Former Deepmind researcher thinks AlphaGo's playbook can cure AI hallucinations
decoder ·
Former DeepMind researcher Thore Graepel has founded Metis Reasoning to address AI hallucinations by applying AlphaGo’s strategic playbook to general reasoning. The proposed system separates knowledge from model weights, storing it in an external ledger that is independently verified against evidence before acceptance. This approach aims to provide source-backed explanations and allow for the correction of errors without retraining the entire model. Graepel argues that simply increasing data and compute does not solve the fundamental issue of verifiable reasoning in large language models. While the method shows promise for robotics and edge devices, the current framework remains theoretical with no working system or benchmarks yet available. The article highlights a broader trend of DeepMind veterans leaving Google to pursue independent ventures focused on reliable AI. Despite the lack of immediate practical implementation, the company is raising significant funding to develop this new paradigm.
decoder