Mastodon Feed
Boosted by glyph ("Glyph"):
datarama@hachyderm.io wrote:
Early on after my workplace started requiring LLM usage, I made the decision that I wasn't willing to compromise on quality. I am a programmer. I want to write good code. I do not want to throw slop at either my colleagues or the outside world, ever. And I still haven't: I've stood firm on writing everything I actually commit myself. (I have *reviewed* thousands and thousands of lines of slop at this point; some of it I'm sure I was the first human being to ever lay eyes on).
I'm pretty sure I've actually become *slower* with LLM assistance. That's because of aforementioned observation about trivial tasks - an LLM can bring the time-and-tedium cost of those from "I'll do that tomorrow" down to "I'll just do it now" - which, in practice, often means that the time spent goes from zero to nonzero (because, in the words of a Black Sabbath lyric*: "live for today, tomorrow never comes"). But this isn't bad: It means more edge cases get handled. And because I insist on writing the code myself, sometimes I discover a pattern that I can fruitfully make an abstraction out of. At least that's how I've tried to use LLMs without sacrificing quality. (And once again, I've found that comparatively tiny and "stupid" models are just as good at this sort of thing as the big ones. Ironically, the heavily engagement-tuned big ones are *worse*). But it's a pretty intense way of working: I find myself "flitting around" dealing with large amounts of trivia, which is cognitively more taxing.
Anyway, I've talked to my manager, and he's OK with me not going all-in on "agentic" development, so it seems I can retain my sanity for a bit longer.
*) "Die Young". I will die on this hill: The Dio era was not as bad as everyone pretends.