Former Google DeepMind research lead and AlphaGo co-creator Thore Graepel is bringing his new AI startup, Metis Reasoning, out of stealth today, claiming its alternative approach to AI could solve the hallucination and reliability problems that continue to dog LLMs. Graepel left Alphabet Inc.’s research lab this summer and is understood to be raising a big round from VCs.
Headquartered in London, Metis is developing AI systems that gather evidence, test hypotheses and verify their conclusions, rather than simply generating plausible answers based on the order tokens are most likely to fall.
Speaking to Pathfounders, Graepel argued that today's AI models are fundamentally limited by their reliance on predicting text rather than establishing verifiable facts.
"I'm a scientist at heart," he said. "I think that the scientific method is really the only way for humans to reliably create new knowledge."
In a white paper which will be published today, alongside its launch, ‘The Age of Reasoning’ (listed and attached below) Metis argues that solving the problem requires a different AI architecture, rather than simply building larger models or extending their chains of thought.
Its proposed approach separates knowledge from the reasoning process, maintaining an explicit record of what an AI system knows, what remains uncertain and what evidence supports each conclusion.
"It knows what it knows. It knows what it doesn't know," Graepel told me, explaining how the system advances its understanding by developing hypotheses, gathering evidence and testing assumptions.
Rather than abandoning neural networks, he said Metis combines their ability to interpret language and images with explicit knowledge representation and structured reasoning. This is an approach he acknowledges broadly falls within the ideas around “neuro-symbolic” AI.
The approach Metis takes draws on his experience developing AlphaGo, the DeepMind system that defeated Go champion Lee Sedol in 2016. AlphaGo combined neural networks with a structured game-tree search to evaluate possible moves and outcomes.
"You can think of Metis Reasoning as a generalisation of AlphaGo," he said. Metis wants to apply similar principles to open-ended problems in science, medicine and robotics, where AI must contend with uncertain information and changing environments.
Graepel argues that current LLMs cannot reliably explain how they reached a conclusion, even when generating elaborate chains of thought.
"More scaling will never give you an auditable reasoning trace that you can actually trust," he said.
Asked if the approach could avoid the problems thrown up by AI agents, Graepel said: “Even in hindsight, we cannot really understand what happened because all we have access to are these chain of thought traces that don't really give away everything about why agents make certain decisions. And also we cannot really give them reliable guardrails… the token-based approach just doesn't allow that… we need a more explicit knowledge representation.”
“You can think of the ‘original sin’ of the transformer architecture is that everything is intermediate in the weights.” He said Metas reasoning achieves that because it separates knowledge and reasoning: “The knowledge of the world and the process, the reasoning process that acts on it. And suddenly that unlocks a much safer, more reliable, trustworthy AI.”
He said the lack of transparency in LLMs could become particularly problematic as autonomous AI agents take on more consequential tasks. Metis believes its architecture could allow developers to inspect and modify a system's knowledge, impose more reliable safeguards and audit its decisions.
“Because now you can inspect the knowledge that you have because it's explicitly represented. You can look at it, you can change it, you can edit it, and you can access control it. But also you can now improve your reasoning process because you've isolated this process,” he said.
The company also claims that accumulating and reusing verified knowledge could reduce computational costs, rather than requiring models to generate answers afresh each time.
Metis says it already has a working prototype. "We have a prototype that can reason about PhD-level scientific problems," Graepel told Pathfounders, adding that scientists connected to the founders' biotechnology ventures have tested it.
He expects an "industry-hardened" system to be available for external testing within three to six months, although the company has yet to publish independent benchmarks demonstrating its claimed advantages.
Robotics is another priority. "Robotics is a reasoning problem," Graepel said, arguing that autonomous machines need to understand their environments and reliably assess the consequences of their actions.
Metis is entering a market dominated by OpenAI, Google DeepMind and Anthropic, whose increasingly capable reasoning models are largely built on scaling existing architectures.
Graepel believes their focus creates an opening for alternative approaches. "They are so focused on the scaling hypothesis that a lot of other research is falling by the wayside," he said.
His ambitions go further than specialist applications, however. "Ultimately, I think this kind of AI system has the potential to replace all existing AI systems because we want that reliability, that safety, and that auditability," he said.
Alongside Graepel, who is also Chair of Machine Learning at UCL, Metis's founding team includes Piers Ingram, a serial technology and biotech entrepreneur whose ventures have raised more than $350 million; Michael Warren, a robotics and autonomous systems expert with experience at DARPA, Boeing and GM; and David Berman, Head of AI at Cambridge Consultants and an adviser to NATO and the UK government.
The company is building its operations around London's King's Cross AI ecosystem, with some of its team based in Singapore, where Ingram is based.
Metis has not yet disclosed its funding, and Graepel declined to discuss financing but confirmed fundraising discussions are ongoing.
According to a recent report in Bloomberg, Graepel is seeking to raise tens of millions of dollars for the startup, although he declined to comment during an interview for the Pathfounders podcast (shortly to be released).
Graepel is understood to be attracting high investor interest, given his involvement with AlphaGo and DeepMind, especially as other prominent researchers involved with AlphaGo, such as OpenAI co-founder Ilya Sutskever and Ineffable Intelligence founder David Silver, have both raised billions for their AI startups.
