On the evening of April 2, 2026, AI expert Andrej Karpathy, a member of OpenAI's founding team and former Tesla AI director, posted on X detailing his recent workflow for building a personal knowledge base using large language models (LLMs). The post quickly sparked global discussion, accumulating over 9 million views and more than 36,000 likes, and has been regarded as a landmark practice signaling AI's move from a "code era" to a "knowledge era."
Karpathy wrote that his token consumption patterns have recently shifted: where tokens were once spent primarily to "manipulate code," they are now increasingly used to "manipulate knowledge" stored as Markdown and images. He observed that the latest generation of LLMs is now strong enough at organizing, summarizing and reasoning over knowledge to serve as a knowledge base's "compiler" and "maintainer," with humans only needing to review intermittently.
Full workflow: Raw data → LLM compiles Wiki → Obsidian visualizes
Karpathy summarizes his workflow as a "Raw → Compile → Query → Enhance" loop, relying on only three core tools: a local folder, an LLM agent and the Obsidian note app.
- Ingest: Research materials (papers, articles, code repositories, datasets, images, etc.) are placed in a unified "raw/" directory. Web articles are converted into clean Markdown with Obsidian Web Clipper in one click, and related images are downloaded locally via hotkeys so the LLM can reference them directly using its vision capabilities.
- LLM compiles the Wiki (Compile): The LLM reads the raw materials in "raw/" one by one and automatically generates a structured Markdown Wiki. The Wiki contains summaries of all documents, backlinks, categorizes content into concepts, and writes topical articles that interlink. The process is incremental: when new material is added the LLM updates relevant entries automatically. Karpathy emphasized: "The LLM writes and maintains all the Wiki data; I rarely edit directly."
- Obsidian as the front end (IDE): Obsidian is used to view raw materials, the compiled Wiki and various derived visualizations (charts, slides). Obsidian plugins (such as Marp) can render Markdown directly into slides. Karpathy noted that keeping his primary Vault clean and letting the agent operate in a separate Vault helps prevent "contamination" risk.
- Complex Q&A and research (Q&A): Once the Wiki reaches roughly 100 articles and around 400,000 words, users can pose complex questions directly to the LLM agent. The LLM maintains index files and document summaries, enabling efficient retrieval within the context window without relying on a complex RAG system. Karpathy found that at this "small scale" the LLM performs research tasks well.
- Output and feedback loop (Output & Filing): The LLM no longer outputs plain text only; it generates Markdown files, Marp slides, Matplotlib charts and other visual results. These outputs can be filed back into the Wiki to further enrich the knowledge base. Karpathy said: "Every exploration and query increases the Wiki's value."
- Linting and health checks: Periodically the LLM runs "health checks" on the Wiki to find inconsistencies, fill missing information (combined with web search), discover new concept connections and even suggest future research directions, enabling continuous optimization of the knowledge base.
He also vibe‑coded a simple Wiki search engine and gave it to the LLM as a tool. In the long term, he plans to synthesize data and fine‑tune models so knowledge is "internalized" into model weights rather than relying solely on context.
Community reaction: from personal practice to product discussion
After the post, figures such as Lex Fridman and Obsidian CEO kepano (@kepano) responded quickly. Lex Fridman said he uses a similar hybrid setup with Obsidian+Cursor+dynamic HTML and even turns mini knowledge bases into voice interactions for learning while running. kepano advised a strict separation between a personal Vault and an Agent Vault to avoid agent‑generated content polluting human knowledge.
Reddit, LinkedIn and VentureBeat followed up rapidly, and developers shared similar workflows or tools built on the idea. Many see this not only as a personal productivity improvement but as a shift from one‑off question answering to long‑term knowledge compounding by AI agents.
Far‑reaching impact: the start of a new era in AI knowledge management
Karpathy's workflow reveals a new direction for AI applications: LLMs are no longer just chat tools or code assistants but can act as a personal knowledge "compiler" and "steward." Its core value lies in the knowledge accumulation effect—each query feeds back into the Wiki so the knowledge base evolves with use rather than restarting from zero each time.
The practice offers insights for developers, researchers and companies:
- Personal productivity: Researchers can quickly build domain‑specific knowledge bases and reduce redundant reading and organizing.
- Agent ecosystem: It promotes evolution from simple RAG to "LLM‑autonomously‑maintained knowledge graphs," laying groundwork for multi‑agent collaboration.
- Product opportunity: Karpathy candidly noted at the end of his post that "behind these hacky scripts lies the potential to build an incredible new product." Several Obsidian+LLM plugins and tools have already appeared, and more commercial knowledge‑base products are expected in 2026.
At the same time, the workflow highlights current AI limits: at very large scale (millions of words), indexing and search technologies remain necessary; data privacy and version control are still challenges.
Karpathy's share reiterates that AI's real value lies not in model parameters alone but in deep integration with human knowledge workflows. The shift from "manipulating code" to "manipulating knowledge" reflects not just a change in personal habits but a microcosm of AI's transition from assistant to intelligent partner. In the future, "LLM knowledge bases" may become standard for knowledge workers, and the Obsidian+LLM combo may quietly define the next generation of personal knowledge management paradigms.
▶️ Enjoy this piece? Creating content is hard!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com