I fear that LLM-optimal code structure is different from human-optimal code structure, and people are starting to optimize for the former rather than the latter.