What are the core concepts of attention in large language models?