Most people use AI the same way every day. Open a chat, ask a question, get an answer, close the tab. The next day, start over. Every conversation resets to zero. The AI never remembers. You never build anything that compounds.
Andrej Karpathy just showed a different way.
Karpathy is not a casual observer of AI. He co-founded OpenAI, led AI at Tesla, and coined the term “vibe coding” — the practice of describing what you want to an AI agent and letting it build. When he shares a workflow, people pay attention. His April 2 post on X, titled “LLM Knowledge Bases,” has already crossed 1.2 million views and sparked a wave of developers rebuilding their entire research systems from scratch.
The System
The idea is deceptively simple. Instead of chatting with an LLM and forgetting everything, Karpathy feeds raw source material — articles, research papers, GitHub repos, datasets, images — into a folder called raw/. An LLM then incrementally compiles that material into a structured wiki: summaries, concept articles, backlinks, index files. All written and maintained by the AI. Karpathy himself doesn’t manually edit or add anything to the wiki. The LLM writes it, updates it, and runs regular “health checks” — scanning for inconsistencies, filling gaps via web search, and suggesting new articles based on what’s missing.
The frontend is Obsidian, a markdown-based note-taking tool. The LLM writes. You read. His current knowledge base on a recent research topic: roughly 100 articles and 400,000 words. Longer than most PhD dissertations. Built without typing a single word.
Also Read: Former Tesla AI Director Andrej Karpathy rejoins OpenAI
Why This Beats RAG
For the past few years, the standard approach to giving AI access to your own documents has been RAG — Retrieval-Augmented Generation. You chunk documents into pieces, convert them into mathematical vectors, store them in a database, and retrieve relevant chunks when you ask a question. It works, but it’s a black box. You can’t read the embeddings. You can’t audit what the AI found. You can’t trace an answer back to a specific source.
Karpathy’s system rejects all of that complexity. Because the wiki is just markdown files, every claim is traceable. Every article is readable. Every connection is visible. He notes he expected to need complex RAG infrastructure, but at personal knowledge base scale, a well-structured markdown wiki turns out to be something a modern LLM can navigate “fairly easily.”
The Follow-Up Was the Real Signal
After the original tweet went viral, Karpathy did something that quietly said more than the workflow itself. He didn’t share the code. He didn’t release an app. He published a GitHub Gist — an “idea file” — and explained: in the era of LLM agents, there’s less point sharing specific implementations. You share the idea. Each person’s agent builds a version customized for their specific needs.
That’s a meaningful statement about where AI development is going. The product is increasingly the concept, not the code.
Developer Farza built a live example of exactly this — “Farzapedia” — a personal Wikipedia compiled from 2,500 entries across his diary, Apple Notes, and iMessages. The result: 400 articles covering research areas, people, projects, and ideas, all interlinked, all maintained by AI. Karpathy highlighted it as proof of concept.
What It Means for You
If you work in data science, AI research, or any field where staying current matters, this is worth paying attention to. The competitive advantage in the next phase of AI isn’t going to come from knowing how to prompt better. It’s going to come from having better systems — structured, compounding, queryable knowledge that gives your AI agents the context they need to do genuinely useful work.
Karpathy ended his original post with a line that’s worth sitting with: “I think there is room here for an incredible new product instead of a hacky collection of scripts.”
He’s right. That product doesn’t exist yet. But the workflow does. And it’s available to anyone willing to set it up today.
