AlphaGo Co-Creator Launches Metis to Prove LLMs Can't Reason
By Daniel Rao · London, October 20, 2026
One of the men who taught a machine to humiliate a world champion at Go has just told the most powerful industry on earth that it is building on sand. Thore Graepel — co-creator of AlphaGo, until this summer a research lead at Google DeepMind — launched Metis Reasoning in London today with a thesis that reads like a punch in the mouth: large language models don’t actually reason, and no amount of scaling will ever make them. If he’s right, the trillion-dollar bet on ever-bigger models is aimed at entirely the wrong target.
Graepel left DeepMind this summer and has assembled a 12-strong team of researchers, mathematicians, robotics specialists and entrepreneurs to pursue what Tech.eu describes as a “compute-efficient” alternative path to the frontier models of Anthropic and OpenAI. The company is headquartered in London, is building its operations around the King’s Cross AI cluster, and keeps part of its team in Singapore. It is understood to be raising tens of millions of dollars, first reported by Bloomberg in September.
We checked metisreasoning.ai on launch morning. The entire pitch fits on a single page — “Intelligence from first principles” — followed by three terse claims and an email address for investors and researchers. No product. No benchmarks. No team page. Right now, Metis is a promise and a white paper. But it is a promise made by someone with an unusual right to make it.
What Metis is actually building
The white paper, titled The Age of Reasoning and published alongside today’s launch, argues that AI’s central failure is architectural, not a matter of scale. Metis’s position is that today’s models generate fluent answers by predicting which tokens should come next — and that this mechanism, however impressive, can never produce reasoning you can trust, because nothing in the system records what it actually knows.
Its proposed alternative separates knowledge from the reasoning process. The system would keep an explicit record of what it knows, what remains uncertain, and what evidence supports each conclusion. As Graepel put it to Pathfounders: “It knows what it knows. It knows what it doesn’t know.”
The approach advances, he explained, by developing hypotheses, gathering evidence and testing assumptions — the scientific method, mechanised. “I’m a scientist at heart,” Graepel said. “I think that the scientific method is really the only way for humans to reliably create new knowledge.”
Rather than abandoning neural networks, Metis wants to keep their ability to interpret language and images while bolting on explicit knowledge representation and structured reasoning — an approach Graepel acknowledges sits broadly within “neuro-symbolic” AI. And the lineage is unmistakable: “You can think of Metis Reasoning as a generalisation of AlphaGo,” he said. AlphaGo, which beat Go champion Lee Sedol in 2016, combined neural networks with a structured game-tree search that evaluated possible moves and outcomes — including the famous Move 37, the shoulder-hit in game two that looked like a glitch and turned out to be genius. Metis wants that style of deliberative search, generalised to open-ended problems in science, medicine and robotics.
That, at least, is the argument. Whether it survives contact with real engineering is the entire question.
The anti-scaling pitch
If the architecture is the thesis, the economics are the provocation. Co-founder Piers Ingram told Tech.eu that Metis’s architecture was “much much more compute-efficient” than frontier models, and was blunt about the direction the industry’s giants are heading: “We are very much not going in this direction of ever larger models and ever larger training budgets.”
Metis claims it has already achieved “encouraging” results on frontier reasoning problems at a fraction of the cost of frontier models. Treat that with the scepticism it deserves: the company has published no independent benchmarks, and Pathfounders notes that none have been demonstrated. What Metis does claim is a working prototype that, in Graepel’s words, “can reason about PhD-level scientific problems” — tested, he says, by scientists connected to the founders’ biotechnology ventures. He expects an “industry-hardened” system to be available for external testing within three to six months.
The deeper cut is about trust, not cost. Graepel argues that current models cannot reliably explain how they reached a conclusion, even when they generate elaborate chains of thought. “More scaling will never give you an auditable reasoning trace that you can actually trust,” he told Pathfounders. His point about AI agents is sharper still: because we only ever see chain-of-thought traces that don’t reveal why decisions were really made, “we cannot really give them reliable guardrails.”
Ingram framed the stakes for the industries Metis wants to sell into: “When you think about higher-stakes cases in engineering or science, or medical or legal reasoning, the arguments that the models come up with basically can’t be trusted. I don’t think that’s a good foundation for what we want our AI to be.”
The people and the money
Graepel is not doing this alone, and the founding bench is conspicuously built for the industries named in the pitch. Alongside Graepel — who also holds the Chair of Machine Learning at UCL — the founders are Piers Ingram, a serial technology and biotech entrepreneur whose ventures have raised more than $350 million; Michael Warren, a robotics and autonomous systems specialist 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. That is a team assembled to sell into defence-adjacent, regulated, safety-critical buyers — the exact customers an auditable AI would appeal to.
Ingram told Tech.eu the pair met at a Cambridge barbecue, where the conversation turned — as these things apparently do — to Turing machines. It is a small detail, but it tells you something about the culture being built here: this is a company founded on first-principles argument, not on a go-to-market spreadsheet.
