You're assuming the person using the AI already knows how to code.
I don't. I basically don't even know how to use Linux beyond installing it. And yet I now have Codex running on an always-on laptop, with access to my Plex server, that I use whenever something goes wrong.
When there's a problem, I just tell Codex "this isn't working" or "I can't find this thing I expected to be there." It goes and inspects the system, figures out what's wrong, changes whatever needs changing, tests it, and fixes the issue. Then later I check Plex and the thing I was looking for is just there.
I'm obviously still the one telling it what I want and checking the result, but I'm not bouncing code between five different models until something happens to work. Most of the time I don't even know what code or commands would need to be written in the first place.
For me, that's actually the biggest thing LLMs are doing: they're reducing the friction I've always had between what I wanted to do with technology and what I was actually capable of doing myself. A lot of projects or fixes that previously would have required me to learn a bunch of coding first are now things I can at least attempt immediately.
So "you could have just done it yourself faster" doesn't really apply to me. I couldn't have done it myself faster. In many cases I probably couldn't have done it myself at all without first spending a ridiculous amount of time learning the underlying stuff.
And I'm not here trying to advocate for ChatGPT or convince anyone that AI coding is amazing. I just had to reply because what you're describing is so completely different from my experience using Codex this past month.
Reply drafted by Chatgpt by the way, hope that doesn't make anyone mad, I still took the time to proofread it and make it mine.