AlphaGo builder's new London lab bets LLMs can't be trusted to reason

Thore Graepel, the DeepMind researcher who helped build AlphaGo, has launched a new London-based AI lab called Metis Reasoning. The company's founding claim is blunt: large language models cannot be trusted to reason. That is the bet behind the startup, which wants to take the techniques that defeated the world's best Go player and apply them to general reasoning.

Metis Reasoning laid out its thesis in a white paper titled "The Age of..." (the source cuts off there). The core argument is that LLMs, for all their fluency, are not reliable reasoning engines. Graepel's move is notable because he is not an outsider. He was part of the DeepMind team that built AlphaGo, the system that stunned the world in 2016 when it beat Go champion Lee Sedol. That victory was powered by a combination of deep neural networks and search algorithms that could evaluate future moves.

For context, AlphaGo's breakthrough was not just raw pattern matching. It used Monte Carlo tree search to look ahead and evaluate positions, guided by neural networks trained on human games and self-play. The lesson Graepel and his co-founders want to carry forward is that reasoning requires structured search, not just next-token prediction. LLMs generate text by predicting the next word, which is very different from exploring a tree of possible conclusions.

The timing is interesting. The AI industry is in a race to scale LLMs, with companies like OpenAI, Google, and Anthropic pushing larger models and longer context windows. But there is a growing chorus of researchers who argue that scale alone will not produce reliable reasoning. Metis Reasoning is entering that debate with a different technical bet: build systems that explicitly search over possible reasoning steps, much like AlphaGo searched over board positions.

For executives and investors, the significance is twofold. First, it signals that some of the most respected AI researchers believe the current paradigm has a ceiling. Second, it suggests a possible direction for the next wave of AI startups: not bigger models, but better reasoning architectures. If Metis succeeds, it could change how companies evaluate AI vendors, moving the conversation from benchmark scores to verifiable reasoning.

The London location is also worth noting. The city has become a serious AI hub, with DeepMind's heritage and a growing cluster of AI startups. Graepel's choice to launch there, rather than in Silicon Valley, reinforces that trend. It also gives the UK a stake in the next phase of AI development, one that may be less about raw compute and more about algorithmic insight.

There are open questions. The white paper's full title is not public, and it is unclear who the other founders are, how the company is funded, and what products it plans to ship. What is clear is the founding claim: LLMs cannot be trusted to reason. That is a strong statement from someone who helped build one of the most famous reasoning systems in history.

For decision-makers, the takeaway is to watch this space. If reasoning becomes the new frontier, the value may shift from companies that scale language models to companies that can make them think. That has implications for procurement, for AI strategy, and for the long-term roadmap of any organization relying on LLMs for critical decisions. Graepel's move is an early signal that the next big AI battle may be about reasoning, not just size.