Andrej Karpathy proposes idea-file LLM wiki approach

An alternative to traditional RAG architectures is proposed, where an LLM compiles incoming sources into an idea file or wiki just once rather than re-deriving knowledge with every query. This approach shifts the reading and reconciling effort to the moment sources arrive, creating maintainable Markdown pages. However, the author notes that this method has scale limitations and works best at moderate scales. Open article →