Thore Graepel launches London AI company Metis Reasoning

Thore Graepel, who helped build AlphaGo at Google DeepMind, has started Metis Reasoning in London to develop AI with reasoning that users can examine. The company outlines its approach in a white paper rather than a finished product.

In the paper, titled “The Age of Reasoning,” Metis argues that large language models generate unreliable claims while leaving users unable to inspect the knowledge and reasoning behind their answers. The founders want to draw on the search and evaluation techniques used by AlphaGo.

Their proposed design would maintain an explicit account of knowledge and uncertainty. A separate system would evaluate evidence before accepting claims, while subsequent steps would be chosen according to their potential to reduce uncertainty. The system would also learn strategies from records of its own inquiries.

Metis says this approach could reduce hallucinations, make reasoning checkable and let users revise or delete knowledge without retraining. It also says using smaller models for many steps could lower computing costs. Those claims are not accompanied by published test results. Co-founder Piers Ingram told Tech.eu that initial work had delivered encouraging results at lower cost, but gave no figures.

The founders have not specified when a first model will be released. Alongside Graepel and Ingram, the founding team includes Michael Warren, a robotics expert formerly at Boeing, and David Berman, head of AI at Cambridge Consultants. Tech.eu puts the workforce at 12 people, while the white paper lists London, Singapore and Los Angeles offices.