A few days ago, Demis Hassabis stepped down as CEO of Google DeepMind. He didn’t leave the company. He moved up — into a new role as Chairman of Google DeepMind and Chief Scientist of Alphabet, handing day-to-day operations to longtime CTO Koray Kavukcuoglu. On the surface, this reads like a routine corporate reshuffle. Underneath it, there’s a much older document that explains exactly why this moment was always coming.
Six years earlier, in March 2019, Hassabis stood in front of an MIT audience and walked through a 55-slide presentation on self-learning systems. He wasn’t unveiling ChatGPT-era hype — that era hadn’t started yet. He was laying out DeepMind’s entire theory of intelligence, including five specific open research problems he believed had to be solved for real progress toward general-purpose AI. Revisit that deck today, and it reads less like an old talk and more like a roadmap — one that maps almost perfectly onto both the AI industry’s last six years and Hassabis’s own career move this month.
This is the story of what that 2019 deck actually said, why it matters now, and what it tells us about why Hassabis is stepping back from day-to-day DeepMind operations at exactly this moment.
Setting the Scene: MIT, March 2019
The talk took place on March 20, 2019, co-hosted by MIT’s Center for Brains, Minds, and Machines and MIT Quest for Intelligence, in front of an academic audience already deeply engaged with AI research. This was well before DeepMind’s merger with Google Brain. Well before AlphaFold 2 solved the protein-folding problem. Well before large language models became a mainstream household topic. Most of the room had no real sense of what was coming next.
Hassabis’s own path to that stage wasn’t a conventional one. A former child chess prodigy, he coded the bestselling simulation game Theme Park at age 17, graduated from Cambridge with a Double First in Computer Science, built and sold a videogame studio, then returned to academia for a PhD in cognitive neuroscience at UCL — research on memory and imagination that made Science magazine’s Top Ten Scientific Breakthroughs of 2007. He founded DeepMind in 2010, and Google acquired it in 2014 in what was then its largest European acquisition. By 2019, DeepMind had already produced AlphaGo, the first AI system to beat a human world champion at Go. For the full story of that journey — including the 2024 Nobel Prize that followed — see our complete Demis Hassabis biography. The MIT talk was Hassabis explaining, in his own words, why that win was never really the point.
The Core Argument: Two Paths to Intelligence
The deck’s central thesis was simple, and Hassabis framed it as a genuine fork in the road for the entire AI field. One path relies on expert systems — hardcoded knowledge, carefully engineered rules, systems that work well within their designed boundaries but break the moment they hit something unexpected. The other path relies on learning systems — architectures that learn from first principles, generalize what they’ve learned to new, unseen tasks, and keep improving as they encounter more data and experience.
DeepMind bet everything on the second path. That bet wasn’t obvious or safe in 2019. Expert systems had decades of institutional trust behind them, particularly in industry applications where predictability mattered more than raw capability. Learning systems were promising but unproven at the scale that would eventually matter. Hassabis’s MIT talk was, in effect, a public argument for why DeepMind was staking its entire research identity on the harder, less certain path — and everything that followed from AlphaFold to Gemini has served as the payoff for that bet.
The Five Open Problems — And What Happened to Each
The part of the 2019 deck that holds up best isn’t the philosophy. It’s the specifics. Hassabis identified five concrete, unsolved research problems standing between the AI of 2019 and genuine general-purpose intelligence: unsupervised learning, memory, transfer learning, imagination-based planning, and language understanding.
Unsupervised learning — the ability for a system to learn useful structure from raw, unlabeled data rather than relying on expensively hand-labeled examples — went on to become the literal foundation of the modern large language model era. The core training approach behind GPT-style models is built explicitly around learning from massive amounts of unlabeled text, treating next-word prediction itself as the unsupervised signal.
Memory — giving AI systems the ability to retain and use information over extended interactions, rather than treating every input as an isolated event — shows up today in everything from the expanded context windows of modern models to dedicated memory systems that let AI assistants recall earlier conversations and maintain continuity over time.
Transfer learning — the capacity to take knowledge learned in one domain and apply it usefully to a completely different one — is now standard practice across the field. Modern foundation models are trained broadly and then adapted, or “fine-tuned,” to specific downstream tasks, a workflow that would have sounded ambitious in 2019 and is now simply how the industry operates.
Imagination-based planning — systems that can simulate future outcomes internally before acting, rather than learning purely through costly trial and error — connects directly to DeepMind’s own later research into model-based reinforcement learning and planning, and to the broader industry shift toward agentic AI systems that reason through multi-step plans before executing them.
Language understanding — arguably the most visibly “solved” of the five from a public perspective — is the problem that produced the entire conversational AI wave the public now interacts with daily, from Claude to ChatGPT to Gemini itself, and that same underlying progress now extends well beyond text into tools like the AI video generators reshaping how content gets made.
Six years is not a long time in research terms, and Hassabis certainly wasn’t the only researcher pointing at these problems. But naming all five, together, as the connected agenda for reaching general intelligence — and then spending the next six years building an organization that systematically attacked each one — is a different kind of achievement than getting lucky on a single prediction.
What This Reveals About the 2026 Transition
Read against that backdrop, Hassabis’s move from CEO to Chairman and Chief Scientist looks less like a retirement lap and more like a return to the work the 2019 deck was actually about.
