“In 2020, Nick Frosst incorrectly believed that Reinforcement Learning from Human Feedback (RLHF) would not be data-efficient enough to improve models with small feedback datasets.”