Understanding the Abilities of AI Systems - Memorization, Generalization, and Points in Between

Watch a 44-minute lecture from Yale University's Tom McCoy exploring how Large Language Models (LLMs) develop their capabilities through the lens of generalization and training data analysis. Delve into the complex relationship between shallow heuristics and deeper abstractions in AI systems, examining specific cases where LLMs demonstrate both impressive abilities and unexpected limitations. Learn how these models can generate novel sentence structures absent from their training data while simultaneously struggling with basic algorithmic tasks like cipher decoding. Gain insights into the fundamental connection between an AI system's training process and its resulting capabilities, helping to better understand the true nature and limitations of artificial intelligence.