On April 3, 2026, Andrej Karpathy introduced the concept of “LLM Knowledge Bases” in a brief post that lit up Twitter and Hacker News. His core proposal: skip the usual retrieval-augmented generation (RAG) step where you fetch chunks from raw files at query time. Instead, have a language model ingest everything—papers, screenshots, repos, docs—and spit out a fully linked markdown wiki. That wiki then becomes the persistent knowledge store, one that gets richer with each update and cuts down query costs over time. Just two days later, Graphify emerged. You point it at any folder—code, docs, images—and run a single command. Under the hood, it uses an LLM to generate summaries, backlinks, concept maps and an index file. No vector database, no custom config, no complex retrieval pipeline. Everything lives as plain markdown, which the model can navigate autonomously on future queries. The tool hit GitHub, got a simple README, and within hours had hundreds of stars. The reaction was immediate. AI researchers and devs praised the radical simplicity. Instead of wrestling with embedding stores or designing prompts to sift through raw text, Graphify offers a ready-made knowledge graph you can inspect, edit or version in Git. People posted screenshots of sleek diagrams showing topic clusters, shared tips for organizing folders, and debated how often to recompile as projects evolve. This sparks a bigger shift in how teams might manage research and docs. With markdown as the backbone, you get full transparency—no hidden index or proprietary search engine. And because the wiki grows cumulatively, each addition compounds previous work rather than creating duplicate embeddings. It’s an experiment in trading retrieval complexity for upfront compilation, and so far, everyone’s intrigued.