The first part of your comment doesn’t seem related to the second part, and neither part seems related to the question of whether most applied problems do or do not benefit from making a distinction between the sampling distribution used for generating training and test data and the “true” distribution.
