The LLM Comments Are Not for You (danilafe.com)

Recent discussions in the AI community have highlighted the ongoing issues with comments generated by Large Language Models (LLMs) in software development. These comments often appear verbose, contextually disconnected, and laden with invented terms, raising concerns about their utility. Despite improvements in model performance as measured by benchmarks, these criticisms point to a deeper problem: the training methodologies for modern LLMs, particularly reinforcement learning (RL) techniques, prioritize outcome-based behaviors over concise communication. As a result, models tend to generate comments meant for agent-to-agent communication rather than for human developers. This trend has significant implications for the AI/ML community, particularly as LLMs are integrated into more complex areas of coding. The verbosity of LLM comments, while possibly beneficial for machine interactions, can create a disconnect from human users, complicating the code review and debugging processes. The author argues that correcting these behaviors without undermining the models’ effectiveness may not be straightforward, as their training favors lengthy, non-human-centric comments. This raises critical questions about how LLM-generated text will evolve within real-world software engineering contexts and whether existing human practices can balance or mitigate these emergent behaviors for better collaboration between humans and AI.