Sorry, when I said "predictive power", I was actually assuming normalization for efficiency. That is, my claim that the total predictive capacity you get for your available computation resources is greatest by taking the physical stance in this case.