One of the most annoying quirks of LLMs for me is the insistence on, after being corrected, loudly noting what wasn't done. After getting it to change something stupid in code, it will leave a comment in the code boasting about how it doesn't do the stupid thing, which to future readers reads like an 'asbestos free!' label on a cereal box.
I’ve left so many PR reviews in the last few months that are essentially “this looks good but delete 80% of the comments.”
Lets assume one of two worlds - one where AI is constantly improving and one where the tooling is a dead end. In the first world the AI generated comments are no better than the commentary that would be generated in the next six months - in the second world the AI generated comments will obscure the human written comments.
In what world are AI generated comments actually value adds?
If the AI (or the person) ran into a non-obvious constraint, or some externality, or there’s a plan that just isn’t implemented yet- then the comment wouldn’t be able to be divined by the AI. But yeah just pointing the AI at something and saying “write me docs” will get you something at best equally good as what’s going to be available.
this. like I feel that even telling sonnet 5 to explicitly not add comments it still produces them even on suppa simple logic
Usually for me it writes a test to verify that the stupid thing isn’t happening.
One of the reasons this is there is the "psychology" of the LLM performs better with these kinds of affirmations. If you've made a mistake once you're likely to make it again, so leave a note.
This indeed pollutes the code with a bunch of nonsense. One of the things I have it maintain instead is a "findings" document with lessons learned and such.
> If you've made a mistake once you're likely to make it again, so leave a note.
In my experience the comments behave as a form of prompt injection. The llm makes a conceptual mistake, writes it in a comment, and now subsequent agents make the same mistake!
I have a contrarian take on this: if most people working on the codebase are using the same model, then these comments serve a very useful purpose: they tell the model's future self that despite its tendency towards this solution, it is incorrect and here's why. It encodes useful context right when the model needs it next to avoid making the same error.
I imagine this style of commenting is partially responsible for model improvement in SWE work, a trained trait through feedback, since it makes the comments actually useful. Not to you or me, but to an LLM.
Some of these errors are obvious to a human, that's why we find the comments unnecessary. But if everyone's using LLMs and said model tends to make this error, then they are very useful. It's like a comment you would leave for another human to help them avoid a footgun. It's just a LLM footgun, not a human one.
The problem is that the list of things one shouldn't do is infinite. The things one should do are narrow and get narrower along two independent axes of clarity/wisdom: system requirements and engineer seniority. LLMs overindex on any words given, so you only want to give them negative guidance around truly repeated, almost common-sense pitfalls. But in a large distributed system often the changes come from all different angles, and each agent will find it's own unique failure modes. Those comments will dilute critical context for diverse agents far more than they will help on average, at least for the systems I'm working on, given current claude code chattiness.
One way to clear some of the low-hanging slop is to just have a separate agent code reviewing and pruning any comments that don't stand on their own purely in the context of the diff, but it still doesn't catch it all.
I agree, but that information should be in the commit, which the model also has access to. Of course it's more token intensive to look through commits, but "good" models might know when some more research about a particular bit of code they find weird would be a good idea. Thoughts?
It doesn't necessarily have to go through commits, it can do bare git blame to see if something stands out. But otherwise agreed.
If LLMs are any good, like many of you claim, they won’t need these comments. They will ingest the code and know what was done. They won’t need comments about what wasn’t done.