Waking up one day, we learned that Demis Hassabis—Nobel laureate and the man who became legend through AlphaGo—had voluntarily handed over day‑to‑day operational control of DeepMind. At the same time, Google veteran Jeff Dean also left to start his own venture, taking three top technical experts with him.
The news had barely been announced when Alphabet's stock dropped 4% that same day. Taking over the entire operation and stepping into the role of Google's AI chief is a man whose name most ordinary observers can barely pronounce: Koray Kavukcuoglu. So who exactly is this new leader, far less famous than Hassabis?
A formidable figure who got into AI while accompanying his wife to study
To be fair, Koray Kavukcuoglu's entry into AI is arguably the most unlikely accident in the field. Koray is Turkish, 49 years old. Neither his bachelor's nor master's degree was in computer science; he studied aerospace engineering at Middle East Technical University (METU) in Turkey. He graduated with his bachelor's around 1999 and his master's around 2003.
What did he do after that? He went to work as a R&D engineer at Roketsan, a Turkish defense company that makes missiles. Yes, you read that right—he started out building missiles. So how did he end up in AI? According to his doctoral advisor, Yann LeCun—one of the "godfathers of deep learning"—the story goes like this: Koray's wife was going to Rutgers University in the US for her PhD, so he followed her and said, "Since I don't have much to do anyway, I might as well get a master's in computer science." And just like that, with a casual thought, the AI world lost a missile engineer and gained a future DeepMind leader.
He enrolled at New York University (NYU), took LeCun's course on machine learning and pattern recognition, and was hooked. In 2005, he officially became a computer science PhD student under LeCun and received his doctorate around 2010, working on sparse coding, hierarchical visual features, and other niche topics at the time. By the way, most of his classmates later became heavyweights—Raia Hadsell and Marc'Aurelio Ranzato, for example, now work with him at DeepMind, one as VP of research and the other as research director. Back then, deep learning was still an obscure field; those who stuck with it basically became the seeds of today's major AI labs.
After his PhD, he joined NEC Laboratories America. In 2012, a small, little‑known startup based in London—still years away from being acquired by Google—came calling: DeepMind. It is said that LeCun encouraged him to go. And thus, the missile engineer's AI career officially began.
From scientist to "fixer": DeepMind's true hidden operator takes the helm
In 2014, Google acquired DeepMind. Over the next decade, Koray climbed the ranks: research scientist → research director → VP of research → CTO. If you think he is just a "manager," you are underestimating him. His Google Scholar profile shows over 345,000 citations, with an h‑index of 90; citations since 2021 alone exceed 264,000, meaning his academic influence is still growing rapidly.
And any one of his technical achievements is textbook‑worthy: in 2015, he was a co‑core author of the Nature paper "Human‑level control through deep reinforcement learning"—Nature explicitly credited Kavukcuoglu, Mnih, and Silver as equal contributors. What is DQN? Simply put, it combines convolutional neural networks with Q‑learning, allowing the same algorithm to learn to play classic Atari games like Space Invaders, Pong, and Breakout just by looking at screen pixels. This is the ancestor of today's ubiquitous "AI agents"—even those agents that can send emails, check calendars, and operate software owe their lineage to this work.
WaveNet: a generative architecture that directly models raw audio waveforms, later deployed in Google Assistant, Maps navigation, voice search, and Cloud text‑to‑speech—the increasingly human‑like voice on your phone has his fingerprints on it. He was also a significant contributor to the early core research system of AlphaGo. Add to that Bayes by Backprop in Bayesian neural networks, and the seminal NLP work "Natural Language Processing (Almost) from Scratch"... So you see, he is not just an office‑bound manager, but a scientist who can still write papers.
Yet what truly elevated him to near‑legendary status inside Google is another role: the fixer. When Google's billionaire co‑founder Sergey Brin wants something done in the AI division—say, allocating more AI chips to a team, hiring a researcher, or moving someone to another department—whom does he message? The answer is obvious: Koray.
According to nine current and former employees who have worked closely with him, over the past two years Koray has become the indispensable firefighter for Google's AI models, specializing in resolving the friction among DeepMind's proud and temperamental top talent. How much authority does he have over people and budgets? By last year, he had 21 direct reports, and the entire organization numbered about 2,100 people. The final say on headcount and budget rests in his hands; any manager who wants to get things done has to come to him. Some call him an "empire builder."
The Information once reported an interesting detail: he can say "no" in a way that makes you feel like you've won. Several former colleagues recall leaving meetings with him feeling they had gotten everything they wanted, only to realize later—hey, I came away empty‑handed. That kind of rhetoric is worth learning for every employee.
