Do n-gram models also require an understanding of the underlying reality?

If not, which architecture level of next-word predictors does?

If so, then imo the “understanding” is inherit in the data, not the transformation of it