Comment by 👻 darkghost
Re: "Been seeing this term around the internet recently, called "…"
I'll be surprised if AI programming is used for anything other than very small bespoke applications, at least in its entirety, for the next few years. While it can regurgitate code chunks I am unconvinced it can produce projects all on its own. With future advancements, I don't see it progressing far beyond the kind of projects you find in how to guides. Maybe I'll make VibesOS, the operating system you don't dare run on bare metal connected to the internet. It's all vibes coding with the only comment in the code being "I don't know what any of this does."
2025-03-27 · 1 year ago
4 Later Comments ↓
🐦 wasolili [...] · 2025-03-27 at 15:15:
I don't think LLMs/AI will be successfully used for entire projects any time soon. The existing AI tools fucking suck. I have tried using them, and will continue to try, but the entirety of my experience can be summed up as "AI generates broken code, guiding it to near-correct code takes longer than if you wrote it yourself."
Reminds me of the phrase "an hour of googling can save you five minutes of reading the documention" - "an hour of prompt engineering can save you 5 minutes of reading the documentation, and lighten your wallet!"
What I am worried about is that LLMs have given people who absolutely have no business being near a single line of code the illusion that they can write code. This results in horrible, broken code that actual developers have to waste time reviewing/refactoring/rewriting.
I am also highly suspicious of the claims companies make that a certain product or feature was developed with heavy use of AI. The people making those claims aren't programmers, they're the marketing people who heard AI was the new big thing, and are probably claiming AI was used to a greater extent than it was in part due to being unfamiliar with the extent the actual programmers used it and (in larger part) because they think "made with AI" will appeal to stakeholders
🚀 clseibold [OP, 🛂] · 2025-03-27 at 15:18:
I've personally had good experiences with Claude generating a function or a very small portion of code pretty decently and correctly, but I haven't tried making whole projects with it, let alone anything more than like a 25 lines max. It sounds kinda like a pain to use it exclusively for a whole project, honestly, lol.
I agree, if you can't read the code generated by an LLM to know whether it's correct or not, then that's really not a great thing.
🌲 Half_Elf_Monk · 2025-03-27 at 15:20:
I would love to have some kind of "code authority" institutions that verify the integrity and safety of code in OS projects. It'd help non-programmers evaluate the safety of code, and would provide some kind of imprimatur for a codebase' safety, LLM-free status, and bloat levels. I wish I could read/understand every line of code on running on my machine, but realistically... that's not going to happen.
👻 darkghost · 2025-03-27 at 15:54:
I'm old enough to remember IRC. The prank we would play on people is "hey you can get IRC Operator status by just pressing Alt+F4!" then watching a bunch of people leave. Using an LLM to code without understanding what you're reading is a bit like that. Is it still social engineering if it's an LLM?
Original Post
Been seeing this term around the internet recently, called "Vibe Coding", and I finally looked it up. What that absolute heck?!? 🤣