I'd expect that any kind of logistics would need tight bounds on min-max 'performance', and not be up to the whims of artificial stupidity. Am I wrong?

"Training a model" is just a different name for "optimization".

The typical logistics optimization system is "smart" because it's optimizing exactly what it is supposed to. LLMs here would not be "smart" as they are optimizing toward a different target (humanlike use of language), and using them for things they're not specifically trained to do is indeed stupid.

I meant sexy in the sense of fashionable. People are impressed if you can solve TSP for a few nodes with a NN, but they don`t really care if you can solve for 100k of nodes with good ol' branch-and-cut.