Google has spent the past two years trying to establish itself as a leader in generative artificial intelligence. After an early period in which it appeared to be caught off guard by Microsoft's aggressive push with OpenAI-powered Copilot, the company has regained ground with its Gemini models and DeepMind research. Yet just as it settles into that position, a significant shift in leadership and a string of departures threaten to complicate its trajectory.
Demis Hassabis, the co-founder and long-time CEO of DeepMind, announced he is stepping back from the day-to-day management of the lab. He will become chair of DeepMind and chief scientist at Alphabet, the parent company of Google. In his new role, Hassabis says he intends to focus on the big-picture challenges that come with the rapid advancement of AI. He will also remain in charge of Isomorphic Labs, the Alphabet subsidiary that uses AI for biomedical and scientific problem solving.
Hassabis helped found DeepMind in 2010, and the company was acquired by Google in 2014. Under his leadership, DeepMind produced a series of landmark research results, including the game-playing AlphaGo system, the protein-folding model AlphaFold, and the architecture that underpins many modern language models. He has also been one of the most prominent voices in the industry predicting the arrival of artificial general intelligence, or AGI, often described as AI capable of human-like understanding and reasoning across a wide range of tasks.
A new leader for DeepMind
The person inheriting daily control of DeepMind is Koray Kavukcuoglu, a veteran of the lab who has served in several research leadership positions. He will take the title of senior vice president of DeepMind, reporting directly to Google CEO Sundar Pichai. Kavukcuoglu has been deeply involved in the engineering and research efforts behind DeepMind's work for years, and his promotion signals continuity rather than a clean break with the past.
In a memo announcing the transition, Hassabis again leaned into the excitement and anxiety surrounding AGI. He said he has been working toward AGI his whole life and now feels it is close at hand. That framing is consistent with the language he has used in public interviews and keynote talks, where he has repeatedly said that advances in large language models are bringing artificial general intelligence closer than many people expect. His new position at Alphabet is designed to give him more room to think about the societal and technical consequences of that transition, while Kavukcuoglu handles the operational work of training and shipping new models.
Senior researchers leave to found Discovery Loop
While Hassabis is remaining inside the Alphabet ecosystem, four prominent Google researchers are leaving the company entirely. Quoc Le, Oriol Vinyals, Jeff Dean, and Sanjay Ghemawat are the co-founders of Discovery Loop, a new startup focused on using AI to automate scientific discovery. The company is structured as a public benefit corporation, meaning it is a for-profit entity intended to have a positive social impact. OpenAI and Anthropic use similar structures, although the practical meaning of that commitment has been debated.
According to reporting from Wired, Google CEO Sundar Pichai made multiple attempts to convince the four researchers to stay, particularly Dean and Ghemawat. He met with them several times before they ultimately decided to move forward with their plans. The fact that Pichai personally intervened highlights the importance of the group. Discovery Loop was able to raise substantial funding largely based on the reputations of its founders, a sign of how much weight these names carry in the AI community.
Jeff Dean is perhaps the most notable of the group. He was Google's 30th employee and has spent decades building the infrastructure that turned Google Search into a global information index. He is widely respected inside Google, to the point that engineers maintain an unofficial list of 'Jeff Dean Facts' on GitHub, filled with humorous and apocryphal claims about his abilities. More recently, Dean and Vinyals played a central role in designing Google's Gemini models, making their departure a significant blow to the company's AI roadmap.
Quoc Le is the co-inventor of several influential neural network techniques, including sequence-to-sequence learning, which became the foundation for many natural language processing systems. Oriol Vinyals, a former DeepMind researcher and Google scientist, was also a key contributor to the transformer architecture and other advances in machine learning. Sanjay Ghemawat is a distinguished engineer and systems researcher who has worked on Google's distributed computing and storage infrastructure. Together, the four bring a combination of systems engineering and deep learning expertise that is hard to replace.
A summer of high-profile exits
The departure of these four researchers is not an isolated event. Earlier in the summer, Google lost Noam Shazeer, a vice president of engineering and co-lead of Gemini. Shazeer was the lead author of the seminal 2017 paper 'Attention Is All You Need,' which introduced the transformer architecture that powers essentially every modern generative AI system. He has since joined OpenAI, a major coup for Google's chief rival. Shazeer had previously left Google in 2021 to found the chatbot startup Character.AI, but returned when Google licensed its technology and brought him back into the fold. That return now appears to have been temporary.
