Programmer Humor

Eh, that's not quite true. There is a general alignment tax, meaning aligning the LLM during RLHF lobotomizes it some, but we're talking about usecase specific bots, e.g. for customer support for specific properties/brands/websites. In those cases, locking them down to specific conversations and topics still gives them a lot of leeway, and their understanding of what the user wants and the ways it can respond are still very good.