but of course AI works great for coding! if you apply this list of proper engineering practices your manager won't let you apply because the point is to go faster, not correctly
no you can't see my allegedly wonderfully engineered AI code
0 Boosts
0 Likes
@davidgerard Yes. The kind of super specific engineering practices that are so commonplace that in my nearly 3 decades in the field I've never seen them. Like, a proper specification instead of having to second guess from some vague description.
And don't forget the preemptive list of please don't make such-and-such mistakes, try to add some security and please don't delete my data, and never being quite sure you've nailed it down enough to stop it from going interesting places.
Yeah, nah.
@davidgerard I tried a spot of llm coding (well, troubleshooting) the other day.
"$problem"
"Oh yeah, do $this"
"That option doesn't exist"
"Oh yeah, in the latest version of $thing that's different"
(WTF do you think I'm doing other than latest version?)
"Still doesn't work"
"Ok yeah that library sucks, try $other"
"Also doesn't work"
"Ok, try $firstsuggestion"
Ok how about we don't go round again...
(Paraphrased, obvs.)
@davidgerard Listen. And understand. That LLM is out there. It can't be bargained with. It can't be reasoned with. It doesn't feel pity, or remorse, or fear. And it absolutely will not stop responding, ever, until you are deathly sick of it. -- Not Quite Kyle Reese
@davidgerard yeah the whole "it works if you just prompt it carefully and build harnesses and verification etc etc" argument misses the point that the *entire* claimed benefit of LLMs - speed - is negated, since you could just code it yourself and actually understand it AND not pay for a stupid tool that makes everyone in the room stupid and depressed.
@davidgerard I was trying to humor a friend so I just prompted the thing to write me a Roslyn analyzer (a simple no context single file problem with 4 short bullet points of acceptance criteria) and even after re-prompting it to split it into sub-methods I can't read this shit.
@davidgerard At my previous job they had a (human-made) 20 year old unreadable spaghetti codebase. It would regularly take me 3 hours of reading multiple files up a stack to figure out how to re-use a simple component. The revolutionary aspect of LLMs is being able to achieve this state in a month.
@davidgerard I'm not even kidding. It's uncanny how this style of no-effort procedural puke I saw there looks like LLM output.
@lidsel@mastodon.gamedev.place @davidgerard@circumstances.run that's probably like 50% of its training data after all, lol
@lidsel you must be prompting it wrong
@davidgerard One has to wonder what time I'm saving if it takes longer to figure out how to prompt it correctly than to just write the damned code yourself. 🤔
![[object Object]](https://media.mas.to/accounts/avatars/109/287/758/133/964/366/original/0f3099c0adfb36a0.png "[object Object]")
@davidgerard my incredible local LLM that works perfectly every time and isn’t slow as shit and bad at everything
no you can’t see it
@davidgerard it’s great! All you have to do as a programmer is the tasks that 70 years of the field have shown programmers are demonstrably horrible at and hate, and do them carefully and without error. If you do then the computer can do the stuff people are actually good at for them.
Wait, hold on…