There are tools which increase human productivity, while still requiring it (barcode scanners, for example). And then there are another class of tools which make human labour obsolete (self-checkout systems).
LLM's (as they exist today) could be considered both. Github copilot would be an example of how GPT can be used as a productivity tool by human programmers, but as the technology progresses AI will become less of a "copilot", and will gradually replace humans as the main decision maker. Then eventually LLM will probably be used by people completely none technical to replace the need for coders entirely.
Now the argument becomes, well this will open up new opportunities. Instead of being a programmer you can be be a user researcher on a project, which could be the case, but this is a much more nuanced argument.
The most well paid jobs are typically those which require years of knowledge retention and require the human to basically serve as an advanced expert system in some domain.
Both "good" and "bad" programmers can write code into a text editor, the difference is that the "good" programmer will make decisions backed up by years of professional experience – same for a good doctor or good lawyer. This is why we pay more for these professions, because that depth of knowledge is hard accumulate.
This is the very thing that GPT attacks. What it can't replace is someone physically laying bricks or plumbing pipes. But there is less depth of knowledge required in jobs like this which limits salaries for these professions.
So sure. Perhaps in the future someone can say, "hey, GPT, build me [x]", but just remember you won't be the only one who can do that and there is no significant depth of knowledge in such a job. So while GPT won't replace all jobs and may even create some new ones, expect it to replace or devalue the majority of "good" jobs like doctors, programmers, lawyers, designers, etc.
So to Carmack's point, he's right you'll be able to build websites and apps faster using GPT as a tool, but you'll probably do so for a fraction of the salary.
We've ran similar experiences over the last several decades with outsourcing. If your labour can be easily out sourced then your ability to retain a good salary drops. It's not that a worker today can't make clothes in the US faster and better than at any point in the past, it's that it makes no economic sense to do so.
-----
Another thing I'd note here is that I'm autistic as are a lot of programmers in my experience. My brain is built to do technical things and I struggle intensely with human interaction. In my opinion it's not that programmers "don't understand" software solves problems for people, it's that a lot of us don't naturally excel in those areas. I think it's fair to say a lot of like to stick to what we're good at, and that's generally writing code and designing complex systems. The more time I have to talk with users about their needs instead of doing technical work, the less useful I am. And I'm guessing designers are also people who want to design rather than type prompts into a chatbox.
So another consequence here is that we might increasingly be forced to do jobs we don't really want to do as AI restricts the areas of labour where humans can still compete.