The further you push language models from their natural output patterns, the worse they perform. Understanding why reveals fundamental truths about how these systems actually work.
