Do you have a link to some of the code that you have produced using this approach? I am yet to see a public or private repo with non-trivial generated code that is not fundamentally flawed.
https://github.com/micahscopes/radix_immutable
I took an existing MIT licensed prefix tree crate and had Claude+Gemini rewrite it to support immutable quickly comparable views. The execution took about one day's work, following two or three weeks thinking about the problem part time. I scoured the prefix tree libraries available in rust, as well as the various existing immutable collections libraries and found that nothing like this existed. I wanted O(1) comparable views into a prefix tree. This implementation has decently comprehensive tests and benchmarks.
No code for the next two but definitely results...
Tabu search guided graph layout:
https://bsky.app/profile/micahscopes.bsky.social/post/3luh4d...
https://bsky.app/profile/micahscopes.bsky.social/post/3luh4s...
Fast Gaussian blue noise with wgpu:
https://bsky.app/profile/micahscopes.bsky.social/post/3ls3bz...
In both these examples, I leaned on Claude to set up the boilerplate, the GUI, etc, which gave me more mental budget for playing with the challenging aspects of the problem. For example, the tabu graph layout is inspired by several papers, but I was able to iterate really quickly with claude on new ideas from my own creative imagination with the problem. A few of them actually turned out really well.
(edit)
I asked it to generate a changelog: https://github.com/wglb/gemini-chat/blob/main/CHANGELOG.md
It's not about lines of code or quality it's about solving a problem. If the problem creates another problem then it's bad code. If it solves the problem without causing that then great. Move onto the next problem.
In other words, it matters whether the AI is creating technical debt.
That has nothing to do with AI/LLMs.
If you can't understand what the tool spits out either; learn, throw it away, or get it to make something you can understand.
Weren't there 2 or 3 dating apps that were launched before the "vibecoding" craze that went extremely popular and got extremely hacked weeks/months in? I also distinctly remember a social network having firebase global tokens on the clientside, also a few years ago.
We went from "this thing is a stochastic parrot that gives you poems and famous people styled text, but not much else" to "here's a fullstack app, it may have some security issues but otherwise it mainly works" in 2.5 years. People expect perfection, and move the goalposts. Give it a second. Learn what it can do today, adapt, prepare for what it can do tomorrow.
LLMs are still stochastic parrots, though highly impressive and occasionally useful ones. LLMs are not going to solve problems like "what is the correct security model for this application given this use case".
AI might get there at some point, but it won't be solely based on LLMs.
Frankly I've seen LLMs answer better than people trained in security theatre so be very careful where you draw the line.
If you're trying to say they struggle with what they've not seen before. Yes, provided that what is new isn't within the phase space they've been trained over. Remember there's no photographs of cats riding dinosaurs but SD models can generate them.
Repeat after me, token prediction is not intelligence.
I have experimented with vibe coding. With Claude Code I could produce a useful and usable small React/TS application, but it was hard to maintain and extend beyond a fairly low level of complexity. I totally agree that vibe coding (at the moment) is producing a lot of slop code, I just don't think Tea is an example of it from what I understand.