I like to vibe code single self-contained pages in html, css, and JavaScript, because there's a very slim chance that something in the browser is going to break my computer.
There will be bugs that the AI cannot fix, especially in the short term, which will mean that code needs to be readable and understandable by a human. Without human review that will likely not be the case.
I'm also intrigued by "see if it works". How is this being evaluated? Are you writing a test suite, manually testing?
Don't get me wrong, this approach will likely work in lower risk software, but I think you'd be brave to go no human review in any non-trivial domain.
This is the problem I have seen a lot. From professionals to beginners, unless you actually carved the rock yourself, you don't really have an insight into every detail of what you see. This is incidentally why programmers tend to like to just rewrite stuff than fix it. We're lazy.
except someone less experienced never gets to try, since all the experienced programmers are now busy shepherding AI agents around, they aren't available to mentor the next generation.
Hmm. And maintenance?
Good code is less visible.
Vibe coding is "don't write any libraries, just copy-paste and use these existing libraries uncritically".
The worst outcome of this will be NPM full of vibe-coded copy-paste code shoved into opaque modules that nobody has ever really read, that other vibe-coding tools will use uncritically.
My concern is that the act of surrendering ownership itself is always worse.
But this is, properly expressed, "see if it works, based on my incomplete understanding of the code that I haven't worked through and corrected by trying to write it, but I have nevertheless communicated directly to an AI that may not have correctly understood it and therefore may not have even the vaguest inkling of the edge cases I haven't thought to mention yet".
Vibe-coded output could only properly be said to "work" if we were communicating our intentions in formal methods logic.
What you mean is "apparently works".
You notice how LLMs do good on small tasks and on small projects? The more LLM code you add to your projects, the slower and worse (read: more tokens) they perform. If this was by design (create bigger, unmaintainable projects, so you can slowly squeeze out more and more tokens out of your users), I'd have applauded the LLM creators, but I think it's by accident. Still funny though.
This still happens quite a bit, and it's just like taking away a hard task from someone less experienced. The difference is there is no point in investing your time teaching or explaining anything to the AI. It can't learn in that way, and it's not a person.