I don't know what model you're using or how you're prompting it, but for me some 60-80% of the time the results require only a little bit of steering to be 'acceptable' (like at least what I would expect from a junior engineer and I'll approve the PR even though it's not quite how I would do it), some 30% of the time the results are pretty much what I would do, and some 10% of the time the results are better than what I would do ("huh, good idea, okay let's do it that way").

They're not perfect by any stretch but if they're being likened to slot machines for code, I'll take those odds almost every day.

Have you seen the kind of code an average developer writes?

I agree that LLMs produce code that is less good than what a good developer could write, but most developers are not good developers, and even the good developer gets tired and must sleep eventually.

Arguments against LLMs are like the old arguments against high level languages like C. People argued that the compiler wrote trash code, that humans could do better, that the costs weren't worth it. None of that mattered and it's the same story here.