> permanently tying up a chunk of capacity spent on producing output that is only useful for a single purpose — feeding a learning model.

The horror.