I wonder if th style shift has anything to do with training for conversation (ie. tuning models to respond well in a chat situation)?