"Enjoy it while you can"
To understand the scale of this upheaval, we need to rewind a few years. For a long time, DeepMind enjoyed a golden existence within Google: relatively independent, focused on research, without worrying about how to turn results into profitable products. LeCun watched this and would tease his former students at academic conferences, saying that pure research in industry would not last long; sooner or later they would have to serve product lines, even if indirectly. His exact words were something like: "Enjoy it while you can."
What followed is known to all. In November 2022, OpenAI unleashed ChatGPT, shaking the entire tech world—and Google was hit hardest because it struck at the very heart of Google Search. Even more painful: the Transformer underlying ChatGPT was precisely a research achievement that Google itself had developed but never turned into a product. Their own child, raised by someone else, came back to slap them in the face.
So in 2023, Google gritted its teeth and merged its two AI labs—Google Brain in California and DeepMind in London—into Google DeepMind, placing Hassabis at the helm. Then came Gemini, going head‑to‑head with GPT. Jeff Dean and Oriol Vinyals led the technical direction, but who was given the job of actually executing this grand vision and coordinating the fragmented teams? Again, Koray.
As we know, Gemini went from being a mocked "chaser" to climbing to the top of industry leaderboards. Last June, Google created a new role for him: Chief AI Architect, reporting directly to CEO Sundar Pichai, with one mission: to embed Gemini into every Google product. After taking office, he made sweeping changes, consolidating scattered model training efforts under himself; about 200 people from the search team alone (including the VP of search quality) were brought under his wing. He also moved from London to the Mountain View headquarters. Oh, and during that time Meta tried to poach him—but failed.
Now, Hassabis is officially stepping back, moving from CEO of Google DeepMind to Chairman of DeepMind while also serving as Chief Scientist at Alphabet. Day‑to‑day operational authority has been formally handed over. But he is not completely retired; he continues to run Isomorphic Labs (a DeepMind spin‑off focused on AI drug discovery). Taking over from the chess prodigy and Nobel laureate is his long‑time subordinate Koray, who now reports directly to Pichai. From firefighter to top leader, Koray has finally made it.
A researcher who cares about metrics and products
Compared with Hassabis, who often talks about AGI, scientific discovery, and the future of humanity, Koray speaks in a much more down‑to‑earth manner. He cares about how models perform on benchmarks, how pre‑training and post‑training work together, and whether new capabilities can become product experiences that users can directly touch.
When discussing Gemini, Koray has repeatedly emphasized that model capabilities can only generate effective feedback once they enter products and reach real users. In his view, the model release itself is only part of the job. The more important task is to turn multimodal understanding, coding, and agent capabilities into interactive components, learning tools, development environments, and enterprise services within search results.
A similar approach is reflected in his attitude toward AGI. He once said in an interview—and it sounds especially telling now: "One very, very important point for me is that we actually don't yet have the recipe for how to build AGI."
While Hassabis excels at defining long‑term goals and embodies DeepMind's tradition of basic research, Koray's background spans research, model training, organizational management, and product deployment, making him better suited to address Google's current pain point: unifying scattered research, compute, and product teams into a coherent system. The job arrangement itself also reveals Google's intent. Koray reporting directly to Pichai means the distance between Google DeepMind and Google's product divisions will continue to shrink. Gemini will also accelerate its evolution into a group‑level infrastructure covering search, office, cloud computing, Android, and developer tools.
After taking over, the real test begins
The Google DeepMind that Koray takes over has ample compute resources, a complete AI technology stack, and vast product entry points. But the problems before him are all tough nuts to crack. He needs to deliver the next‑generation Gemini on time, keep the model competitive, and rapidly deploy the technology across Google's product lines. At the same time, he must retain top researchers, reduce redundant construction and departmental friction within the huge organization, and prove that massive AI investments can ultimately generate visible commercial returns.
Making matters worse, early this morning, one of the most important employees in Google's history, former Chief Scientist Jeff Dean, also left to start a new company called Discovery Loop, focused on automating machine learning research—letting AI improve its own technology. And he didn't go alone; he took at least three heavyweight DeepMind figures with him: Sanjay Ghemawat, Quoc Le, and Oriol Vinyals.
Former Google Brain researcher Jeremy Nixon put it bluntly. He said Dean's departure is "almost unimaginable" and could be the first real crisis moment for this long‑stable company. If it were just Dean leaving, one could say "people have different aspirations." But the problem is, this is not the first. Earlier this year, Nobel laureate John Jumper jumped to Anthropic; less than two months ago, Noam Shazeer, a core author of the Transformer, also announced he was joining OpenAI.
So the aftershocks of this earthquake are far from over. Veterans are leaving one after another, and the new leader has barely settled into his seat. Koray's first major exam will be Gemini 4. Whether he can carry the burden left by Hassabis—let's wait and see.
