Devil's advocate though. With things like 4GLs, it was still all on the human to come up with the detailed spec. Best case scenario was that you work very hard, write a lot of things down, generate the code, see that it didn't work and then ???. That "???" at the end was you as the programmer sitting alone in a room trying to figure out what a non-responsive black box might wanted you to have said instead.

It's qualitatively different if you can just talk to the black box as though it were a programmer. It's less of a black box at that point. It understands your language, and it understands the code. So you can start with the spec, but when something inevitably doesn't work, the "???" step doesn't just come back to you figuring out with no help what you did wrong. You can ask it questions and make suggestions. You can run experiments. Today's LLMs hit the wall pretty quick there, and maybe they always will. There's certainly the viewpoint that "all they do is model text and they can't really learn anything".

I think that's fundamentally wrong. I'm a pretty solid programmer. I have a PhD in Computer Science, and I've worked as a software engineer and an architect throughout a pretty long career. And everything I've ever learned has basically been through language. Through reading, writing, speaking, and listening to English and a few other languages. I think that to say that I can learn what I've learned, but it's fundamentally impossible for a robot to learn it is to resort to mysticism. At some point, we will have AIs that can do what I do today. I think that's inevitable.