[name] asked: Why can attention in LLMs disperse over long contexts? What do you know about this?