Thore Graepel, one of the researchers behind Google DeepMind's AlphaGo, has launched a new artificial intelligence company in London. Metis Reasoning went public on October 20 with a white paper titled "The Age of Reasoning" and a blunt founding claim: today's large language models cannot be trusted to reason, and no amount of scaling will fix that.
Who is behind Metis
Graepel, a co-creator of AlphaGo - the DeepMind program that made history by beating human champions at Go - left Google DeepMind this summer, according to Tech.eu. His co-founders are Piers Ingram, a serial technology and biotech entrepreneur who has raised more than $350 million; Michael Warren, a robotics and autonomous systems expert with experience at Boeing; and David Berman, head of AI at Cambridge Consultants and an advisor to the UK government and NATO on AI.
The company - named after the Greek goddess of wisdom - has assembled a 12-strong team of researchers, mathematicians, robotics experts and tech entrepreneurs, with the white paper listing offices in London, Singapore and Los Angeles. Ingram and Graepel met, by Ingram's telling, at a Cambridge barbecue where they ended up chatting about Turing machines.
The problem Metis says scaling cannot solve
The white paper lists three failures in current models. LLMs hallucinate, and there is no way to audit how they reached an answer. The knowledge behind an answer is locked inside model weights, so users cannot inspect, update or remove it. And the step-by-step chain of thought a model shows often does not reflect how it actually arrived at its conclusion.
"There exists, however, a significant set of problems that no amount of scaling can solve," the paper argues. Metis frames this as a barrier to wider adoption: many industries still cannot verify an AI system's work, so they cannot rely on it. Ingram made the business version of the point to Tech.eu: in high-stakes settings - engineering, science, medical or legal reasoning - "the arguments that the models come up with basically can't be trusted."
Reasoning versus pattern matching
At the heart of the pitch is a distinction between two things modern AI does. Pattern learning is what lets a model produce a fluent answer from what it has seen before. Reasoning, as Metis defines it, is the process of working through a problem and connecting information to reach a conclusion. The company's case is that frontier models are very good at the first and structurally weak at the second - and that the weakness shows up precisely where mistakes are most expensive. That, Metis argues, is why adoption in many industries remains constrained: organizations cannot verify how a fluent answer was reached, so they cannot stake anything on it.
The AlphaGo lesson, applied to reasoning
The paper opens with Move 37, the move AlphaGo played against Lee Sedol in 2016 that commentators first read as a mistake. AlphaGo's own policy network had estimated the chance of a human expert choosing it at roughly one in ten thousand. The founders argue the move came from search and evaluation - what they call reasoning - rather than from pattern prediction alone, and that general AI needs the same separation.
Metis's proposed architecture keeps an explicit record of what the system knows and how certain it is. A separate judgment component accepts a new claim only when the evidence supports it, and each next step is chosen by how much it reduces uncertainty. The paper models this on philosopher Jaakko Hintikka's framing of reasoning as a game between an inquirer and nature. As in AlphaZero's self-play, the system would learn better inquiry strategies from its own recorded sessions.
What that would buy
If it works, Metis says the design would mean fewer hallucinations and reasoning that can be checked step by step. Knowledge could be updated or deleted without retraining the model. Smaller models could carry out many reasoning steps, which the company says would cut costs and make reasoning practical on devices and robots - the industries it expects the thesis to suit best.
Ingram calls the architecture "much, much more compute-efficient" than frontier models and says Metis is "very much not going in this direction of ever larger models and ever larger training budgets." Asked why Anthropic and OpenAI are not pursuing a similar path, his answer was simply: "I think they are thinking about it differently."
What Metis has not shown yet
The white paper describes an approach, not a product, and it publishes no benchmark results. Ingram says the team has achieved "encouraging" results on frontier reasoning problems at a fraction of the cost of frontier models, but offered no figures, and it is unclear when Metis will publicly launch a first model. On funding, he declined to comment on reports that the company is initially raising tens of millions of dollars, saying only that Metis has been able to fund itself "more than adequately to date."
Part of a London wave
Metis joins a growing cluster of London AI labs started by DeepMind alumni and other prominent researchers, as investors pour money into alternatives to LLM-only approaches. Ineffable Intelligence was founded by David Silver, Graepel's former AlphaGo colleague, and Richard Socher recently co-founded Recursive Superintelligence. Asked about overlap, Ingram said neither side knows exactly what the other is doing - but argued Metis's approach is "really a very different, almost philosophical approach." Whether that philosophy turns into working systems is now the thing to watch.