Comment by fpgaminer

> This is powerful evidence that even though models are trained to output one word at a time

I find this oversimplification of LLMs to be frequently poisonous to discussions surrounding them. No user facing LLM today is trained on next token prediction.