CONCEPT
World Model Architecture
Yann LeCun’s proposed foundation for genuine machine intelligence: an internal simulation of how reality behaves that would allow a system to predict, plan, and reason about the consequences of its actions rather than merely completing patterns in text.
The world model is the missing component that separates pattern-matching from understanding, in
Yann LeCun’s diagnosis of where current AI systems fall short. A system with a world model carries within itself a working simulation of how its environment behaves: given a current state, the model predicts what will happen next, and given an intended action, it predicts how the state would change. This capacity for internal prediction enables planning—the ability to imagine the consequences of an action before taking it, to search through possible futures for the one that best serves a goal, to reason about cause and effect rather than retrieve text that sounds like causal reasoning.
Large language models can describe how a dropped glass breaks without modeling it—they have absorbed vast quantities of text about falling objects and can generate statistically plausible sentences about them—but they do not simulate the falling glass, and the difference is not academic. A system with a world