前DeepMind研究员拟用AlphaGo思路治理AI幻觉
原标题：Former Deepmind researcher thinks AlphaGo's playbook can cure AI hallucinations
AI 摘要
前DeepMind研究员Thore Graepel创立Metis Reasoning，提出借鉴AlphaGo的搜索与验证思路来减少AI幻觉：将知识从模型权重中分离，存入外部“知识账本”，由独立模块依据计算、形式证明或经验证据核验每个假设，并据此重新评估依赖该假设的结论。该方法目前仅为理论框架，尚无可用系统、基准测试或与现有推理模型的对比。Graepel称正为Metis融资数千万美元，后续融资估值更高。
正文节选
Former Deepmind researcher thinks AlphaGo's playbook can cure AI hallucinations Key Points - Former Deepmind researcher Thore Graepel founded Metis Reasoning to make AI systems more reliable and verifiable, using AlphaGo's approach as a blueprint. - The method separates knowledge from model weights and stores it in an external ledger, where an independent module checks every assumption against evidence. - The approach aims to cut hallucinations and provide source-backed explanations, but so far