Thore Graepel, CEO & Co-Founder at Metis Reasoning: AI needs to reason like a scientist before we can trust it

Metis bets that separating knowledge from reasoning makes AI trustworthy

Metis Reasoning is betting that trustworthy AI needs a system that reasons the way a scientist does. Large language models predict tokens and bury knowledge and reasoning together in their weights, so nobody can audit how an answer was reached. Metis keeps knowledge explicit and treats reasoning as a separate process that reduces uncertainty against evidence. That makes answers inspectable in domains where people must see why.

Takeaways

• Language models predict tokens and do not reason. A model has no notion of truth and holds no beliefs it can update. Graepel defines reasoning as starting from fuzzy beliefs and cutting uncertainty by forming hypotheses and testing them against evidence. He calls it a big open problem in AI.
• Separating knowledge from reasoning makes AI inspectable. Knowledge entangled in model weights is the original sin of the transformer. Metis keeps knowledge explicit, so it can be edited and access-controlled like a database, while reinforcement learning improves reasoning separately. Verified answers stored with evidence could make later answers orders of magnitude faster.
• AlphaGo's Move 37 shows reasoning creates knowledge. The policy network gave the move about a one in 10,000 chance, and commentator Michael Redmond called it a mistake. Search-based deliberation, Kahneman's system two, dared to play it. Graepel concludes that new knowledge needs explicit reasoning.
• Trust matters more than hallucination. Chain-of-thought traces are not the real causal path to an answer, and models usually do not reach 100% even on benchmarks. In medical diagnosis, engineering, air traffic control and robotics, people must see what evidence was used, how it was judged and how it was synthesised.
• Good reasoning cuts uncertainty with few moves. Metis rests on a persistent epistemic state of beliefs, reasoning moves that change it, and an objective of reducing uncertainty, tracked with probability theory. Good reasoning gets there with few, clever, low-cost moves. Go is deterministic, so the general problem is harder.
• Language models interpret while evidence supplies answers. Models handle semantic interpretation and can propose hypotheses, even nonsensical ones, because evidence later weeds them out. For the capital of France, Metis questions assumptions, forms hypotheses, checks sources such as Wikipedia and stores the result. Reasoning must work on propositions, not tokens.

What Metis does next

• The team includes physicists, mathematicians and roboticists, based in Singapore, the West Coast and London.
• London is planned as headquarters.