At least my experience is that a vibe coded app gets to ‘early alpha’ then stays there. If AI is augmenting a developer who is otherwise competent and they don’t let it get ahead of their own understanding, it can work out though not as ‘miraculously fast’ as this. If AI can oneshot an application then it can also work out. If a project vibes itself in a week or two to ‘looks like what it should be but isn’t good enough’, then thus far every one I see is doomed to sit there, because the AI can’t finish it and if you let the GenAI produce that much code, it’s generally not serviceable even for very capable developers.
I spent a lot of time last week basically peeling back an overly vibed attempt at an application by a colleague. It was too flaky and incomplete and deleting large sections of it was frankly more practical than trying to rework it. It had gotten to the point where every attempt to try to have GenAI fix it one way would just break something else. That project was, essentially a fairly small command line utility, I fear to think what something the level of a Photoshop knockoff would be like.
I found it doing quite decently when using it with strongly opinionated web frameworks and some rules. After all, most web projects are mostly CRUD anyway. For desktop apps it seems mixed, probably because there aren’t many frameworks that give structure. It did however throw together a working app for me so i can map my mouse buttons, but it’s in golang. I’m more of a python / php guy, so i can’t really tell if it’s coded correctly. It didn’t delete my computer yet, though.
With this photoshop clone however, i’m surprised it starts at all. Last time i’ve tested something more complex for desktop, it failed spectacularly. So i’m quite curious how far they can push this thing. What i don’t like, is that GenAI stuff they have to generate images. Coding is one thing, but using it for “automated art” is not so great…
Yeah, there are scenarios that it can ‘one shot’, either very straightforward variations of common themes or a use case that is comprehensively instrumented in a way to “know” if it is correct or not automatically, and having the solution iterate against that criteria evaluation until all criteria pass (though in the latter case, I have seen it change the criteria deciding the criteria was a mistake instead of fixing a problem).
If you spend enough tokens, let it rebuild and test on its own, and let it churn and churn, then I’m not too surprised it could make this launch and be superficially intact. The real pain starts when trying to close the gaps the project has in this sort of scenario.
I will say that GenAI makes golang a lot less tedious. Lot’s of obvious but required boilerplate to force the developer to abide by consistent structure. Though I still prefer reading python code since I don’t have to peruse a ton of boilerplate there.
jj4211@lemmy.world 3 weeks ago
At least my experience is that a vibe coded app gets to ‘early alpha’ then stays there. If AI is augmenting a developer who is otherwise competent and they don’t let it get ahead of their own understanding, it can work out though not as ‘miraculously fast’ as this. If AI can oneshot an application then it can also work out. If a project vibes itself in a week or two to ‘looks like what it should be but isn’t good enough’, then thus far every one I see is doomed to sit there, because the AI can’t finish it and if you let the GenAI produce that much code, it’s generally not serviceable even for very capable developers.
I spent a lot of time last week basically peeling back an overly vibed attempt at an application by a colleague. It was too flaky and incomplete and deleting large sections of it was frankly more practical than trying to rework it. It had gotten to the point where every attempt to try to have GenAI fix it one way would just break something else. That project was, essentially a fairly small command line utility, I fear to think what something the level of a Photoshop knockoff would be like.
BlindPenguin@piefed.social 2 weeks ago
I found it doing quite decently when using it with strongly opinionated web frameworks and some rules. After all, most web projects are mostly CRUD anyway. For desktop apps it seems mixed, probably because there aren’t many frameworks that give structure. It did however throw together a working app for me so i can map my mouse buttons, but it’s in golang. I’m more of a python / php guy, so i can’t really tell if it’s coded correctly. It didn’t delete my computer yet, though.
With this photoshop clone however, i’m surprised it starts at all. Last time i’ve tested something more complex for desktop, it failed spectacularly. So i’m quite curious how far they can push this thing. What i don’t like, is that GenAI stuff they have to generate images. Coding is one thing, but using it for “automated art” is not so great…
jj4211@lemmy.world 2 weeks ago
Yeah, there are scenarios that it can ‘one shot’, either very straightforward variations of common themes or a use case that is comprehensively instrumented in a way to “know” if it is correct or not automatically, and having the solution iterate against that criteria evaluation until all criteria pass (though in the latter case, I have seen it change the criteria deciding the criteria was a mistake instead of fixing a problem).
If you spend enough tokens, let it rebuild and test on its own, and let it churn and churn, then I’m not too surprised it could make this launch and be superficially intact. The real pain starts when trying to close the gaps the project has in this sort of scenario.
I will say that GenAI makes golang a lot less tedious. Lot’s of obvious but required boilerplate to force the developer to abide by consistent structure. Though I still prefer reading python code since I don’t have to peruse a ton of boilerplate there.