brea on Nostr: #[0] #[1] there's a hard to maintain balance regarding what is worth relying on it ...

bless (npub1nxn…wn2k) pistolero (npub1ch8…sw60) there's a hard to maintain balance regarding what is worth relying on it for in my very very limited experience where if you tilt into trusting it too much it results in a net increase of the time and effort you're expending because now you have to untangle the previous bullshit it gave you and come up with a prompt that will not make it hallucinate libraries or methods (and then you get sunk cost fallacie'd) but where tilting back juuuust right will in fact manage to avoid you a lot of boring grunt work

what it's generally okay for:
- there's a stupid dumb trivial error in this code and you know it's something incredibly minor, but your eyes are really tired and unfocused because you've been at this for hours. it will catch that error. it will also give you extra suggestions you didn't ask for because it thinks too highly of itself, but, whatever, you can just alt tab and let it ramble after you see where you forgot to capitalise that variable

- you get a bizarre runtime error, googling it is giving you people not quiiiite dealing with the same problem you're having, or maybe it's a relatively obscure error for a relatively obscure program (when i was troubleshooting pleroma i couldn't find many hints anywhere); it's generally not bad at giving you a lead to pull on if the error text itself isn't particularly informative

- it's not bad at suggesting libraries to do things, at least in web stuff? i wanted to implement a drag and drop image area and it suggested me a perfectly good library. this is inherently devalued advice the more we get away from the 2021 cutoff point, but, hey. it'll also likely know why something's not working when something in the library fails without giving you any feedback (this is hedged usefulness: if the library has good documentation it's not really anything you couldn't have found out yourself, but, yknow, big if)

but as to writing actual code itself, like, anything that you haven't already done 90% of the work for? nnnnnnnnooooot really?

it's... theoretically doable? but the problem is that you have to rest on the assumption that the amount of time and effort you spend on getting better at prompting it is meaningfully lesser than the amount of time and effort you'll end up save from not writing and looking over stuff like usual - and there's a soft cap on how good results you can get that just, comes much sooner than you tend to think so

idk you could ask it to implement you a hybrid quick-insertion sort loop, that sort of little snippet of code thing, but if you're really that concerned about sort performance you're likely going to immediately get annoyed at it doing it not quite as efficiently as it could have

it can be useful, there's just such a narrow slice where it will pay off, and it's hard to consistently rest on that slice, even if you do manage to get there sometimes