LLMs are still just next token predictors. Blasting out words in a forward-only way is how you write a first draft. It's a form of thinking out loud. The ideas are still forming while you generate the text. Generating the text forms the ideas.
Communicating well requires editing. Editing requires going backwards, not just forwards. It requires turning things over in your head. It requires rethinking connections. It requires knowing when one word says more than a sentence. It requires remembering what you were trying to say. It requires remembering.
LLMs don't have much of that. They can mimic some bits with loops, but usually not very well.
It seems like thinking should be the place where it can iteratively draft before creating an output. I've wondered if text diffusion models could in theory write better, thanks to iteratively rewriting/refining the text. It seems to me at least that avenues like that are worth pursuing, so I'm happy to see efforts like Mercury 3 and Diffusion Gemma.
But LLMs don’t think. Neither do diffusion models.