Karpathy cofounder of OpenAI spends more tokens...
TIKTOK

Karpathy cofounder of OpenAI spends more tokens on knowledge than code now His setup is 3 folders Raw for dumping everything Wiki where the LLM organizes and links concepts automatically Outputs where answers get filed back in At 100 articles it answers complex questions across all your documents No RAG No vector database Just markdown Knowledge that compounds karpathy openai llm ai obsidian

Apr 07, 2026
250 words 80% confidence
So Karpathy, one of the co-founders of OpenAI, just said that he spends more tokens on knowledge than on code itself. Just three folders, no RAC, not VectorDBs, and this is very interesting knowing that RAC was such a big thing not too long ago. First, your REL folder, you just dump everything, articles, screenshots, notes, anything you need, just dump it, don't organize it. The second is the wiki folder, your LLM will read everything you dumped into the first folder, start creating articles, and relate a lot of these concepts with each other. Third is the output folder, you start asking questions to LLM, the LLM searches the wiki, you save all the answers, and as you start asking more questions, the system just gets better. What he said is that at 100 articles, the LLM just answers questions so much better because it has already connected a lot of these concepts with each other in your wiki. For the brain, he uses Codex, Cloud Code, or whatever of these tools, and for visualizing everything he uses Obsidian. You also could build any interface you want with whatever the output of those folders are. I really think that as you start building agents and you want to customize a lot of the things, you can simply just use MD files and start structuring things in a way that will make it easier for the brain like Cloud Code or Codex to actually output the stuff that you need.

Summary

Karpathy's workflow emphasizes knowledge management using three folders. He dumps information in Raw, organizes it in Wiki, and stores answers in Outputs. This method enhances the LLM's ability to answer complex questions effectively.

Key Points

  1. Karpathy prioritizes knowledge over code in his workflow.
  2. He uses three folders: Raw, Wiki, and Outputs for organization.
  3. The Raw folder is for dumping all types of information.
  4. The Wiki folder organizes and links concepts automatically.
  5. The Outputs folder stores answers from the LLM.
  6. At 100 articles, the LLM improves its answer quality significantly.
  7. He utilizes Obsidian for visualization and organization.

Tags

knowledge-management openai llm obsidian workflow

Repurpose Ideas

  • LinkedIn post: How to structure knowledge for LLMs
  • Tweet: 3-folder system for effective knowledge management
  • Checklist: Set up your own knowledge management workflow

Save videos. Search everything.

Build your personal library of inspiration. Find any quote, hook, or idea in seconds.

Create Free Account No credit card required
Original