Former Google DeepMind top researcher launches London “compute-efficient” AI company

Former Google DeepMind researcher Thore Graepel has launched a London-based AI company called Metis Reasoning, aiming to develop a "compute-efficient" alternative to the large language models (LLMs) used by companies like Anthropic and OpenAI. Graepel, who co-created AlphaGo, the historic AI program that defeated human players at the Chinese board game Go, is co-founding Metis with other experts in robotics, engineering, and AI.

The startup, named after the Greek goddess of reasoning and wisdom, is led by a team of 12 members, including Graepel, Piers Ingram, Michael Warren, and David Berman. Ingram, a serial technology and biotech entrepreneur, met Graepel at a Cambridge barbecue and believes that current LLMs lack the ability to be trusted for higher-stakes cases in engineering, science, medical, or legal reasoning due to their unreliability and lack of an audit trail.

Metis's approach is based on Graepel's work at AlphaGo and seeks to create a more verifiable, robust, and safer AI system. Ingram claims that Metis's architecture is significantly more compute-efficient than frontier models and has achieved encouraging results in frontier reasoning problems at a fraction of the cost of these models. The company is currently raising tens of millions of dollars to fund its operations.

Other recent AI companies in London, such as Ineffable Intelligence and Recursive Superintelligence, are also pursuing different paths in AI development. Ingram is confident that Metis's approach is unique and philosophical in nature.