Metis Reasoning 在倫敦啟動，主張採用更具計算效率的 AI 路線

A London-based AI company co-founded by a former Google DeepMind research lead and AlphaGo co-creator is launching today, pursuing a “compute-efficient” alternative path to Anthropic and OpenAI’s frontier AI models in progressing AI.

Thore Graepel, who left Google DeepMind this summer, is launching AI company Metis Reasoning, which is understood to be initially raising tens of millions of dollars, marking the latest AI company springing up out of London.

Graepel, who is a co-creator of AlphaGo, an AI programme developed by Google DeepMind which made history by defeating human players at the Chinese board game of Go, and his co-founders have put together a 12-strong team.

The team is made up of researchers, mathematicians, robotics experts and tech entrepreneurs to pursue a different approach to AI reasoning, the process whereby an AI system works through a problem, connecting information, rather than producing an answer based on learning patterns.

The launch of Metis (named after the Greek goddess of wisdom), which believes its thesis will lend itself well to industries such as robotics and engineering, comes as investors pour millions into alternative AI systems to LLMs.

Metis’s founding team is Graepel; Piers Ingram, a serial technology and biotech entrepreneur who has raised more than $350m; Michael Warren, robotics and autonomous systems expert with experience at Boeing; and David Berman, head of AI at Cambridge Consultants and a UK government and NATO AI advisor.

Metis argues that LLMs have made significant progress and can often generate fluent answers, infer meaning, follow instructions and perform some multi-step tasks. But Metis argues that LLMs have a fundamental limitation in how they reason and can be unreliable and lack an audit trail.

Metis argues that this is a barrier to wider AI adoption, saying its adoption in many industries remains constrained by the difficulty of verifying its work.

Ingram, who met Graepel at a Cambridge barbecue where they chatted about Turing machines, says: “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.”

Metis’s thesis builds on Graepel's work at AlphaGo, saying that a similar architectural lesson should be applied to AI reasoning.

Metis argues its approach is more verifiable, robust and safer than current frontier models.

Asked why Anthropic and OpenAI were not pursuing a similar path to Metis, Ingram said: "I think they are thinking about it differently.”

Ingram would not comment on reports that Metis was raising tens of millions of dollars, but said: "We have been able to fund the company more than adequately to date.”

He said that Metis’s architecture was “much much more compute-efficient" than frontier models. He said: “We are very much not going in this direction of ever larger models and ever larger training budgets."

He said Metis had “achieved” some “encouraging” results in frontier reasoning problems at a fraction of the cost of frontier models.

It is unclear when Metis will publicly launch its first AI model.

Metis is joining other new London-based AI companies, including Ineffable Intelligence, founded by Graepel's former AlphaGo colleague David Silver; and Recursive Superintelligence, co-founded by high-profile German computer scientist Richard Socher.

Asked if there was any overlap between Metis and Ineffable and Recursive, Ingram said: “We obviously don’t know exactly what they are doing; they don’t know exactly what we are doing.” But added that he thought Metis’s approach was “really a very different, almost philosophical approach”.