On money, both founders are careful. Bloomberg reported in September that Graepel was seeking tens of millions of dollars from a small group of initial backers, with talks described as preliminary. Graepel declined to discuss financing in the Pathfounders interview but confirmed fundraising discussions are ongoing; Ingram wouldn’t comment on the reports either, but said: “We have been able to fund the company more than adequately to date.” Tech.eu notes it is unclear when Metis will publicly launch its first AI model.
Graepel’s ambitions, at least rhetorically, go far beyond a niche product for engineers: “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.”
London's AI moment — and why America should care
Metis is the latest — and arguably the most philosophically pointed — in a wave of DeepMind-alumni startups now defining London’s AI moment. Graepel’s former AlphaGo colleague David Silver raised $1.1 billion at a $5.1 billion valuation in April for Ineffable Intelligence, with Sequoia and Lightspeed co-leading and Nvidia and Google participating. Richard Socher has co-founded Recursive Superintelligence in the city. OpenAI co-founder Ilya Sutskever, meanwhile, has raised billions of his own. Investors, as Tech.eu puts it, are pouring millions into alternatives to LLMs — and the DeepMind diaspora is where much of that money is landing.
That exodus is itself becoming a political story. The talent drain from DeepMind is unfolding against the political firestorm over DeepMind’s year-long non-compete “paid exile”, which has drawn fire from politicians including Andy Burnham — a reminder that London’s AI boom and its labour battles are two sides of the same coin.
For an American reader, the geography is almost beside the point. What matters is the argument, because it is aimed directly at the business model of the US labs. OpenAI, Anthropic and Google DeepMind have all converged on the same playbook: scale the architecture, extend the chain of thought, spend more. Graepel’s claim is that this playbook has a structural ceiling — that you cannot spend your way to an auditable machine, and that the industries where AI’s economic value is largest (robotics, engineering, medicine, science) are precisely the ones where unauditable systems cannot be deployed. He told Sifted, in an interview reported last month: “We need to go back to the architecture and fundamentally redesign it so that it does proper reasoning.”
He is not entirely alone in doubting the consensus. The scepticism about what LLMs can really do has been gathering from unexpected quarters — including Yann LeCun’s argument that AI takeover fears are science fiction, part of a broader establishment pushback against both the hype and the doom. Metis is that pushback’s most aggressive commercial expression yet: not a critique, but a company.
Our take: the reasoning insurgents have the right target. The odds are still brutal.
Let’s be blunt about what Metis is up against. Neuro-symbolic AI — the broad church Graepel admits this belongs to — has been colossally overpromised for the better part of forty years. Every decade produces a beautiful architecture for trustworthy machine reasoning, and every decade the statistical methods eat its lunch. A twelve-person startup with no benchmarks and no launched model is taking on three of the best-capitalised research organisations in history. On paper, this is a rounding error picking a fight with a hurricane.
But Graepel has chosen the one line of attack that doesn’t require him to win on capabilities at all. His argument isn’t “our model is smarter.” It’s “your model can’t be audited, and the customers with the deepest pockets won’t buy what they can’t audit.” That is a go-to-market argument disguised as a research thesis, and it’s aimed at the exact seam where the current boom is weakest: regulated industry, physical systems, anything where a wrong answer kills someone or voids a warranty. If Metis can demonstrate even a narrow, verifiable reasoning system that survives independent testing in the next six months, the conversation in boardrooms changes — not because the labs lose, but because “fluent but unverifiable” stops being an acceptable answer for the highest-value deployments.
Whether Graepel can convert AlphaGo’s lesson into a general machine for reliable thought is unproven, and the graveyard of elegant AI architectures should keep everyone humble. But he is, without serious dispute, one of the handful of people on earth most qualified to try. The industry spent a decade learning that scale produces fluency. The next decade’s argument — the one Metis just walked into the middle of — is whether fluency was ever the right goal.
What happens next
The white paper The Age of Reasoning is out today. Graepel says an industry-hardened system should be ready for external testing within three to six months — which means the first independent verdict on the “much much more compute-efficient” claim should arrive well before next summer. Until then, everything Metis says about performance is the company’s own word, and the honest position is to hold it at arm’s length.
Watch the hiring, too. If serious robotics and formal-methods researchers start joining a twelve-person London startup over offers from the frontier labs, that will tell you more than any white paper. And watch whether the labs respond: the moment OpenAI or Anthropic starts publishing about explicit knowledge representation and audit trails, you’ll know Graepel’s barbecue conversation landed.
Sources: Tech.eu’s launch-day coverage (October 20, 2026); the Pathfounders interview with Thore Graepel and its attached white paper The Age of Reasoning; Bloomberg’s September reporting on the fundraise, via syndicated relays; Graepel’s remarks to Sifted, via Tech Times; and Metis Reasoning’s launch site, checked on publication day.
Daniel Rao is TechPulse USA’s London-based AI correspondent, covering frontier research, model releases and the startup wars reshaping the industry. He writes about the business of intelligence — who’s building it, who’s funding it, and whether any of it actually works.
Comments
Post a Comment