Jeff Dean, one of Google's earliest employees and an architect of its artificial-intelligence strategy for the past 15 years, is leaving the company to launch an AI startup focused on scientific discovery.
Dean, who joined Google as its 30th employee in 1999, will be chief executive at a new public-benefit corporation called Discovery Loop that aims to accelerate technology by automating complicated, multistep science and engineering tasks.
The startup will initially focus on automating the process of machine learning research and engineering via its own internal technology stack. Eventually, it plans to apply its technology to a variety of areas, including hardware design, drug discovery and global challenges such as clean energy, according to a news release.
Joining Dean in the endeavor are fellow Google veterans Oriol Vinyals, Quoc Le and Sanjay Ghemawat, who between them account for high-profile advances in mathematics and protein structure and are some of the most cited researchers in the world. Their departures, along with Dean's, add to a string of recent talent losses for the company.
"The four of us have been chatting about this a little bit and we got excited," Dean said in an interview at his Palo Alto, Calif., home Tuesday. "We think by automating many of the classical experimental-loop aspects with machine learning models, augmenting human scientists and engineers, we'll be able to make it possible to explore more space in broad science areas to accelerate discovery."
The scientific process, he said, is about coming up with an idea for an experiment, setting up and running the experiment and then evaluating the results.
"Particularly in a lot of domains, you can fully computerize that whole loop," he said -- the inspiration for the startup's name.
Alphabet is providing cloud and computing capacity to Discovery Loop and is also an investor in the company, the news release said. Radical Ventures and Khosla Ventures are co-leading its seed round. The round isn't closed, and the startup declined to share its valuation.
Dean said he began seriously contemplating the startup five weeks ago and reached out to Vinyals, a longtime colleague and main contributor to AlphaFold, a famous AI program that predicts protein structures and won the Nobel Prize in 2024. They then connected with Le and Ghemawat and started discussing building the infrastructure to allow more kinds of science to be automated and what would be needed to do that.
The four have all known each other for more than a decade. Le co-wrote and advised Google's AI systems that won gold-medal status at the International Mathematical Olympiad and have solved Olympiad-level geometry problems.
Ghemawat was also an early Google employee. Alongside Dean, he helped create Google's distributed systems in the early 2000s and in 2012 he and Dean were recognized by the Association for Computing Machinery for their efforts.
The scientists decided they needed to leave Google to pursue their new mission.
"We wanted to build something differently than how things are built at Google right now," Ghemawat said, explaining that Google's infrastructure works well for big consumer apps, large ad systems and search. But, he said, it has "very different requirements than the type of infrastructure we want to build for research."
Google Chief Executive Sundar Pichai said Dean and Ghemawat helped drive some of the company's most significant technology transitions over their 27-year tenure and that he wished them the best.
Dean co-founded Google Brain, the company's pioneering AI deep-learning research team, in 2011. Le was a founding member of Google Brain and Vinyals joined it two years later. In 2023, Hassabis's Google DeepMind, parts of Google Brain, and the broader Google research team merged into one organization.
Dean also spearheaded development of the company's tensor processing units, the chips that power its large language models and other technology bets and are now being sold to other firms.
In recent months, Google has lost other AI top talent. Noam Shazeer, the Character.AI co-founder and co-author of the famous 2017 transformer paper that led to the current AI revolution, recently left for OpenAI. And John Jumper, who shared in the AlphaFold Nobel Prize, left to join Anthropic.
After vaulting ahead in the race to develop powerful frontier AI models with an updated version of Gemini last fall, Google has since fallen behind OpenAI and Anthropic. In some cases, new models from Meta are outperforming Gemini on certain benchmarks.
Google still hasn't released its latest frontier model, Gemini 3.5 Pro, and investors are getting impatient. The company's stock initially dropped more than 7% two weeks ago after it said it would boost spending on AI and didn't give an update on when the new model would be launched. Its shares have since rebounded and are up 17%.
News Corp, owner of The Wall Street Journal, has a commercial agreement to supply content on Google platforms.
Comments