Thore Graepel’s bet — the physicist-turned-machine-learning pioneer whose name appears on the 2016 Nature paper that introduced AlphaGo to the world, has left Google DeepMind to launch a startup built around structured search and planning — the same class of techniques he helped pioneer a decade ago. His departure, first reported Sunday by Sifted, adds another senior researcher to what is shaping up to be the most concentrated single-lab talent exodus the AI industry has ever recorded, and his technical thesis makes it one of the more pointed ones. “We need to go back to the architecture and fundamentally redesign it so that it does proper reasoning,” Graepel told Sifted.
Graepel’s career has tracked the arc of probabilistic machine learning from research curiosities into infrastructure that billions of people touch. At Microsoft Research Cambridge, where he worked from 2003 to 2015, he co-created TrueSkill, the Bayesian player-ranking system that has matched Xbox Live gamers since 2005, and AdPredictor, the click-through prediction model behind Bing’s advertising business. Both systems leaned on factor graphs and message-passing inference — methods for turning messy, uncertain inputs into calibrated probabilistic outputs at industrial scale.
He joined DeepMind in 2015 and was part of the four-name team — alongside David Silver, Aja Huang, and Demis Hassabis — that published the AlphaGo paper in January 2016. A month later, AlphaGo defeated 18-time world champion Lee Sedol 4–1 in Seoul, in a match broadcast to millions. Graepel went on to contribute to AlphaGo Zero in 2017, AlphaZero that same year, and MuZero in 2019. Between two DeepMind stints, he served as SVP and Global Head of AI at Altos Labs, the Bezos-backed cellular-rejuvenation biotech, before returning to Google DeepMind in 2025 as a Distinguished Research Scientist working on what the lab called post-AGI futures. He has held a concurrent appointment as Chair of Machine Learning at University College London and sits on the boards of the Partnership on AI and the Cooperative AI Foundation.
Why Graepel Thinks LLMs Cannot Get There Alone
Graepel’s bet is that the industry’s current frontier model paradigm — bigger transformers trained on more tokens — has a ceiling, and that ceiling is reasoning. Large language models generate each token in a single forward pass. They predict; they do not search. AlphaGo worked differently. It combined a policy network that proposed candidate moves, a value network that estimated each position’s win probability, and an explicit Monte Carlo tree search procedure that looked ahead before committing to a move. The architecture is the intellectual ancestor of nearly every “reasoning” release the industry has shipped in the last 18 months, even when the labs marketing those releases do not always say so.
“We need to go back to the architecture and fundamentally redesign it so that it does proper reasoning,” Graepel said — a statement that lands as a quiet critique of the scaling-does-everything orthodoxy that has dominated frontier-lab roadmaps since GPT-3. His personal project website now frames the new venture as bringing “AlphaGo-style reasoning to frontier AI, so that machines can plan and act under real-world uncertainty.” The implication is that LLM token prediction, even at trillion-parameter scale, treats each next-step decision as a single forward pass with no lookahead — workable for chat, brittle for any domain where the cost of a wrong move compounds.
Another Senior Exit From DeepMind
Graepel’s move is the latest in a string of senior departures from the London-based lab. Co-founder Mustafa Suleyman left in 2022 to co-found Inflection AI, then moved to Microsoft to lead its AI products division. Co-founder Shane Legg has remained, but a steady drip of principal researchers and team leads has followed them out the door — most recently to Anthropic, where former DeepMind staff now hold senior roles, and to a cluster of well-funded reasoning-focused startups. The pattern has been visible enough that DeepMind publicly restructured its research divisions in mid-2025 to try to slow it.
What makes Graepel’s exit distinctive is the explicit technical argument attached to it. Most departures have been framed in personal or commercial terms — Suleyman’s clashes with product priorities, for example, or compensation packages at competing labs. Graepel is instead arguing that the dominant architecture is the problem, and that structured search is the missing primitive. That is a harder claim to dismiss as poaching, and a more uncomfortable one for DeepMind’s own research roadmap, much of which still runs on transformer scaling. If his startup gains traction, it will be read as evidence that the frontier-lab consensus on how to reach AGI has a real internal dissent.
What the Startup Is Building
Graepel has not yet disclosed a company name, funding round, or launch date. His website describes the mission in broad terms: porting AlphaGo-style search, planning, and probabilistic reasoning into frontier-scale systems that can act under uncertainty in the physical world, not just on a 19-by-19 board. The bet is that structured search and planning — augmented by modern learning — can carry the field past the reasoning ceiling that pure scaling has hit, and that the team that builds that primitive first will define the next era of AI research.