Large language models are very good at predicting static systems but poor at predicting adaptive systems like stock markets or cultural trends, because they average training data rather than capture emergent dynamics.