Graepel, a key figure in developing Google DeepMind's AlphaGo, has established Metis Reasoning as a new London-based venture. The startup's central thesis contradicts prevailing industry wisdom: scaling up language models will not resolve their fundamental reasoning failures. Rather than pursuing ever-larger models, Metis proposes building AI that maintains an explicit, verifiable account of its knowledge and reasoning steps.
Graepel departed Google DeepMind this summer, according to Tech.eu. Joining him are Piers Ingram, a technology and biotech entrepreneur credited with raising over $350m; Michael Warren, a robotics specialist formerly at Boeing; and David Berman, who leads AI initiatives at Cambridge Consultants and advises both the UK government and NATO on artificial intelligence matters. The team currently numbers 12 people, with operations spanning London, Singapore and Los Angeles.
The company outlined its position in a white paper titled "The Age of Reasoning". Ingram articulated the stakes for high-consequence domains: "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 told Tech.eu.
The case against scaling
Metis identifies three core limitations in contemporary language models. First, these systems generate false information—hallucinations—with no mechanism for auditing how they arrived at their conclusions. Second, the knowledge embedded in a model's weights remains opaque and inaccessible to users. Third, the reasoning paths models report through techniques like chain-of-thought do not necessarily correspond to their actual decision-making process.
"There exists, however, a significant set of problems that no amount of scaling can solve,"
Metis Reasoning white paper
Learning from AlphaGo
The white paper begins with Move 37, the controversial play AlphaGo executed against Lee Sedol in 2016. Initially dismissed as an error by observers, the move carried only a one-in-10,000 probability rating from AlphaGo's own policy network. Metis contends this move exemplified reasoning—the product of search and evaluation—rather than mere pattern prediction.
Building on this insight, Metis proposes systems that maintain explicit records of their knowledge and confidence levels. A separate evaluation mechanism would validate new claims only when evidence justifies them. The system would then select subsequent steps based on their capacity to reduce uncertainty. This framework draws inspiration from philosopher Jaakko Hintikka's conceptualization of reasoning as a dialogue between an inquirer and the natural world.
Following the self-play model pioneered by AlphaZero, Metis's system would refine its strategies through learning from its own documented reasoning chains. This architecture promises fewer hallucinations and fully transparent, step-by-step reasoning that users can verify. Knowledge could be modified or deleted without requiring complete model retraining. Smaller models could execute numerous reasoning steps, potentially reducing computational demands and enabling reasoning capabilities on edge devices and robotic systems.
Ingram emphasized the efficiency gains: "We are very much not going in this direction of ever larger models and ever larger training budgets," he stated, describing the approach as "much much more compute-efficient" than current frontier models.
Early stage, no public results
The white paper presents a conceptual framework rather than a finished product, with no published experimental data. Ingram disclosed that Metis has achieved "encouraging" results on complex reasoning benchmarks at substantially reduced cost, though he declined to provide specifics. The company has not announced a timeline for releasing its first model.
Regarding funding, Ingram deflected questions about reports of a multi-tens-of-millions fundraise, stating only that Metis had secured capital "more than adequately to date".
Metis joins an emerging ecosystem of London-based AI laboratories founded by DeepMind alumni and industry veterans. This includes Ineffable Intelligence, launched by David Silver, Graepel's former colleague on the AlphaGo project, and Richard Socher's Recursive Superintelligence. Ingram characterized Metis's direction as "really a very different, almost philosophical approach" within this competitive landscape.
Source: The Next Web
