Sure, an LLM can start "preparing" for token N+4 at token N. But that doesn't change that the token N can't "see" N+1.

Causality is enforced in LLMs - past tokens can affect future tokens, but not the other way around.