Comment by nl

Modern LLMs are post trained for tasks other than next word prediction.

They still output words through (except for multi-modal LLMs) so that does involve next word generation.