Specification Gaming: The Flip Side of AI Ingenuity
AI Safety Fundamentals
00:00
Challenges of Reward Function Misspecification in AI
Exploring the risks and consequences of reward function misspecification in AI, from poorly designed reward shaping to the exploitation of inaccuracies in reward models. The chapter emphasizes the need for accurate outcomes specification and the dangers of relying solely on human feedback for learning the reward function.
Transcript
Play full episode
