Code World Model - Building World Models for Computation

Learn about a revolutionary approach to neural models for code that goes beyond syntax to understand program execution and computational dynamics in this 17-minute conference talk by Jacob Kahn from FAIR Meta. Discover how the Code World Model (CWM) incorporates data from program execution to implicitly predict behavior while generating code, moving beyond traditional token-based learning that only captures code structure. Explore how this world-model paradigm enables true reasoning about programs by understanding execution flow and computational behavior rather than just syntactic patterns. Gain insights into how CWM opens new capabilities for AI-driven reasoning about code and establishes a foundation for future research and prototyping in AI-powered software systems, representing a significant shift from current neural code models that learn primarily from code sequences to models that understand the underlying computational processes.