But well over a decade ago, we had high-level frameworks for translating "models" into all the gritty details of a crud app. Model description goes in. API endpoints, database migrations, admin UI, all auto-generated. I even developed something like it based on Django 0.96 (now abandoned).
It's not a stretch of the imagination to generate your model.py with model.txt containing the problem description in plain text.
My experience with every one of them is that getting the model to the point where it will generate something useful is more work than just writing the useful part myself.
Your second point re: how model-driven code generation can be a dead-end... I totally agree. I didn't say it was a good idea! But it was motivated by a real need: generating lots of web apps in a particular domain.
It is quite a stretch to say that lay people can effortlessly write application only with AI whatever it is copilot or chatgpt.
You still have to know a lot of things to build simple crud to store data about stamp collection.
LLMs are not changing that.