That is literally the exact promise of CASE tools in the 80s and the early 90s; UML code generation tools in the 2000s, and "low-code/no-code" platforms in the 2010s. It turned out to be a disaster every time, especially when the Idea Persons chucked their creations over the wall to SWEs to bash them into actual products because the Idea Persons had Far More Important Things To Do than maintain their coalesced brain farts.
We're repeating history but with more energy consumption.
LLMs unlock a fundamentally different paradigm of interaction, in my experience a non-technical person with a good humanities background can describe what they want adequately. This is without needing to master the arbitrary grammar of a no-code system. Does often inevitably turn out to be a 'toy' version of what a real business needs? Yes, but it's still strictly better than previous ways of working.
I do think that there’s a difference in kind here - we’re not producing UML graphs that require programmer time to implement (or sending the diagrams to SE Asia and then code reviewing).
The code ‘works’ - and the folks who are improving the prototype can also benefit from the tools that the Idea Person used.
The code worked for the examples the OP gave as well. They weren’t talking just about UML graphs, but about automated tools to turn those graphs into code.
And in the case of low code/no code, those produced working prototypes as well. And you could in most cases export them to raw code.
I wasn't around for the second millennium versions. At some point, doesn't there exist a kind of activation energy threshold where enough money/promise etc is gained from the prototype that this pattern works for good ideas and not for bad ones?
We're repeating history but with more energy consumption.