Karpathy, an engineer at Anthropic, former co-founder of OpenAI, and former AI director at Tesla, proposed treating notes as immutable source code and letting the LLM act as the compiler, sparking a migration across the Obsidian, Claude, and Cursor communities. He argues that saving does not equal ownership and highlighting does not equal understanding, calling meticulously organized databases "cyber mummies." Before this LLM-Wiki framework, the mainstream solution was RAG (Retrieval-Augmented Generation), which Karpathy criticizes as a "courier" that can only handle local information, cannot understand the big picture, suffers from "split personality" and self-contradiction, and leads to "graph rot" with broken links and exponentially declining search efficiency. Karpathy's intuition is that search and retrieval are signs of human limitation, and what is needed is "consensus," "structure," and "truth." He compares knowledge to source code and the LLM to a compiler, where the compilation process is difficult once but every subsequent run is extremely fast, with the cost amortized over thousands or even millions of future uses. Someone has already expanded their wiki to hundreds of pages and hundreds of thousands of words, and automation plugins are beginning to emerge.