Ask HN: AI is destroying my software

This is maybe a bit of a rant, but I am quite exhausted recently. Customers demand extremely fast delivery of new features as they are getting used to AI-assisted development and can vibe-code stuff themselves, so naturally they expect us as their software provider to also react to new feature requests with a new release within days instead of weeks.

That forces us to use heavily AI-based workflows, but if we're honest the speed advantage we can get through that only materializes if we are skipping detailed manual code review of all the AI-changed code, as that produces tons and tons of revisions and edits.

What I found after around a year is that the codebase is riddled with inconsistencies and hacks that were quitely introduced by the AI, really ridiciulous stuff sometimes. I'd say the codebase is close to the point where it will become unmaintainable and we will have to consider a large scale refactor or cleanup, only that with the new pace everyone seems to be going at this becomes very hard. So I'm feeling a bit trapped, and I become more and more skeptical of AI for software development. Don't get me wrong for some things it works great but I think I have to stop using it for large-scale feature implementation or anything that is more complex than glorious refactoring or autocomplete within very narrow bounds, as it's just not there yet.

And yeah I'm using the latest and greatest models, even they are not able to produce something that is consistent over the long run it seems. I have also tried writing research papers with Fable and all these other glorified models and that was a disaster as well, can't get how people think these things accelerate development by 10x, even 2x seems hard to get outside of specific areas where you have a very well specified problem that you can exactly describe to an agent. On the large scale, these things are still so far from the general ability of a developer it's laughable. They are very smart for keeping a ton of context and producing extremely precise solutions, but thinking about software code on a large scale over an extended period of time is something they can't do.