I can think few great tools like t3code (a software that does a lot) that have been mostly coded by LLMs and yet have a solid architectural foundation and implementation. The software is genuinely gargantuan in its features and capability, you really cannot appreciate it from quickly skimming the readme.

At work, LLMs helped a lot in doing the kind of work you were always tired or had no budget to do, and the codebases have never been healthier, more tested and better documented.

In any case, it's always about the effort and capabilities of the people using it. If you want to vibecode, then you're still stuck in prototype land and I think that even better models won't help, in fact they may have negative value.

The naivety of your post is in any case appalling, you seem to be missing the forest for the tree.

What’s naive about it? Maybe a bunch of people are finding great new software only possible because of LLMs. Good for them. I’m simply stating I haven’t.

(I note in your comment you don’t mention one piece of end user software or how the user has benefitted…)