| Quote: "This line of thinking boils down to, “Humans are valuable if they produce high-quality output.” This argument dangerously depends on the existing-but-narrowing human-AI capability gap. The gap certainly existed in the past (2023-era ChatGPT). It may still exist now. I do not know if it will hold in the future." I still don't agree with this, because even the best engineers are producing poor output from AI, unless they spend a reasonable amount of time refactoring and cleaning up the output. To expect a non-engineer to have the same wisdom to architect and structure their output in a sustainable way that will stand the test of time (and inevitable future changes), they're just fooling themselves. For anyone who's been using LLMs long enough for development, we all know how sloppy the output is unless you have a mountain of formatting and organizational rules in place, and even then, it's anyone guess how well LLMs can actually follow them as a project scales. Yes, value the human first, but there are strong reasons we've always emphasized mature patterns and idioms in codebases that help them remain maintainable so they can scale and transform with as little friction as possible, and without reaching a point where nobody has the stomach to keep working on it. Just because we can try to have AI unwind the hairy mess, doesn't mean it's any easier for AI to navigate it than we do. It's a liability that the AI mistakes one naming conflict for another, when the context rot hits the fan — and after months of tossing unreviewed PRs over the finish line without a single thought, it's only a matter of time before your non-engineers are the worst offenders for introducing bugs and eroding the quality. It doesn't matter how good the models appear to get (it truly is a facade), without judgment and wisdom, they still don't produce nearly the same quality as an experienced engineer, unless they're heavily steered. Do they catch more potential bugs and edge cases than a human might? Definitely, but it's going to vomit its solutions in all the weirdest places doing it, and now suddenly you've got a 100k LOC pull request to (not) read. All I can say is, good luck! |