Yann LeCun Forecasts World Models to Replace LLMs Within Five Years

AI pioneer Yann LeCun predicts "world models" will dominate AI architecture within three to five years, making current large language models (LLMs) obsolete. This shift represents a move towards AI systems that understand the physical world, not just language.

Large language models (LLMs) currently define much of public AI discourse, demonstrating impressive abilities in generating text and code. However, a significant architectural shift is now emerging, moving beyond these word-predicting systems. This new paradigm focuses on "world models," AI systems designed to learn and simulate the physical world. These models aim to develop a "common sense" understanding of how objects interact and environments behave, a capability LLMs lack.

Yann LeCun, a leading figure in AI research, forecasts that world models will become the dominant AI architecture within three to five years. He states that current LLMs, which largely rely on pattern recognition from vast datasets, will become obsolete for advanced AI applications. LeCun demonstrated this conviction by leaving his role as Meta's chief AI scientist to establish his own organization, dedicated to advancing world models. This shift emphasizes AI systems that grasp the underlying physics and dynamics of an environment. Early examples of this technology are already emerging. Runway's GWM-1 (General World Model 1) illustrates this progression by generating video frame-by-frame. This model simulates real-world physics, marking a step toward fully physics-based world models.

World models represent a significant divergence from current LLM capabilities. While LLMs predict the next word in a sequence, they operate without inherent understanding of physical reality or causal links. World models, conversely, learn how environments function, allowing them to simulate outcomes, reason about constraints, and adapt to new situations. This capability extends to understanding object permanence and cause and effect, capacities absent in language-only models. This foundational understanding is considered a crucial step toward artificial general intelligence (AGI), where AI systems can learn and apply knowledge across diverse, unfamiliar situations. The future will show how effectively these new models can transition from simulating environments to making complex, real-world decisions.