i watch people open merge requests on github with 2,000+ lines of changes and an entirely LLM generated explanation and then i watch a maintainer for that repo reply back also with an entirely LLM generated response and can't tell if i'm rooting for the maintainer to fight fire with fire to defend his shit, or disappointed just seeing everyone come out worse from such behavior.
Timeline
Post
Remote status
@mischievoustomato anyone trying to merge 2,000 lines of changes in a single pr needs to calm the fuck down
Replies
10i mean, yeah, you could ask it to separate stuff, but if all parties are ok with using agents, which can analyze, run, test, verify on their own, is there much of an upside, provided all parties are ok with the results?
@mischievoustomato just because "agents" exist doesn't mean you get to throw away the basic concepts of version control and maintainability.
@7666 @mischievoustomato Ive been studying StGit to solve a very very specific branch management problem im having, downloaded it, discovered its completely fuckin busted on windows. So i had claude work to find the root cause until all the test cases passed.
Not im learning how to use the tool proper and how it works underneath so i can understand how to deal with edge cases.
I feel like the last sane man on earth sometimes
Not im learning how to use the tool proper and how it works underneath so i can understand how to deal with edge cases.
I feel like the last sane man on earth sometimes
the >vsc thing gets resolved with breaking it down, and maintainability... well, too subjective, no?
@mischievoustomato why does one break changes down into meaningful commits. why does one try to work maintainability into their code. why does one shoot for modularity, managing dependencies, creating functions to prevent re-implementing the same thing three times over three different ways in three different spots.
if you can answer that question you should see the problem. if you can't, you're not qualified to be doing this stuff and will make garbage.
if you can answer that question you should see the problem. if you can't, you're not qualified to be doing this stuff and will make garbage.
it makes things easier to manage, roll back, contain, et cetera (it's 1am). But again, I reckon LLMs trivialize a lot of this.
@mischievoustomato they don't unless you deliberately ask them to. they don't know when to stop and think unless you know when to ask them to.
the problem is LLMs have no opinions. they'll just do "industry standard" things based on the gamut of what they've been trained on. that gamut is generally sort of shit, and it magnifies the operator's shit for a shit sandwich. that's how you get 2,000 line commits.
some of the best software engineers i've met are very opinionated people doing things solidly differently from others because they've been through the trenches and stubbed their toe a billion times on what not to do and their way ends up saving everyone time and effort even if it's a bit baroque. i'm just as opinionated in the same way. if i can channel that opinionated behavior into LLM operation then it *can* work out fine. but you gotta stub your toe before you know why you shouldn't do X or Y.
the problem is LLMs have no opinions. they'll just do "industry standard" things based on the gamut of what they've been trained on. that gamut is generally sort of shit, and it magnifies the operator's shit for a shit sandwich. that's how you get 2,000 line commits.
some of the best software engineers i've met are very opinionated people doing things solidly differently from others because they've been through the trenches and stubbed their toe a billion times on what not to do and their way ends up saving everyone time and effort even if it's a bit baroque. i'm just as opinionated in the same way. if i can channel that opinionated behavior into LLM operation then it *can* work out fine. but you gotta stub your toe before you know why you shouldn't do X or Y.
@7666 @mischievoustomato modern agents have the concept of a memory and sometimes it will be like "we tried it this way five times and every time it didn't work right away so now we're doing it this other way we've seen work the first time" its not perfect but
@sun @7666 @mischievoustomato i used qwen 3.5 and it remembers lot than keeps freaking out because i said things
YAW SO MISRUBLE N RAYCISSSS 