What this means is that the consistency gained in programming languages would be lost. A computer program would no longer generate repeatable results.
I think both of our points are somewhat moot though, as I don't see NLP code as something that you would necessarily put up on github. I see it as playing a similar role to DSLs or metaprogramming nowadays, and similarly, it wouldn't always be a good solution.
Simple use case: I recently finished an HTML5 puzzle minigame. The layout is pretty simple and has to adjust to window resizes, but the layout can't be done with straight HTML/CSS. So, my resize function is 120 lines of dense javascript, despite being able to describe in a simple paragraph what I would want. Indeed, a simple paragraph is what was given to me in the first place to describe the layout.
"Hello Computer! I need an inverse square root function. It should take a float as a parameter and..." - Not practical
"Hello Computer! I need this image to pan across the screen. It should take one second. When it gets there, I want it to stop and become clickable. When you click it..." - Very practical
The practical solution here could even be iterative...
"No computer, no! The image should ease into position, not just go straight there!"
The point is you wouldn't rely on it for precise logic, just "good enough" behavior. You can definitely argue that this isn't programming though.