I think OpenAI has already published some research showing humans preferred smaller/fewer parameter models that were "better trained" through the use of human f

I think OpenAI has already published some research showing humans preferred smaller/fewer parameter models that were "better trained" through the use of human feedback. If there were a model that could adequately replace the role of the human, then that approach would probably work well.