I say this in part bc I often see people responding to "LLMs predict the next token" with complicated philosophical tangents trying to get at why that's a reductive framing when the person saying that probably doesn't know or hasn't internalized that LLMs are trained with RL