Assessing the Risk of Advanced Reinforcement Learning Agents Causing Human Extinction

In this 58-minute lecture, Michael Cohen from UC Berkeley examines the potential existential risks posed by advanced reinforcement learning systems. Explore the methodologies for evaluating whether highly capable AI agents might pose extinction-level threats to humanity, with particular focus on safety guarantees and risk assessment frameworks. Learn about current research approaches to ensuring that reinforcement learning systems remain aligned with human values even as they become increasingly powerful and autonomous. Part of the Safety-Guaranteed LLMs series at the Simons Institute, this talk provides critical insights into one of the most pressing challenges in AI safety research.