#170 Richard Sutton on Pursuing AGI Through Reinforcement Learning

This podcast features an interview with Richard Sutton, a leading AI researcher, focusing on his "Alberta Plan"—a five-year research agenda aiming to build embodied AI agents capable of learning and planning through environmental interaction. The discussion covers the crucial role of increasing computational power in AI advancements, contrasting the limitations of current large language models with the Alberta Plan's focus on continual learning, world modeling, and the development of more efficient reinforcement learning algorithms. Sutton highlights the importance of off-policy learning and the creation of "dynamic learning nets" with adaptive learning at multiple levels. The interview concludes with Sutton's perspective on the potential benefits and challenges of advanced AI, dismissing overly pessimistic "doomer" viewpoints. A key takeaway is Sutton's emphasis on building AI agents that understand and respect goals, unlike current large language models.