Former DeepMind researcher Thore Graepel wants to apply in a startup Metis Reasoning the search principle used in AlphaGo to general reasoning tasks. The idea envisions a separate knowledge registry so that every conclusion can be checked against related evidence, rather than storing knowledge only in model weights.
At the moment, according to the available description from The Decoder, there is no working system or published benchmarks. Therefore, the practical effectiveness of the approach and its ability to reduce hallucinations are not yet proven.
To evaluate the idea, at least a working implementation and comparable tests of conclusion accuracy and verifiability of evidence will be required. The source is presented by a single independent publication and is based on metadata, so the conclusions should be considered preliminary.
editorial commentary
Why it matters
If the concept is implemented, the next noticeable signal will be the publication of a working prototype and independent benchmarks with verifiable conclusions. Until then, there remains uncertainty about the accuracy, speed and practical utility of the approach.