AlphaGo co-creator Thore Graepel launches Metis Reasoning, a compute-efficient AI startup that wants machines to show their work

• Who: Thore Graepel — AlphaGo co-creator, former Google DeepMind research lead (left this summer), and Chair of Machine Learning at UCL.
• What: Metis Reasoning, a London startup building 'compute-efficient' AI that keeps an explicit, editable record of what it knows, what it doesn't know, and what evidence backs each conclusion.
• When: launched October 20, 2026, alongside the founding white paper 'The Age of Reasoning.'
• Team: 12 people; co-founders Piers Ingram (tech/biotech entrepreneur, ventures raised $350M+), Michael Warren (robotics, ex-Boeing), and David Berman (head of AI at Cambridge Consultants, NATO/UK government advisor).
• Funding: fundraising discussions ongoing; Bloomberg has reported the company is seeking tens of millions of dollars.
• Milestones: working prototype reasons about PhD-level scientific problems; 'industry-hardened' system expected for external testing within 3–6 months; no independent benchmarks published yet.

Thore Graepel, the Google DeepMind researcher who co-created AlphaGo, is back — this time with a London startup that says the path to reliable AI runs through reasoning, not scaling. Metis Reasoning, launched October 20 with a 12-person team and a white paper titled The Age of Reasoning, is betting that machines which keep an auditable record of what they know can succeed where ever-larger language models stall.

Graepel left Google DeepMind this summer. His new company's pitch is blunt: current LLMs cannot reliably explain how they reach conclusions, which makes them unfit for the high-stakes work — science, engineering, medicine, robotics — where correctness actually matters.

The case against scaling

Metis's founding argument is that reasoning, not pattern-matching, is the missing ingredient. In an interview with Pathfounders, Graepel called today's models fundamentally limited by their reliance on predicting text rather than establishing verifiable facts. "I'm a scientist at heart," he said. "I think that the scientific method is really the only way for humans to reliably create new knowledge."

His co-founder Piers Ingram put it more bluntly to Tech.eu: "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. I don't think that's a good foundation for what we want our AI to be."

Graepel argues the flaw is architectural — what he calls the "original sin" of the transformer: everything is intermediate in the weights. Metis instead separates knowledge from the reasoning process, keeping an explicit, editable record of what the system knows, what remains uncertain, and what evidence supports each conclusion. "It knows what it knows. It knows what it doesn't know," he said. The approach, broadly neuro-symbolic, is "much much more compute-efficient" than frontier models — deliberately not a race toward ever-larger training budgets.

An AlphaGo for open-ended science

The name Metis comes from the Greek goddess of wisdom, but the blueprint is the 2016 system that made Graepel famous. AlphaGo combined deep neural networks with structured game-tree search to defeat Go champion Lee Sedol 4–1. "You can think of Metis Reasoning as a generalisation of AlphaGo," Graepel told Pathfounders — applying the same search-plus-learning lesson to open-ended problems in science, medicine and robotics, where information is uncertain and environments change.

The company says it already has a working prototype. "We have a prototype that can reason about PhD-level scientific problems," Graepel said, adding that scientists connected to the founders' biotechnology ventures have tested it. An "industry-hardened" system should be ready for external testing within three to six months — though Metis has yet to publish independent benchmarks backing its claims.

London's post-LLM cluster

Metis is the latest in a London exodus from the scaling playbook. Alongside Graepel, the founding team includes Ingram, a serial technology and biotech entrepreneur whose ventures have 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 adviser to NATO and the UK government.

The 12-strong team is building around London's King's Cross AI cluster, with some members based in Singapore, where Ingram is based. On funding, Graepel declined to discuss specifics but confirmed fundraising discussions are ongoing; Bloomberg has reported the company is seeking tens of millions of dollars. Investor appetite for DeepMind-alumni alternatives is real: David Silver, Graepel's former AlphaGo colleague, raised $1.1 billion at a $5.1 billion valuation for Ineffable Intelligence in April.

What happens next

Metis published its founding white paper, The Age of Reasoning, alongside today's launch, arguing the field needs a different architecture rather than simply larger models or longer chains of thought. The company says its approach suits industries such as robotics and engineering, where verifiable work is a prerequisite, not a luxury.

When the public gets to test it remains the open question — it is still unclear when Metis will publicly launch its first model. Graepel's ambition, though, is total: "Ultimately, I think this kind of AI system has the potential to replace all existing AI systems because we want that reliability, that safety, and that auditability."

Sources: Tech.eu, Pathfounders, metisreasoning.ai