At then end they briefly mentioned how they started a service to post-train LLMs on producing more human, less formulaic-obviously-AI text.