I'm pretty sure LLMs have to generate text left-to-right, but there was a screen recording floating around that had "multiple cursors" of text generation at once, where it looked like the LLM was filling in a grid of text, so to speak. I'll see if I can find it.