Former DeepMind Researcher Launches Startup Aimed at Overcoming AI Hallucinations

Tore Grepel, a former DeepMind researcher, has launched a startup called Metis Reasoning aimed at resolving AI "hallucinations," the problem of generating information that contradicts facts. The company's goal is to apply the concept of "search," which has proven successful in chess and Go AI, to general-purpose reasoning.

What Grepel seeks to incorporate into the model is the design philosophy of "AlphaGo," the Go AI developed by DeepMind. AlphaGo possesses a mechanism for deriving the best move from vast options through "search." If this method could be leveraged for general reasoning beyond the realm of games, it would enable more systematic control over the process by which AI reaches conclusions—this is the starting point of Metis Reasoning.

As a concrete approach, Grepel envisions the implementation of an explicit mechanism he calls a "knowledge ledger." Current large language models take the structure of implicitly embedding knowledge obtained from training data into the model's internal parameters (weights). In contrast, the knowledge ledger proposes to explicitly record externally the evidence that serves as the basis for conclusions, making them verifiable later. However, at the present stage, neither concrete system implementation nor benchmarks (performance evaluation metrics) exist.

Hallucinations have been recognized as a significant practical challenge since generative AI became widely used. This problem—where models generate content that "sounds plausible" but contradicts facts—has made application difficult in fields requiring reliability, such as healthcare, law, and education. Metis Reasoning's approach can be seen as directly confronting the fundamental issue of "knowledge opacity" underlying this problem.

On the other hand, caution is warranted given that this remains at the conceptual stage. With no functional system in place and no objective evaluation having been conducted, the effectiveness of this approach remains unconfirmed. While search-based methods have succeeded in domains like Go, which have clear rules and evaluation criteria, reasoning with natural language involves ambiguous problem boundaries and diverse definitions of correct answers. Therefore, the extent to which the same methodology will apply remains uncertain.

Nevertheless, it is noteworthy that a researcher deeply involved in AlphaGo's development is attempting to repurpose that expertise toward hallucination mitigation. This direction of making reasoning transparent and verifiable through explicit evidence shares problem consciousness with existing approaches like RAG (Retrieval-Augmented Generation) and may play a role in the ongoing discussion on improving AI reliability. The release of a functional system and the publication of benchmark results will serve as the next basis for judgment.