from some of the engineers I've debated this over, I think some of them have just dug their heels in at this point and have decided they're never going to use LLM tools period, and are just clinging to the original arguments without really examining the reality of the situation. In particular this "The LLM is going to hallucinate subtle bugs I can't catch" one. The idea that LLMs make subtle mistakes that are somehow more subtle, insidious and uncatchable compared to any random 25 pull requests you get from humans is simply ridiculous. The LLM makes mistakes that stick out to you like a sore thumb, because they're not your mistakes. The hardest mistakes to catch are your own, because your thinking patterns are what made them in the first place.
The biggest problem with LLMs for code that is ongoing is that they have no ability to express low confidence in solutions where they don't really have an answer, instead they just hallucinate things. Claude will write ten great bash lines for you but then on the eleventh it will completely hallucinate an option on some linux utility you hardly have time to care about, where the correct answer is "these tools don't actually do that and I dont have an easy answer for how you could do that". At this point I am very keen to notice when Claude gets itself into an endless ongoing loop of thought that I'm going about something the wrong way. Someone less experienced would have a very hard time recognizing the difference.
> The idea that LLMs make subtle mistakes that are somehow more subtle, insidious and uncatchable compared to any random 25 pull requests you get from humans is simply ridiculous.
This is plainly true, and you are just angry that you don't have a rebuttal
I didnt say the LLM does not make mistakes, I said the idea that a reviewer is going to miss them at some rate that is any different from mistakes a human would make, is ridiculous.
Missing in these discussions is what kinds of code people are talking about. Clearly if we're talking about a dense, highly mathematical algorithm, I would not have an LLM anywhere near that. We are talking about day-to-day boilerplate / plumbing stuff. The vast majority of boring grunt work that is not intellectually stimulating. If your job is all Carnegie-Mellon level PHD algorithm work, then good for you.
edit: I get that it looks like you made this account four days ago to troll HN on AI stuff. I get it, I have a bit of a mission here to pointedly oppose the entrenched culture (namely the extreme right wing elements of it). But your trolling is careless and repetitive enough that it looks like.....is this an LLM account instructed to troll HN users on LLM use ? funny
The biggest problem with LLMs for code that is ongoing is that they have no ability to express low confidence in solutions where they don't really have an answer, instead they just hallucinate things. Claude will write ten great bash lines for you but then on the eleventh it will completely hallucinate an option on some linux utility you hardly have time to care about, where the correct answer is "these tools don't actually do that and I dont have an easy answer for how you could do that". At this point I am very keen to notice when Claude gets itself into an endless ongoing loop of thought that I'm going about something the wrong way. Someone less experienced would have a very hard time recognizing the difference.