Sakana and Berkeley's MASS has one model run teams of copies of itself, distill what the team figured out, and train on that, so the student, the teacher and the grader are now the same model and nobody human is in the loop.