The thesis: from weights to a verifiable registry
Former DeepMind researcher Thore Graepel has founded the startup Metis Reasoning with the goal of building artificial intelligence systems capable of tracking what they know and verifying every conclusion against the available evidence. According to The Decoder, Graepel presented the approach at Nvidia’s GTC conference in Berlin. His conviction is that the same principle that in 2016 enabled AlphaGo to beat Go professional Lee Sedol could make AI systems more reliable, including robots.
Metis argues that current language models amount only to AlphaGo’s prediction network, the one that proposed possible moves. AlphaGo paired it with a systematic search of game sequences, finding the famous Move 37, which the network would have assessed as having a probability of about one in ten thousand for a human player. Metis’s idea is to extend that kind of search to general reasoning.
How the system would work
The mechanism revolves around an explicit knowledge registry, which the company compares to the game board. The registry records what is known, the system’s degree of confidence and the questions still open. A separate module verifies new entries against calculations, formal proofs or empirical evidence; if an assumption fails, all conclusions that depended on it are reassessed. A “move” can consist of a question, an experiment or a challenge to a premise, and the system chooses the one that most reduces uncertainty about the initial question, arriving at an answer or a reasoned abstention.
As in AlphaGo, two neural networks propose and evaluate moves: they are trained on recorded prior investigations, but according to Metis they cannot violate the verification rules. The company claims that the setup should reduce hallucinations, produce explanations with source citations and make it possible to modify or delete knowledge without retraining the model. Because reasoning is broken down into small steps, the approach could favor lighter models on edge devices and robots.
The theoretical framework rests on a critique of current models: more data and more computing power improve them, but do not solve fundamental problems such as fabricated answers, the non-verifiability of reasoning chains and knowledge trapped in weights. In Metis’s thesis, hallucinations arise from the training process itself, because models learn to produce plausible rather than true text, and the chains of thought of reasoning models often do not show how an answer was reached. Hence the company’s argument that AI has so far had limited economic impact: in fields such as medicine, finance, chip design and robotics, intrinsic verifiability would be needed.
What is missing and existing reference points
The white paper, however, describes only a research program: there is no working system, no benchmarks and no comparisons with current reasoning models. The team admits that the real world, unlike Go, offers only partial information, that obtaining evidence can be costly and that some questions have no answer. The most open question concerns how a verified knowledge registry could take shape in messy domains beyond textbook examples, considering that symbolic AI systems failed precisely because all knowledge had to be entered by hand.
Metis is not alone on the ideas front, however: Microsoft’s GraphRAG organizes knowledge outside the model’s weights in a graph, while Google’s AlphaGeometry2 pairs a language model with a symbolic engine that verifies mathematical proofs. Metis says it wants to extend these principles to open-ended real-world questions.
On a personal level, The Decoder recalls that Graepel earned a doctorate in machine learning in Berlin in 2001, contributed to TrueSkill at Microsoft Research, worked at DeepMind on AlphaGo, AlphaZero and MuZero and was on Google DeepMind’s post-AGI team, which he left in the summer; he also holds the chair in machine learning at University College London. According to Bloomberg, he is raising a round in the tens of millions of dollars, with a follow-on of several hundred million at a higher valuation already in preparation. The material does not allow these financial figures to be independently verified or the effectiveness of the approach to be assessed.
