Don't you need to do reinforcement learning through human feedback to get non gibberish results from the models in general?

1900 era humans are not available to do this so I'm not sure how this experiment is supposed to work.