“RLHF does not make a model fundamentally smarter, but rather "bridges the gap" to help the model better communicate its existing intelligence to humans.”