Former Deepmind researcher Thore Graepel has founded Metis Reasoning to address AI hallucinations by applying search-based logic inspired by AlphaGo. The company aims to replace current language models, which prioritize plausible text over factual accuracy, with systems that track knowledge and verify conclusions against evidence.
The proposed architecture uses an explicit knowledge ledger to record known facts, confidence levels, and open questions. A separate module checks new information against formal proofs or empirical data. If an assumption changes, the system re-evaluates all dependent conclusions, aiming to provide verifiable answers with citations.
Metis Reasoning suggests that this modular approach could enable smaller models to run on edge devices and robotics. By breaking reasoning into small, verifiable steps, the system avoids the need for full retraining when knowledge is updated or deleted, addressing issues that currently limit AI adoption in fields like finance and medicine.
The project is currently in the research phase, with no working system or benchmarks yet available. Graepel is reportedly raising a funding round in the low tens of millions of dollars, with plans for a subsequent round of several hundred million dollars at a higher valuation.
Graepel, who previously worked on AlphaGo and AlphaZero, joins a trend of former Deepmind researchers launching independent startups. Other notable departures include the founders of Ineffable Intelligence and Recursive, the latter of which is valued at $4.65 billion.
