AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
AI Model Evaluation for Security and Safety Measures
The chapter delves into the development of an AI model evaluation metric for potential malicious use scenarios like designing chemical, bio weapons, and cyber attacks. It focuses on creating a benchmark to assess models' capability to unlearn dangerous knowledge, highlighting the challenges of balancing evaluation procedures and risks. The chapter also discusses a government commissioned report warning of extinction-level threats from advanced AI and emphasizes the importance of addressing risks through safety measures and global cooperation.