This podcast episode explores the importance of uncertainty and exploration in reinforcement learning (RL). It discusses the challenge of bridging the gap between Bayesian and machine learning methods and highlights the potential benefits of combining these approaches. The concept of information-directed sampling is introduced as a framework for balancing exploration and exploitation in RL, providing more efficient learning compared to traditional methods. The episode also delves into the concept of joint prediction and its relevance in decision making, as well as the limitations of current approaches to uncertainty estimation in deep learning. Epistemic neural networks are presented as a more flexible and efficient approach to making joint predictions, and the EpiNet architecture is introduced as a solution for approximating Bayesian computation. The speaker emphasizes the importance of joint prediction in deep learning advancements and the applications of RL in real-world problems. The episode concludes with insights on the future of artificial intelligence and RL, emphasizing the need for AI systems to prioritize their own learning and align with human goals.