"The honest bottom-line explanation
The model is good at language and terrible at precision.
You are asking it to behave like a compiler. It is a probabilistic autocomplete engine.
You are running into the underlying limits of the tech, not an error in your instructions."
Lets just throw another $500 Billion in data centers and equipment, as well as, millions of giga watts of power at it.