Comment by nucleative

Is there no value in how the training was done such that it's accessible via inference in a particularly useful way?