TL;DR

Thore Graepel, a co-creator of DeepMind’s Go-playing system AlphaGo, launched a London AI company called Metis Reasoning on 20 October, Tech.eu reports. He left Google DeepMind in the summer. The 12-person start-up says it is pursuing reasoning that can be checked and that needs far less compute than the frontier models from Anthropic and OpenAI, and is reported to be seeking a first round in the tens of millions of dollars.

The pitch

Metis accepts that large language models now produce fluent answers and handle some multi-step tasks, but argues they reason unreliably and leave no audit trail. That, the company says, is what holds back adoption in fields where the work has to be verified. It wants to carry an architectural lesson from AlphaGo over to reasoning, and sees robotics and engineering as natural fits.

Co-founder Piers Ingram, who has started several tech and biotech companies, put the problem bluntly: in engineering, science, medicine or law, he said, the arguments current models produce “basically can’t be trusted”. He described the Metis design as “much much more compute-efficient” than frontier systems and said the team is “very much not going in this direction of ever larger models and ever larger training budgets”.

The other founders are Michael Warren, a robotics and autonomous systems specialist with Boeing experience, and David Berman, Cambridge Consultants’ head of AI, who has advised the UK government and NATO.

What is still unknown

There is little to judge yet. Ingram declined to discuss the reported funding round beyond saying the company has been funded “more than adequately to date”, and claims of “encouraging” results on frontier reasoning problems come with no figures attached. Tech.eu says there is no date yet for a first public model.

Looking forward

Metis joins a cluster of London labs started by senior researchers, such as David Silver’s Ineffable Intelligence (Silver also worked on AlphaGo) and Recursive Superintelligence, where Richard Socher is a co-founder. In our view, the compute-efficiency claim is the one that matters most for UK businesses. If reasoning systems can deliver checkable answers without hyperscale training budgets, more organisations could use them without frontier-scale spending. Until Metis publishes results, though, that remains a promise rather than a finding.