Ex-DeepMind’s Graepel Launches Metis Reasoning to Fix LLM Flaws
Thore Graepel and co-founders launch Metis Reasoning in London, aiming to replace untrustworthy LLM outputs with auditable, step-by-step reasoning systems.
Key points
- Thore Graepel, former Google DeepMind researcher, founded Metis Reasoning in London this summer.
- The company argues large language models cannot be trusted for high-stakes reasoning due to hallucinations.
- Metis proposes a system that records knowledge and uncertainty, allowing users to audit every step.
Thore Graepel, a former researcher at Google DeepMind known for his work on AlphaGo, has launched a new artificial intelligence company called Metis Reasoning in London. According to a report by The Next Web, Graepel departed Google DeepMind this summer to lead this new venture, which aims to address fundamental flaws in current large language models.
The company’s core thesis, outlined in a white paper titled “The Age of Reasoning,” is that large language models cannot be trusted to reason accurately. The founders intend to apply the technical lessons learned from building AlphaGo to the broader field of general reasoning. They argue that existing models lack the necessary transparency and reliability for critical applications, necessitating a new architectural approach that prioritizes verifiable logic over probabilistic prediction.
The Case Against Current Models
The white paper identifies three primary problems with current large language models: they hallucinate information, their reasoning processes are not auditable, and users cannot inspect the underlying knowledge because it is embedded within opaque model weights. Piers Ingram, a co-founder and technology entrepreneur who has raised more than $350 million, told Tech.eu that model arguments are untrustworthy in high-stakes fields like engineering, science, medicine, and law. The document states that no amount of scaling can solve these specific issues.
Metis Reasoning argues that the step-by-step “chain of thought” often generated by models does not accurately reflect how the system arrived at its answer. The white paper opens with the example of Move 37 from AlphaGo’s 2016 match against Lee Sedol, a move experts initially dismissed as a blunder. The founders contend that this move resulted from rigorous search and evaluation—what they define as reasoning—rather than simple prediction, demonstrating the power of explicit logical processes.
Who Is Behind Metis Reasoning
Graepel is joined by three other co-founders: Piers Ingram, a biotech and technology entrepreneur; Michael Warren, a robotics expert formerly with Boeing; and David Berman, head of AI at Cambridge Consultants. Berman also advises the UK government and NATO on artificial intelligence matters. According to Tech.eu, the team currently consists of twelve people. The white paper indicates that the company maintains offices in London, Singapore, and Los Angeles, signaling an international operational footprint from its inception.
What should you do?
Security and technology professionals should monitor developments from Metis Reasoning as it releases more details about its proposed architecture. Organizations relying on large language models for critical decision-making should continue to implement rigorous human-in-the-loop verification processes. Since no patch or fix exists for the inherent reasoning flaws in current models, users must treat model outputs as suggestions rather than facts until auditable alternatives become widely available.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Metis Reasoning
- Audit current large language model outputs for hallucinations in high-stakes workflows.
- Implement manual verification steps for any AI-generated content used in legal, medical, or engineering contexts.
- Monitor vendor announcements for new reasoning architectures that offer step-by-step auditability.
- Review internal policies to ensure no AI system makes autonomous decisions without human oversight.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Who founded Metis Reasoning?
Metis Reasoning was founded by Thore Graepel, Piers Ingram, Michael Warren, and David Berman.
What is the main problem with large language models according to Metis?
The founders argue that large language models hallucinate, cannot be audited, and hide their knowledge source within model weights.
How does Metis plan to improve AI reasoning?
Metis plans to use a system that keeps an explicit record of knowledge and uncertainty, allowing for step-by-step auditing and updates without retraining.
