Jeff Dean Leaves Google to Start Company; What's the Impact on Gemini?

On August 5, Jeff Dean and Oriol Vinyals simultaneously left Google to co-found Discovery Loop with Sanjay Ghemawat and Quoc Le. Google's announcement only mentioned Dean and Ghemawat, but the official website lists four founders. Dean describes himself on LinkedIn as 'Gemini overall co-technical lead,' and Vinyals' X bio reads 'Gemini co-lead.' On the same day, Koray Kavukcuoglu was promoted to Senior Vice President at Google DeepMind, reporting directly to Pichai, responsible for Gemini model development.

Earlier, on June 17, Noam Shazeer joined OpenAI as head of AI architecture research. Shazeer is a co-author of the Transformer paper and the proposer of sparse MoE and MQA. In August 2024, Google brought him back through the Character.AI deal for about $2.7 billion, but he left again in less than two years. Thus, all three Gemini technical co-leads departed within seven weeks.

The delay of Gemini 3.5 Pro is unrelated to these departures. On May 19 at I/O 2026, Pichai promised delivery 'next month,' but it was not released in June. On July 21, Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, but not 3.5 Pro, with official statements saying it is still being tested with partners. Bloomberg reported the delay was due to the model not meeting internal quality standards. All this happened while Dean and Vinyals were still at Google.

Gemini 4 pretraining has started, the training recipe is largely locked, and the execution-level researchers and engineers have not left in large numbers. Koray has been at DeepMind for 13 years, led teams that developed WaveNet and DQN, and previously served as DeepMind CTO and Google's Chief AI Architect. His appointment has not raised questions. Therefore, a difficult birth for Gemini 4 is unlikely.

The first layer is architectural judgment. Shazeer, Vinyals, and Le's representative works are structural innovations like Transformer, sparse MoE, MQA, and seq2seq. Such work cannot be measured by KPIs and relies on the intuition of the same group of people over more than a decade, making it hard to hand over. The second layer is the vertical chain. Google's complete self-developed stack from TPU to compiler to Pathways to models is an advantage over OpenAI and Anthropic. Hassabis called it 'the only company with a complete tech stack.' Dean and Ghemawat are among the few who can judge across layers.

The third layer is organizational gravity. Vinyals joined Google because of Dean's email, and now follows Dean out, which may prompt other senior researchers to reconsider their future. Discovery Loop's website describes the four founders as having 'full-stack depth across chips, hardware infrastructure, software infrastructure, ML models, and products,' which highly overlaps with Google's self-positioning.

Hassabis has become Chairman of Google DeepMind and Chief Scientist at Alphabet, shifting to long-term AGI strategy and Isomorphic drug development, with an office at Platform 37 in London. Koray takes over daily operations as SVP, reporting directly to Pichai, managing Gemini model development, frontier AI research, and Gemini application and developer teams. The change in rank means Google DeepMind has shifted from a semi-autonomous entity to a department with a shorter reporting line and closer to products.

This is not necessarily bad: Google's difficulty in the past year has been delivery pace. Pichai wants future Gemini releases to be 'almost monthly,' with Gemini 4 as the baseline. But the risk is that when the most influential research voices leave, the organization may over-optimize for 'on-time releases,' causing long-term directions without schedules to continue to be sidelined. Google's most valuable historical outputs—Transformer, TPU, MoE—were not scheduled outputs.

Discovery Loop aims to automate the scientific discovery loop, running thousands of loops in parallel, with the first step being the automation of machine learning research itself. Quoc Le told Wired that it might discover architectures different from Transformer. Dean told The New York Times that outside a public company there is space to make decisions not fully aligned with financial interests, implying that such space is lacking inside a public company.

Alphabet is a founding investor in Discovery Loop, and Google Cloud provides compute. The seed round was led by Radical Ventures and Khosla Ventures, with the amount undisclosed. This is a rational hedge: if successful, Google has a stake and priority access; if it fails, the loss is limited. But Vinyals pointed out the biggest flaw: how models can propose new ideas worth trying is exactly the weakness of current models.