Sergey Brin is back at Google, working on Gemini

Sergey Brin holds no executive title at Google. He has nonetheless become increasingly influential over Gemini, the company’s flagship AI model, Business Insider reported, citing eight current and former employees.

The Google cofounder works from a converted microkitchen in Gradient Canopy, a building at the company’s Mountain View headquarters. Some of the people said recent leadership changes could give him even more sway.

TNW has not independently verified the report, and Google declined to comment to Business Insider.

The report follows a reshuffle in August. Demis Hassabis gave up the title of Google DeepMind chief executive and became Alphabet’s chief scientist. Koray Kavukcuoglu, who moved from London to the US last year, became senior vice president of Google DeepMind. Jeff Dean, Google’s longtime chief scientist, left after 27 years to start his own company.

The announcements did not mention Brin’s role.

How the kitchen works

According to Business Insider, Kavukcuoglu now sits alongside Brin in the microkitchen, at a U-shaped arrangement of desks. Chief executive Sundar Pichai visits several times a week. A team working on Gemini’s coding abilities, led by research engineer Sebastian Borgeaud, has desks there too.

One former employee told the outlet that Brin wants to run Gemini like a startup. Another said the promise of the kitchen was that it could help cut through the politics.

Access to computing power is one reason his influence matters. Getting Google’s chips can be difficult, even for the Gemini team. The official route is a formal request that needs sign-off from senior managers. Two people told Business Insider that teams also trade work with other groups in exchange for their chips.

The less official route, they said, is to go to Brin.

The calls he has made

Brin has involved himself mainly in the technical side of Gemini, according to the report. That includes model sizes, release windows and paths to artificial general intelligence. In 2025 he cut a chip project led by Dean, known internally as Frozen, which would have built a version of the Gemini architecture into the silicon.

Google later revived it as Frozen v2, then stripped away some of its resources this year, one person said. Dean declined to comment to Business Insider.

Earlier this year, Brin pushed an idea to monitor some Google employees as they coded, using the data to train Gemini’s coding skills. He has also pushed staff to focus on recursive self-improvement, the point at which AI systems can improve themselves. Current and former employees described him to Business Insider as very invested in the idea.

In 2025 he told staff in an internal post that about 60 hours of work a week was the sweet spot for productivity. The New York Times first reported that message. Reuters had previously reported on some of his remarks on the subject.

Brin has spoken about the role in public. At a question-and-answer session at AGI House in San Francisco, he said he would often poke and prod teams, and called himself “a little bit of a rabble-rouser”.

Business Insider said Brin has spent less time in the kitchen in recent months, sometimes appearing on alternate weeks.

Last December he moved some personal assets out of California. That came ahead of a proposed one-time billionaires’ tax, which goes to a ballot in November. Business Insider said Brin could be liable for $14bn if he were considered a resident.

Where Gemini stands

The report says Google has fallen behind the frontier, which puts pressure on its next model, Gemini 4. The company has lost researchers including John Jumper and Noam Shazeer in recent months. It has also rehired Barret Zoph, a cofounder of Thinking Machines Lab.

Google says its latest Gemini model shows significant improvements in coding. Gil Luria, head of technology research at the investment bank D.A. Davidson, told Business Insider that Google needed a big change that would cut through its bureaucracy.

DeepMind’s science work continues alongside the product race. This week it released AlphaGenome Atlas, a 1-petabyte database of precomputed predictions for all 9 billion possible single-letter changes in the human genome. According to DeepMind’s blog post, the database lets researchers rank which variants are most likely to affect a trait or disease.

In one test with the GREGoR Consortium, researchers used it to find a variant in DNM1, a gene linked to a rare and severe form of epilepsy.

The database is free for researchers. Commercial users, including the drug discovery company Isomorphic Labs, will pay a licence fee, Pushmeet Kohli, DeepMind’s vice president of research, told Bloomberg.

Kohli said the company believed fundamental advances would have an impact on the commercial side. Alessia David, of Imperial College London, told Bloomberg the tool was not yet at the level of AlphaFold, but would speed up research.

Alumni take the methods elsewhere

Some former DeepMind researchers are starting their own companies. Federico Felici and Jonas Buchli founded Fusionality, a Lausanne startup that builds control systems for fusion reactors, this year. It raised a $3.7m (CHF 3m) pre-seed round from Founderful and Playfair, TechCrunch reported.

The seven-person company is starting with magnetic confinement, the approach that uses powerful electromagnets to hold fusion fuel in place. Felici said that about 80 percent of each fusion company’s control system is the same. The two met while teaching AI to control an experimental tokamak at EPFL in Switzerland.

Felici, the chief executive, told TechCrunch that AI would complement parts of a reactor’s control system rather than run the whole thing.