Jeff Dean, one of the most influential engineers in Google’s history, is leaving the company after nearly 27 years to launch an independent artificial intelligence startup called Discovery Loop. He will be joined by senior researchers Sanjay Ghemawat, Oriol Vinyals and Quoc Le, taking decades of engineering and AI experience outside Google at a critical moment in the industry.

The departures coincide with a wider leadership reorganisation. Demis Hassabis is stepping away from the daily management of Google DeepMind but is not leaving Alphabet. He will become chairman of DeepMind and chief scientist of Alphabet while continuing to lead its AI drug-discovery company, Isomorphic Labs.

DeepMind chief technology officer Koray Kavukcuoglu will assume operational responsibility as a senior vice-president reporting directly to Google CEO Sundar Pichai. His priorities will include frontier AI research and delivering Google’s Gemini roadmap as competition with OpenAI and Anthropic intensifies.

Why Jeff Dean’s exit matters

Dean joined Google in 1999 as one of its earliest employees, when the company reportedly had about 25 workers. He helped develop the large-scale computing infrastructure behind Google Search and later co-founded Google Brain, which advanced deep learning before being combined with DeepMind.

His work has covered neural networks, machine translation, speech recognition and specialised Tensor Processing Units used for demanding AI workloads. Google’s official profile of Jeff Dean credits him with driving innovations supporting its infrastructure and global AI development.

The three researchers leaving with him add considerable experience. Ghemawat collaborated with Dean on foundational Google systems. Vinyals served as a DeepMind vice-president and Gemini co-lead, while Le co-founded Google Brain and contributed to major advances in machine learning.

The move is therefore more than a routine executive departure. Google is losing a closely connected group whose work stretches from its early search infrastructure to its latest Gemini models.

What Discovery Loop plans to build

Discovery Loop will operate as a public benefit corporation. Google will invest in the company and provide cloud services, allowing the startup to remain independent while retaining access to the computing infrastructure required for advanced AI research.

The company wants to automate the experimental loop used in science: develop a hypothesis, design an experiment, analyse the result and decide what should be tested next. Its systems could eventually generate ideas, perform digital experiments, evaluate their findings and plan improved tests with limited human intervention.

The founders are particularly interested in recursive self-improvement, where AI helps develop and test better AI systems. Discovery Loop is expected to begin with machine-learning research before considering applications in computer hardware, drug discovery, biology, new materials and clean energy.

These remain ambitions rather than proven products. The company must still demonstrate that automated experimentation can produce reliable scientific discoveries outside controlled settings.

Hassabis changes focus as Kavukcuoglu takes control

Dean is leaving Google, but Hassabis is moving into a broader strategic role within Alphabet. As chairman of DeepMind and Alphabet’s chief scientist, he will focus on artificial general intelligence, scientific research and the consequences of increasingly capable AI systems.

Hassabis said he has worked towards AGI throughout his life and believes it may be close. His scientific background includes the AlphaFold protein-structure project, work that helped Hassabis and John Jumper receive the 2024 Nobel Prize in Chemistry.

Kavukcuoglu, meanwhile, becomes the key operational leader inside DeepMind. As its former technology chief and Google’s chief AI architect, he has experience connecting research with products. He has said his main priority is creating a clear path for the Gemini roadmap while maintaining DeepMind’s responsible-AI mission.

Gemini now extends beyond a chatbot into Search, Workspace, Android, Cloud and developer products. Google’s expansion of Gemini agents, TPUs and enterprise AI tools shows how heavily its product strategy depends on DeepMind’s models and infrastructure.

Gemini delays add pressure to the reshuffle

The changes follow employee pushback and several high-profile departures from Google’s AI teams. Axios reported that Gemini 3.5 Pro was months behind schedule, with some company sources partly attributing the delay to low morale. Its final release timing could still change while development continues.

Google must now retain experienced researchers, complete its model roadmap and turn expensive AI development into dependable products while competitors release new systems. Alphabet nevertheless retains major advantages through its AI chips, cloud platform and ability to distribute Gemini across products used by billions of people.

Those resources help explain why Alphabet remains a formidable competitor despite the leadership disruption, an issue explored in an earlier assessment of Alphabet’s position in the 2026 AI race.

Tomorrow will be my last day at Google after 27 years, and watching it grow from 25 people to 190,000+ has been an amazing journey. Below is a note I shared with many people internally at Google today. An excerpt is:

It has been an absolute pleasure to work with you and to help… pic.twitter.com/5sHj1aDiH7

— Jeff Dean (@JeffDean) August 5, 2026

Why Google is investing in the departing team

Google’s backing of Discovery Loop softens the separation. Providing cloud infrastructure may keep the startup’s computing demand within Google’s ecosystem, while an investment gives Alphabet exposure to any valuable research or technology the company develops.

The relationship also gives Dean and his colleagues greater freedom without completely severing their connection to Google. They can pursue scientific projects that may not offer an immediate commercial return while still collaborating on selected research and infrastructure.

Users are unlikely to see an immediate change in Gemini or Google Search. The real consequences will emerge through the quality and speed of future model releases, DeepMind’s ability to retain talent and Discovery Loop’s success—or failure—in turning automated experimentation into practical scientific advances.