LLM Wiki is Like Adding a Hard Drive to the Model; RAG is Like Giving it a Library

Imagine you're standing in a massive library with all the books on shelves, but you can only borrow one at a time, reading it before swapping it out. This is daily life for RAG—retrieval, concatenation, generation, every step timed to the second. Now someone tells you, why not just engrave the entire library's content directly into your brain so you can recite any passage instantly? That's the promise of Andrej Karpathy's push for LLM Wiki: let the model remember knowledge itself instead of flipping through books every single time.