Who ever came up with the idea of training Language Models to be conversational?

I just think it's such an interesting step in logic, to take a basically an advanced autocomplete model, and ask it to "autocomplete" on chat responses rather than simply arbitrary text blocks.

I think it's *super* fucking cool, that you can sort of "goad" a text prediction machine into replying to you, with responses that actually take into account the semantics and logic of what came before. *Super* fucking cool. I feel like I've only ever seen that part of the whole process glossed over.