I've found two personal use cases for LLM generated code:

(1) I have an idea for some app, but either I feel it won't be useful enough/save me enough time to justify developing it, or I simply don't feel the problem is interesting enough to be motivated by it. In that case, a vibe coded tool is perfect. It generally does one simple thing, and I don't care about long term maintenance, because it just needs to keep doing that thing.

(2) Adding a feature to an open source project. Again, it's a case of "I want this feature, but am not willing to spend the time needed to implement it." Even a relatively simple open source project can take a day or two just to get a basic understanding of the code and where I need to make the changes. Now I can often just get a functioning vibe-coded implementation within a few hours.

(2) leaves me with some unsettling feelings about how this will affect the future of open source software. Some of the features I've implemented this way may very well be useful to other users, but I can't in good conscience just dump a vibe coded pull request on a project and except them to do the work of vetting it. But if I didn't have the energy to implement the change myself, I'm definitely not going to bother doing the work of going through all the LLM generated code, cleaning it up to the standards of the project, etc. Whereas before I didn't have a choice, and the idea of getting the change ready for a PR was much less daunting since I understood the problem space and solution well.

So at least for myself, I can see a future where many of the apps I use are bespoke forks of popular applications. Extrapolate that to many, many people and an interesting landscape emerges.