangement in which the company is funding a venture by its own former, longtime employees, apparently betting on access to the results of their work and on maintaining the business relationship through Google Cloud services.

Dean's departure fits into a broader trend of top AI researchers leaving Google. In June 2026, Noam Shazeer, co-author of the landmark paper "Attention Is All You Need," which in 2017 introduced the transformer architecture underlying most of today's large language models, moved to OpenAI. A few days later, John Jumper, the Nobel Prize in Chemistry laureate for AlphaFold, joined Anthropic. According to industry reports, all eight authors of the transformer paper have now left Google.

Cloud Grows While Research Loses People

Dean's departure coincided with very strong financial results at Google Cloud, whose revenue grew 82 percent year over year in the second quarter of 2026. Analysts point to mounting internal tension: researchers complain of limited access to computing power, as Google allocates a significant share of its TPU chips to cloud customers, including rival Anthropic, rather than to its own research teams.

That tension raises a question increasingly heard across the industry: does Google even need to build the best AI models, given that its cloud business is growing faster than Amazon's and Microsoft's offerings and delivering tangible profits without the risk of costly fundamental research. Some commentators argue that current models are already "good enough" for 90 percent of commercial use cases, which reduces the pressure for breakthrough discoveries and makes it easier for ambitious researchers to drift toward smaller, more agile projects.

What's Next for Discovery Loop

Discovery Loop has not yet disclosed a final valuation or a closing date for its funding round, but the scale of the amount being negotiated, close to $1 billion at a valuation of around $10 billion, puts the company among the best-capitalized AI startups at a very early stage. For Polish companies and investors tracking the AI market, it's a signal that capital is still flowing readily to teams with a proven track record, even when the product is still at the concept stage.

For Google itself, the exodus of the people behind key technologies, from transformers to distributed infrastructure, is a test of the company's credibility as a leader in AI research, regardless of how well its cloud segment performs. The coming months will show whether Discovery Loop actually accelerates scientific discovery or becomes another example of a project that promised more than it could deliver.