AlphaGo Co-Creator Leaves DeepMind to Build AI Using Structured Search Over LLM Scaling

Thore Graepel, a co-creator of AlphaGo, has left Google DeepMind to found a new startup that applies AlphaGo-style structured search and learned world models to real-world AI problems. His departure is part of a broader talent exodus from DeepMind, but his specific technical bet -- explicit search-based planning over pure token prediction -- signals a potential architectural shift for embodied AI and robotics.

What Happened: From Go Boards to Physical Worlds

Graepel told Sifted that "we need to go back to the architecture and fundamentally redesign it so that it does proper reasoning." His personal project page describes the new venture as bringing "AlphaGo-style reasoning to frontier AI, so that machines can plan and act under real-world uncertainty." He has not yet disclosed investors, team members, or a company name. A scheduled talk at LSE in November, billed as "From AlphaGo to AGI," will be his first major public engagement. TechTimes

Graepel was instrumental in the 2016 AlphaGo Nature paper and went on to contribute to AlphaGo Zero, AlphaZero, and MuZero. Between DeepMind stints, he led AI at Altos Labs (the cellular-rejuvenation biotech) and holds a part-time Chair at University College London. He also sits on the board of the Cooperative AI Foundation. TechTimes

Why the Architecture Choice Matters for Builders

AlphaGo did not work like GPT or Gemini. An LLM generates each token in a single forward pass -- it predicts, it does not search. AlphaGo combined a policy network (which proposes actions), a value network (which estimates success probability), and Monte Carlo Tree Search (MCTS), which runs thousands of simulated lookaheads at inference time to select the best action. This explicit search procedure is what Graepel is betting will generalize beyond board games. TechTimes

Current reasoning models like OpenAI o1 and Gemini Thinking approximate lookahead through extended chain-of-thought token generation, but they do so implicitly. Graepel's bet is that the explicit version -- principled structured search with a learned world model -- is what it will take for AI to operate reliably in physical environments where state spaces are continuous, dynamic, and unforgiving. MSN coverage

Practical Implications for Embodied AI

MuZero (2019) is the architectural ancestor most relevant here: it learned its own internal model of environment dynamics without needing the rules of the game. If that approach can be extended to robotics, a system could model its own physical environment, run structured lookahead over possible action sequences, and select the action its value function judges most likely to succeed. This is fundamentally different from the fine-tuned language models that currently power most robotics AI, which rely on imitation learning and struggle with novel physical configurations. TechTimes

The target applications are clear: robotics, drones, surgical instruments -- anywhere continuous state spaces and uncertain sensors dominate. A robot dropping a package, a surgical instrument approaching tissue: these problems do not have a token stream to continue. They require reasoning through state spaces. Lookonchain

The Hard Problem and Caveats

MCTS works beautifully in board games because each rollout is computationally cheap and the environment model (game rules) is exact. Real-world physical environments have continuous state spaces, imperfect sensors, and irreversible consequences. Building a learned world model accurate enough to support reliable multi-step lookahead without catastrophic degradation over long planning horizons is an unsolved research problem. MuZero demonstrated that a system can learn its own dynamics model, but it was tested in environments far simpler than the open-ended physical world. TechTimes

Graepel's departure is also part of a wider trend. DeepMind has lost an extraordinary concentration of founding talent: David Silver (Ineffable Intelligence, $1.1B seed), John Jumper (to Anthropic), Noam Shazeer (to OpenAI), and Jeff Dean (Discovery Loop). DeepMind's EMEA research talent share fell from 49% in 2022-23 to 18.6% in 2025-26, per Zeki Data. But DeepMind still retains massive resources and a strong research bench. The VC environment for AI startups is also extremely favorable, with $18.8 billion invested in AI startups founded since early 2025, per Dealroom. TechTimes

For builders, the key takeaway is that one of the architects of AlphaGo is now betting his career on the idea that explicit search with learned world models will beat implicit chain-of-thought reasoning for physical AI. If he succeeds, the next leap in AI capability may look more like Go than GPT.