London gains another AI lab as AlphaGo veteran launches Metis Reasoning

Metis Reasoning, a 12-person start-up reported to be seeking tens of millions of dollars, says machines should accept a conclusion only when evidence supports it, and wants to borrow the architecture that beat the world’s best Go players to get there.

Thore Graepel, a co-creator of AlphaGo and a former research lead at Google DeepMind, has launched Metis Reasoning, a London company developing AI systems designed to show how they reach their conclusions. Graepel left Google DeepMind this summer, and the company emerged from stealth on Tuesday alongside a white paper, The Age of Reasoning, that sets out its research programme.

In announcing the launch, Graepel said the company wants to build systems that can “establish what they know, test conclusions against evidence” and improve their reasoning through experience. He is due to present the ideas at Nvidia’s GTC conference in Berlin on Tuesday.

Graepel and his co-founders have assembled a 12-strong team of researchers, mathematicians, robotics experts and tech entrepreneurs. The other founders are Piers Ingram, a serial technology and biotech entrepreneur who has raised more than $350m; Michael Warren, a robotics and autonomous systems specialist with experience at Boeing; and David Berman, head of AI at Cambridge Consultants and an adviser on AI to the UK government and NATO. The white paper lists the company in London, Singapore and Los Angeles.

The company’s case rests on a familiar complaint about large language models. Metis accepts that they can produce fluent answers and handle some multi-step tasks, but argues they reason unreliably and leave no audit trail, which it says holds back adoption in industries where work must be verified. Ingram put it more sharply: in engineering, science, medicine or law, the arguments the models produce “basically can’t be trusted.”

The white paper extends that critique. It cites research showing that the step-by-step “chain of thought” traces models produce often fail to reflect the computation behind the final answer, and argues that because training stores knowledge and reasoning together in a model’s weights, the provenance of that knowledge cannot be established or curated. The authors link this to the modest economic returns from AI so far, pointing to coding as the exception, since software output is unusually easy to test, verify and roll back.

Their proposed remedy draws on AlphaGo’s 2016 victory over Lee Sedol. The paper notes that AlphaGo’s policy network gave its famous Move 37 roughly a one in ten thousand chance of being played by a human expert; the move came instead from a search process that weighed alternative lines of play over an explicit game tree. Metis argues that today’s language models resemble a powerful version of AlphaGo’s policy network, without the machinery around it.

The company’s design would supply that machinery. It centres on an explicit, persistent “epistemic state” recording what has been established and with what certainty, with any change to a claim’s status approved by an independent judgment system using criteria set separately from the component that proposes it. The system’s objective during reasoning is to reduce uncertainty about the question at hand. The authors model the process on the philosopher Jaakko Hintikka’s interrogation game, in which an inquirer questions nature and draws deductions, and propose that the strategy for playing it be learned from recorded inquiries, much as AlphaZero learned games through self-play.

The claimed payoffs are commercial as well as scientific. Because knowledge sits apart from the reasoning process, the paper says, it could be updated, removed or access-controlled without retraining, and simpler reasoning steps need not be handled by large, expensive neural models. Ingram said the architecture was far more compute-efficient than frontier models and that the team was deliberately avoiding ever-larger models and training budgets.

Much of this is still unproven. The white paper is a statement of thesis and contains no experimental results. Ingram’s claim of encouraging results on frontier reasoning problems came with no figures attached, and it is unclear when Metis will release its first model.

The company’s finances are similarly opaque. Bloomberg reported in September that the talks were at an early stage, with Graepel seeking initial capital from a small group of investors and possibly a later round of hundreds of millions of dollars at a higher valuation. Ingram declined to comment on those reports, saying only that the company had been funded more than adequately to date.

Metis enters a crowded field in London. Other recent arrivals include Ineffable Intelligence, founded by Graepel’s former AlphaGo colleague David Silver, and Recursive Superintelligence, co-founded by the German computer scientist Richard Socher. Asked about overlap, Ingram said neither side knew exactly what the other was doing, while describing the Metis approach as very different and almost philosophical.

The first independent test of those claims will come when Metis publishes results or a model that outside researchers can examine. The company has given no date for either.