Thore Graepel, one of the creators of AlphaGo, has launched Metis Reasoning, a London-based artificial intelligence company building a fundamentally different kind of AI. Rather than predicting the next word in a sequence, Metis is developing systems that gather evidence, test hypotheses, and verify their conclusions — an architecture inspired by the search-and-planning methods behind AlphaGo’s famous 2016 victory over Lee Sedol.
Graepel, who left Google DeepMind this summer, co-founded the 12-person team with serial entrepreneur Piers Ingram, robotics specialist Michael Warren, and Cambridge Consultants AI head David Berman. According to reports, the company is initially raising tens of millions of dollars, with potentially much larger rounds to follow.
The central argument in Metis’s launch-day white paper, “The Age of Reasoning,” is that today’s language models reason unreliably. They generate fluent answers but leave no audit trail — a serious barrier to adoption in sectors like engineering and robotics, where every conclusion must be verifiable. Metis aims to fix this by separating what a system knows from how it reasons, keeping an explicit record of evidence, uncertainty, and the support behind each decision.
Graepel describes the approach as broadly neuro-symbolic: neural networks interpret language and images, while structured reasoning handles the logic. The company believes this design is more robust, more verifiable, and safer than scaling large models further — a pitch that will be tested against frontier systems like Gemini and DeepSeek.
See how today’s leading models compare in our Claude vs. Gemini comparison.
Source: Tech.eu; Pathfounders
