I thought refining the model with its own outputs was a huge part of training them.