RL with the harness inputs and outputs of users is one of the primary improvers of model performance, a self perpetuating flywheel.