Comment by bitexploder

There is nothing inherent in the minimization process that says you must distill from a larger model. It is just the thing we know how to do right now.