In his own public statement announcing the change, Hassabis said he’s been working toward AGI his entire life, and that this moment — with AGI feeling closer at hand than ever — is exactly why he wants to hand over day-to-day operational responsibility. That framing matters. Running Google DeepMind’s daily operations means managing product timelines, org structure, and the extraordinary scale that Gemini and Google’s broader AI push now demand. It’s a fundamentally different job from the one described in a 55-slide research talk about unsolved problems in intelligence.
Koray Kavukcuoglu, who has been DeepMind’s CTO, now steps up as Senior Vice President to run daily operations — a continuity choice rather than an outside hire, suggesting Alphabet wanted operational stability rather than a strategic pivot. Hassabis, meanwhile, keeps his hands directly on two things: Alphabet’s broader scientific direction as Chief Scientist, and Isomorphic Labs, the AI drug-discovery company he founded in 2021, which sits closer to the “science and humanity’s grandest challenges” framing that has defined his public messaging since long before this reshuffle.
The timing lines up with another data point worth noting: Hassabis won the 2024 Nobel Prize in Chemistry for AlphaFold, the system that predicted the structure of over 200 million proteins and solved a genuine 50-year grand challenge in biology. That’s not a language-model achievement — it’s a direct, real-world payoff of the “imagination-based planning” and “transfer learning” threads from the 2019 deck, applied to biology instead of games or text. If anything, the Nobel likely reinforced exactly the direction Hassabis is now formally stepping toward: less daily management, more direct scientific leadership on problems with that kind of stakes.
The Wider Shakeup Around Him
This transition didn’t happen in isolation. The same reorganization saw Jeff Dean — a 27-year Google veteran and, until this shakeup, Google DeepMind’s chief scientist — depart the company entirely, alongside several colleagues, to launch a new public-benefit corporation called Discovery Loop. Alphabet’s stock reportedly dropped around 4% on the news, reflecting how seriously markets took a change this senior at the company driving Google’s AI strategy.
Internally, though, multiple reports describe the change as landing with far less shock than the outside market reaction suggested — closer to a formality that had been anticipated than a sudden strategic reversal. Google’s own framing, through CEO Sundar Pichai, positioned the move as an acceleration play: freeing Hassabis to focus on frontier AGI research and scientific breakthroughs while Kavukcuoglu drives the operational execution needed to keep pace with competitors.
Why the 2019 Deck Is the Right Lens for This Moment
It would be easy to read Hassabis’s transition as just another AI-industry leadership story — a founder stepping back as a company scales, a familiar pattern across tech. What the 2019 MIT deck adds is evidence that this specific move isn’t reactive. It’s the natural endpoint of a research agenda Hassabis described in explicit, public detail six years before it needed to happen.
He named the problems before most of the field had organized around them. He built DeepMind’s strategy around solving them systematically rather than chasing whichever benchmark was trending that year. And now, with several of those five problems substantially advanced — some largely solved, some still very much open — he’s restructuring his own role to spend more time on the parts of the mission that never fit neatly into a CEO’s calendar: direct scientific work on AGI and on using AI to solve concrete, high-stakes problems like disease.
The deck was never really a pitch in the traditional sense. It was closer to a public research thesis. Six years later, it’s also turned out to be a fairly accurate description of the career move Hassabis just made.
Frequently Asked Questions
What did Demis Hassabis present at MIT in 2019?
On March 20, 2019, Hassabis gave a talk on self-learning systems at MIT, co-hosted by the Center for Brains, Minds, and Machines and MIT Quest for Intelligence. The 55-slide presentation argued that learning-based AI systems would outperform hardcoded expert systems, and identified five open research problems — unsupervised learning, memory, transfer learning, imagination-based planning, and language understanding — as the path toward general-purpose AI.
What is Demis Hassabis’s new role at Google?
Hassabis stepped down as CEO of Google DeepMind and now serves as Chairman of Google DeepMind and Chief Scientist of Alphabet. Koray Kavukcuoglu, formerly DeepMind’s CTO, has taken over daily operations as Senior Vice President. Hassabis continues to lead Isomorphic Labs, the AI drug-discovery company he founded in 2021.
Did Hassabis’s 2019 predictions actually come true?
Largely, yes. The five research problems he identified in 2019 map closely onto the core technical foundations of today’s AI systems — unsupervised learning underpins large language model training, memory now appears in extended context windows and dedicated recall systems, transfer learning is standard practice through fine-tuning, imagination-based planning connects to modern AI agents and reasoning systems, and language understanding produced the conversational AI systems now used daily by hundreds of millions of people.
Who else left Google DeepMind in this leadership change?
Jeff Dean, Google DeepMind’s chief scientist and a 27-year Google veteran, departed the company as part of the same reorganization, along with several colleagues, to launch a new public-benefit corporation called Discovery Loop.
Why did Alphabet’s stock drop after this announcement?
Alphabet shares reportedly fell around 4% following the news, reflecting market sensitivity to a change this significant at the top of the company’s AI division, even though internal accounts suggest the shift was viewed within the company as an anticipated, orderly transition rather than a sudden strategic shock.
Note: Details in this article reflect publicly available information as of mid-2026, based on Demis Hassabis’s public statements, Google’s official announcements, and contemporaneous reporting. As with any recent corporate transition, some specifics may be clarified or updated as more information becomes available.
Related reads:
