A new research paper "Don't Do RAG: When Cache-Augmented Generation is All You Need for Knowledge Tasks" (arxiv.org/abs/2412.15605) suggests we might be overcomplicating how AI systems access knowledge.