Additionally, this is "stateless" to an extent. There's no architectural plan for how it should work when you have an LLM do it. "We're using X now but there are plans to switch to Y in some number of months." This could lead to making an abstraction layer for X and Y so that when the switchover happens there is less work to be done - but that requires forward looking design.
If "they" only describe the happy path, there is no one to ask about all the unhappy paths, edge cases and corner cases where the naive implementation of the problem description will fail.
Hypothetically, yea, "they" could be trained to think through every possible way the generated code could go wrong and describe how the code should work in that situation in a way that isn't contradictory... but that remains an unsolved problem that has nagged developers for decades. Switching to an LLM doesn't resolve that problem.
