Why Karpathy’s LLM Wiki Is Sparking a New Knowledge‑Building Approach

LLM Wiki concept

Andrej Karpathy released an “idea file” (a GitHub gist) that describes a meta‑framework for building a personal knowledge base that is maintained by a large language model (LLM) agent. The framework is model‑agnostic and treats the LLM as a programmer that reads raw material, generates a structured Markdown wiki, and continuously updates it.

Gist URL: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Closed‑loop workflow

Collect raw sources (papers, articles, code, images) in a raw/ directory.

Prompt an LLM to compile the sources into a structured wiki of Markdown files with backlinks and concept classifications.

Browse the wiki with Obsidian (or any Markdown viewer).

When the wiki reaches a moderate scale (Karpathy’s example: 100 articles, ~400 k words), pose complex questions that span the whole collection.

Archive each Q&A as a new wiki page, thereby strengthening the knowledge base.

Periodically run LLM‑based health checks to surface contradictions, fill gaps, and suggest new research directions.

Three‑layer architecture

Raw data layer : immutable source files that the LLM only reads.

Wiki layer : LLM‑generated Markdown pages (summaries, entity pages, concept pages, comparative analyses, overviews) that the LLM creates, updates, and cross‑links.

Schema layer : a configuration document (e.g., CLAUDE.md or AGENTS.md) that tells the LLM how to ingest data, answer questions, and maintain the wiki, turning a generic chat model into a disciplined wiki maintainer.

Operational steps

Ingest : Add a new source to raw/, let the LLM read it, discuss key points, write a summary page, update indexes, and modify related pages (typically 10–15 pages per source). Users may process one source at a time for close supervision or batch multiple sources for speed.

Query : Pose a question to the wiki; the LLM searches relevant pages, synthesizes an answer, and can output the result as a Markdown page, comparison table, slide deck, chart, or canvas. Valuable answers are re‑archived as new wiki pages.

Lint (quality check) : Periodically have the LLM scan the wiki for contradictions, outdated conclusions, orphan pages, missing concepts, absent backlinks, or data gaps, and suggest new research directions or sources.

Scale and RAG comparison

Karpathy notes that at a moderate scale the system does not depend on traditional Retrieval‑Augmented Generation (RAG). As long as the LLM can maintain an index and summaries, it can support effective retrieval and reasoning without re‑searching the raw sources for every query.

Future extensions

The idea can be extended by generating synthetic data and fine‑tuning the model so that knowledge becomes embedded in model weights rather than being fetched from a context window, moving toward a self‑enhancing knowledge system.