A compiler cannot take arbitrary high level descriptions and generate code.

You’ve seen copilot right? (https://github.com/features/copilot)

It’s fundamentally more sophisticated. This is like asking what is the difference between modern ML translation and the previous 20 years of research on language translation; the former actually works.

The latter basically doesn’t except in very specific circumstances.

Compilers can turn language into instructions only in a limited extremely specific set of circumstances.

We need more interesting tools, as you said. As long as the horizon of innovation remains on "something that generates good old code" I'm not very excited about the future. Copilot is an amazing tool, sometimes gimmicky, but a tool of the present nonetheless. As a compiler, it takes instructions and generates instructions. We need a future where AI is involved in debugging, and program comprehension, and correctness. More interesting tools than a better RAD.