It still screws up even simple tasks, and it often takes more time to accurately describe what you want than just writing the code.
And most importantly: it cannot understand complex contexts.
I am working on a corporate product that is essentially a big monolith, with much of the logic being complex because of legal reasons (the product is about real estate), and some bad past design decisions.
The features I am currently working on cannot be vibe-coded, because no AI can understand the context. And I have worked with similar projects in the past -I know that corporate software of monstrous complexity is usual.
The same applies to a caching library I am developing on and off on my spare time: I asked Claude to make some improvements, and it didn't even compile. The result was a disaster.
An actual question, to all of you who 'mourn your craft' and LLMs replacing you: what on earth have you been working on?
Am I the only one who doesn't see LLMs replacing us anywhere soon? Or am I just very lucky?