Comment by adastra22

What I mean is that all processing in an LLM occurs in state space. The next-token prediction is the very last step.