Thore Graepel, a key contributor to DeepMind’s AlphaGo, has co-founded Metis Reasoning in London. The new AI lab challenges the reliability of large language models’ reasoning and proposes a more explicit, auditable approach inspired by AlphaGo’s strategy-driven intelligence.
- Metis targets trustworthiness and auditability lacking in current large language models.
- The team includes experts from DeepMind, biotech investment, robotics, and UK government AI advisory.
- Their method models reasoning as a query game to reduce hallucination and improve explainability.
What happened
Thore Graepel, known for his work on Google DeepMind’s AlphaGo, has launched Metis Reasoning, a new AI laboratory based in London that challenges the established paradigm around large language models. The founding team includes Piers Ingram, Michael Warren, and David Berman, combining expertise in technology entrepreneurship, robotics, and AI policy advising for governments and NATO.
Metis has published a white paper titled “The Age of Reasoning,” outlining their critique of current AI models’ inability to produce reliable, inspectable reasoning. They emphasize that large-scale models hallucinate and lack transparent reasoning chains, which undermines trust, particularly in high-stakes applications. The company seeks to create AI capable of explicit and auditable reasoning, inspired by AlphaGo’s strategic search and evaluation processes rather than mere predictive modeling.
Why it matters
Current large language models operate as black boxes where the reasoning behind answers is embedded in model weights and cannot be easily inspected or audited. This opacity leads to hallucinations or false conclusions, which is a significant concern in critical fields such as medicine, law, and engineering where incorrect reasoning can have severe consequences.
Metis’s approach could set a new standard for AI systems by introducing a separate judgment system that validates claims based on supporting evidence and maintains a transparent record of knowledge and certainty. This method promises fewer hallucinations, better trustworthiness, and the ability to update or remove knowledge without costly retraining. Additionally, their focus on efficiency suggests smaller, more capable models that can run on devices and robots, reducing reliance on large-scale compute resources.
What to watch next
Metis Reasoning has not yet disclosed a release date for its first AI model, but co-founder Piers Ingram described the early results on complex reasoning tasks as encouraging, emphasizing significantly lower computational costs compared to leading models. The company claims to have sufficient funding to advance their research but has not confirmed ongoing fundraising efforts reportedly worth multiple tens of millions.
The founding of Metis adds to a growing cluster of London-based AI ventures led by former DeepMind researchers, such as Ineffable Intelligence and Recursive Superintelligence. Given its philosophical and methodical approach, Metis will be closely watched for how well it can deliver on its promise of transparent and trustworthy AI reasoning, potentially influencing the future design of AI systems beyond mere language generation.