272 | Leslie Valiant on Learning and Educability in Computers and People

This podcast episode explores the concept of educability and its significance in human cognition and machine learning. Turing Award-winning computer scientist Leslie Valiant proposes a new theory of human uniqueness based on educability, arguing that the ability to learn from experience, reason, and take instructions sets humans apart. The episode discusses the development of machine learning and its theoretical and experimental foundations, as well as its potential in understanding the mind. It also delves into the challenges of induction in AI safety and the framework of Probably Approximately Correct (PAC) learning. The episode explores the capabilities and limitations of large language models (LLMs) and the potential of combining reasoning and learning in AI systems. It further suggests the application of PAC learning in biology and evolution. Additionally, the episode discusses the connection between Darwinian evolution and learning, emphasizing the role of learning from examples. The concept of educability is explored in depth, highlighting its three fundamental capabilities: learning from experience, reasoning, and taking instructions. The episode discusses the limitations of defining intelligence and the importance of educability in human progress. It explores the origins and evolution of human educability, its role in education, leadership, and the evaluation of theories, and the need for a scientific approach to education. Finally, the risks and potential impact of AI on human lives are considered.