Around the same time, John Jumper, a DeepMind vice president who led the development of AlphaFold, announced he was leaving for Anthropic. AlphaFold is widely considered one of the most important applications of AI in science because it can accurately predict protein structures, accelerating research in biology and medicine. Jumper was quickly followed by senior DeepMind researchers Jonas Adler and Alexander Pritzel, who were reportedly instrumental in the early development of Gemini. Their move to Anthropic also represented a direct transfer of knowledge from Google to its competitor.
These departures have been accompanied by growing concern that Google's recent AI progress has slowed. The company spent much of last year closing the gap with OpenAI after initially being caught flat-footed by ChatGPT and Microsoft's Copilot. At its I/O developer conference this spring, Google said it planned to release Gemini 3.5 Pro in June. Months have passed, however, and the company has only shipped lower-power Flash models, while the more advanced Gemini 3.5 Pro remains absent. OpenAI and Anthropic, meanwhile, have both released new frontier models with improved reasoning and coding abilities, keeping the pressure on Google.
Pressure on Google's AI ambitions
Google CEO Sundar Pichai attempted to reassure employees and observers with a statement emphasizing the company's strengths. He said Google has amazing talent, world-class compute, and products that bring AI to more people than any other company. There is truth to that: Google still operates one of the largest cloud infrastructures in the world, has deep integration of AI into Search, Android, and Workspace, and continues to make substantial research contributions. But the talent drain is a real problem, and the loss of researchers who were directly involved in Gemini's design raises questions about the next generation of models.
DeepMind remains one of the most respected AI research laboratories in the world. The lab has a long tradition of academic publishing and produced some of the most fundamental advances in deep learning. It is also larger and more well funded than most startups, giving it the resources to attract new researchers and continue pushing forward. However, the culture at DeepMind has changed as it has become more product-focused, and some researchers have expressed discomfort with the military and government contracts that Google has pursued in recent years. The increased closeness between Google and the Trump administration has reportedly been another point of friction, and Jeff Dean in particular was said to be among those who pushed back against that closeness.
The broader AI landscape is also more competitive than ever. OpenAI has moved from research lab to product powerhouse, with ChatGPT becoming a household name and its models being integrated into everything from Microsoft Windows to enterprise software. Anthropic has positioned itself as a safety-focused alternative, attracting top talent from DeepMind and elsewhere. Meta has released its Llama models as open weights, and startups like X.ai and Mistral are also competing for attention. In that environment, retaining the people who built Google's Gemini models is essential to maintaining a competitive edge.
Discovery Loop's public benefit corporation structure is interesting because it suggests the four founders are thinking about scientific discovery rather than consumer AI products. They join a crowded field of AI-for-science initiatives. DeepMind's AlphaFold already demonstrated the potential of AI in biology, and other research groups are exploring ways to use large language models for drug discovery, materials science, and automated laboratory experiments. The startup's founders believe their combined expertise can accelerate this trend significantly.
Hassabis's move to Alphabet also reflects a broader shift in the industry as the most senior AI figures move into higher levels of strategic oversight. Many of the defining technical questions are no longer purely about model architecture; they are about compute, data, regulation, and governance. As AGI becomes a more plausible prospect, the people who invented these technologies are increasingly spending their time on safety, alignment, and the societal implications of the systems they helped build. Google will still benefit from Hassabis's advice and research vision, even if it loses some of its most hands-on builders.
The coming months will show whether Google can maintain its position despite the departures. The company still has a deep bench of researchers and engineers who were not part of the recent exodus, and Kavukcuoglu is an experienced leader who knows the inner workings of DeepMind. But the loss of Jeff Dean, Oriol Vinyals, Quoc Le, and Sanjay Ghemawat, plus the earlier exits of Noam Shazeer and John Jumper, means that the people responsible for some of Google's most important AI breakthroughs are now working elsewhere. With Gemini 3.5 Pro delayed and competitors shipping increasingly capable models, the pressure on Google's AI team has never been greater.
Source: Ars Technica News