Comment by mikestaub

I think a more interesting experiment would be to train an LLM on a dataset that had zero references to "first person POV" sentences. And then see if the final model somehow develops that concept.