Artificial Intelligence
Thore Graepel pitches AlphaGo-style AI reasoning
Thore Graepel's Metis Reasoning proposed an external knowledge-ledger approach to AI reasoning, saying it could reduce hallucinations and produce checkable explanations, but the company described it as research without a working system.
- Thore Graepel founded Metis Reasoning and outlined its approach at Nvidia's GTC conference.
- Metis said the system would keep knowledge outside model weights and verify assumptions against evidence in a separate module.
- The company said the design was inspired by AlphaGo and should reduce hallucinations while adding source citations.
- The white paper said the work was only a research programme, with no working system, benchmarks or comparisons with current reasoning models.
- Bloomberg reported that Graepel was raising a low tens-of-millions funding round for Metis, with a larger follow-on under discussion.
Thore Graepel founded Metis Reasoning and presented its plan at Nvidia's GTC conference. The company said it wanted AI systems to keep track of what they knew and to check each conclusion against evidence.
Metis said its design treated current language models as only one part of AlphaGo's method. It proposed an external knowledge ledger, with a separate module that would test assumptions against calculations, formal proofs or empirical evidence.
The company said two neural networks would suggest and assess moves, but verification rules would remain fixed. It said the setup should lead to fewer hallucinations, source-backed explanations and knowledge that could be changed without retraining.
The white paper described the work as a research programme and said there was no working system, no benchmarks and no comparison with existing reasoning models. Bloomberg reported that Graepel was seeking a low tens-of-millions funding round for Metis, with a larger follow-on already being planned.
Named in this story
People
- Thore Graepel
- founded Metis Reasoning and presented its approach
- Lee Sedol
- the Go professional beaten by AlphaGo in 2016
Companies
- Metis Reasoning
- the startup proposing an external knowledge-ledger system
- Nvidia
- hosted the GTC conference where Graepel spoke
- Altos Labs
- an employer Graepel had previously joined
Organisations
- Google Deepmind
- Graepel's former employer before he left over the summer
- Microsoft Research
- where Graepel previously helped build TrueSkill
- University College London
- where Graepel held the machine learning chair
- Bloomberg
- reported that Graepel was raising funding for Metis
Products and systems
- AlphaGo
- the system whose search method Metis used as a model
- TrueSkill
- the matchmaking system Graepel helped build
- Xbox Live
- the platform behind TrueSkill's matchmaking use
How the source tells it
The article was mostly explanatory, but it used a promotional, speculative register around a startup's untested proposal.
- hype the headline and repeated claims cast an unproven research programme as a potential fix for a major AI problem
- speculation presented as fact the company's own paper was used to support wide claims about AI's limits and future usefulness despite no working system