[–] whatthejesus 1 day ago
[–] Groxx 1 day ago
[–] busssard 1 day ago
"down to earth" and "hobby-project" does a good job. its still not perfect but better than the vanilla output. eg. https://github.com/busssard/OpenRing
[–] varispeed 1 day ago
[–] whatthejesus 1 day ago
This is further evidence of the need for small models doing specific things well.
Where I would say models fit in the process is establishing the general layout of documentation and iteratively refining things out. Interestingly, it’s the same for image generation too. But 9 out of 10 it creates incomplete work that needs refinement.
Go wide, then hone in, refine. There’s no such thing as getting AI to single shot creative tasks. Which is funny because everyone thought writers are on the chopping block initially, but that is unequivocally NOT the case.
Even technical tasks, agents doing it their own way yields garbage in some right. It’s just a question of how long it runs before it goes down a path of no return.
Human in the loop is the only way. And, to be honest, what’s the f*king rush anyway? We all need to slow down. Good ideas with unravel in time. :)
[–] 0x69420 1 day ago
like, if people are gonna be turning the crank to generate prose-like text from statistical models and calling it a day, aren't "subagents" a hip new thing nowadays? there do exist models that aren't overcooked to impressively bad long-tail in the pursuit of maximum coding benchmark scores. such other models generate prose-like text that is significantly less miserable to wade through.
if you're a vibecoder who's terminally committed to the bit, for god's sake, at least tell a local gemma to draft your readme or something. an occasional outright falsehood is less damaging to usability than the experience of reading feeling like trudging through a pool filled with treacle. (mind, both are a violation of trust, and a failure to deliver proof of work, and please stop, but at least if the falsehoods are comprehensible and identifiable there's light at the end of the tunnel)
[–] stuaxo 1 day ago
Edit: I found reading those examples in the most sanctimonious voice I could imagine seemed very plausible - maybe LLMs sound like they are up their own arse.
To the project maintainer: don’t ask people to read a terrible AI generated sales pitch novella that quantifies your AI generated port against every UI framework out there.
Here’s the obvious giveaways:
“ 4. The many ways in” “ Python, four ways”
“ tkinter: all 7,846 public names of tkinter and ttk. Nine real programs behave identically on the standard library and on BUI.”
“ Rust: the web platform, natively”
Ugh it’s excruciating. You’re better off with a readme that shows a few screenshots and installation instructions.