Thore Graepel, one of the engineers behind DeepMind’s AlphaGo, has launched a new artificial intelligence company in London called Metis Reasoning. The startup’s founding argument is blunt: large language models cannot be trusted to reason. Its founders want to take the search-and-evaluation lessons that made AlphaGo famous and apply them to general-purpose reasoning systems.
Metis set out its case in a white paper titled “The Age of Reasoning.” Graepel left Google DeepMind this summer. He founded Metis with Piers Ingram, a technology and biotech entrepreneur who has raised more than $350 million; Michael Warren, a robotics expert who worked at Boeing; and David Berman, head of AI at Cambridge Consultants, who also advises the UK government and NATO on AI. The team is listed as 12 people, and the white paper names offices in London, Singapore and Los Angeles.
Key facts at a glance
- Thore Graepel, a co-creator of AlphaGo, has founded Metis Reasoning in London.
- The company argues that large language models cannot be trusted to reason.
- Metis wants to apply AlphaGo-style search and evaluation to general reasoning.
- Its white paper, “The Age of Reasoning,” identifies hallucination, unauditable reasoning and inaccessible internal knowledge as core LLM problems.
- Co-founders include Piers Ingram, Michael Warren and David Berman.
- The company has 12 people and lists offices in London, Singapore and Los Angeles.
- Metis has not released a model or published test results.
Why Metis says LLMs fall short
The white paper lists three problems. Large language models hallucinate, and nobody can audit their reasoning. Users also cannot inspect the knowledge behind an answer, because that knowledge sits inside the model’s weights. The founders argue that these are not merely engineering inconveniences that will disappear as models get bigger. “There exists, however, a significant set of problems that no amount of scaling can solve,” the white paper says.
Metis also takes aim at the step-by-step “chain of thought” that many models produce. The company argues that this narrative often does not reflect how the model actually reached its answer. A model may present a tidy sequence of reasoning steps while its underlying computation followed a different path. For low-stakes tasks, that might not matter. For high-stakes tasks, it matters a great deal. Ingram put the business case more bluntly. “When you think about higher-stakes cases in engineering or science, or medical or legal reasoning, the arguments that the models come up with basically can’t be trusted,” he said.
The AlphaGo lesson
The white paper opens with Move 37, the move AlphaGo played against Lee Sedol in 2016. Experts first took it for a blunder. AlphaGo’s own policy network had rated the chance of a human expert playing it at about one in 10,000. The founders argue that the move came from search and evaluation, which they call reasoning, rather than from prediction alone. That distinction is central to Metis’s pitch. AlphaGo did not simply pattern-match its way to a move. It searched through possible futures, evaluated positions and selected an action under uncertainty. Move 37 became a symbol of what can happen when a system reasons through a problem rather than relying on intuition alone.
The background helps explain why the founders see AlphaGo as a template. AlphaGo combined deep neural networks with Monte Carlo tree search. A policy network narrowed the search toward promising moves, while a value network estimated who was ahead. The search then did the heavy lifting, exploring lines of play and weighing outcomes. When AlphaGo defeated European champion Fan Hui in 2015 and then Lee Sedol in 2016, it demonstrated that machine reasoning could outperform human intuition in one of the most complex board games ever devised. Move 37 was not a routine prediction from human data. It was, in the founders’ telling, the product of a system that searched, evaluated and chose.
How Metis wants reasoning to work
Metis wants AI to keep an explicit record of what it knows and how certain it is. A separate judgment system would accept a new claim only when the evidence supports it. The system would then pick each next step by how much it cuts uncertainty. The paper models this on philosopher Jaakko Hintikka’s idea of reasoning as a game between an inquirer and nature. Hintikka, a Finnish philosopher and logician, developed an interrogative model of inquiry in which reasoning is treated as a structured process of asking questions, receiving answers and updating a body of knowledge. Metis is translating that philosophical framework into an engineering blueprint.
As in AlphaZero’s self-play, the system would then learn better strategies from its own recorded inquiries. AlphaZero famously learned chess, shogi and Go by playing against itself, improving through repeated cycles of search and evaluation. Metis wants to apply a similar loop to general reasoning: the system records its inquiries, judges which steps reduced uncertainty, and uses those records to improve how it reasons next time. The company says this would mean fewer hallucinations and reasoning that anyone can check step by step. Users could also update or remove knowledge without retraining the model. Smaller models could handle many steps, which Metis says could cut costs and put reasoning on devices and robots.
Ingram said the design is “much much more compute-efficient” than frontier models. “We are very much not going in this direction of ever larger models and ever larger training budgets,” he said. That is a direct challenge to the prevailing scaling narrative. Many leading AI labs have bet that bigger models, more data and more compute will produce better reasoning. Metis is betting that the architecture of reasoning itself matters more than raw scale. If that bet is right, the company could build systems that are cheaper to run, easier to audit and more suitable for deployment in settings where errors are costly.
No model yet
The white paper describes an approach, not a product, and it publishes no test results. Ingram said Metis had achieved “encouraging” results on frontier reasoning problems at a fraction of the cost, without giving figures. Metis has not said when it will release its first model. Ingram would not comment on reports that Metis is raising tens of millions of dollars. He said Metis had raised money “more than adequately to date.”
Metis joins a cluster of new London AI labs led by DeepMind alumni and other big names. They include Ineffable Intelligence, founded by Graepel’s former AlphaGo colleague David Silver, and Richard Socher’s Recursive Superintelligence. The concentration of talent in London has made the city a hub for ambitious AI startups, particularly those founded by researchers who worked on reinforcement learning, games and reasoning at DeepMind. Metis is trying to stand out by arguing that the next leap in AI will not come from simply scaling up language models. Ingram called Metis’s approach “really a very different, almost philosophical approach.”