In a move that could reshape how we interact with artificial intelligence, Andrej Karpathy—the AI visionary behind Tesla’s Autopilot, a founding member of OpenAI, and the mind behind the term "vibe coding"—has quietly launched LLM Wiki, a groundbreaking experiment that turns large language models into living, evolving knowledge repositories.

For years, AI researchers and developers have relied on Retrieval-Augmented Generation (RAG)—a method where LLMs pull snippets from databases to answer questions. But Karpathy’s new approach flips the script. Instead of treating AI as a static question-answering machine, LLM Wiki positions it as a research librarian, one that doesn’t just retrieve information but compiles, refines, and maintains an ever-growing body of knowledge in real time.

Why This Matters

Traditional RAG systems suffer from a critical flaw: they treat knowledge as fragmented, temporary, and context-bound. Once an LLM’s session resets, so does its understanding of your project. Karpathy’s frustration with this statelessness led him to rethink the entire workflow.

His solution? A self-sustaining Markdown wiki, where an LLM doesn’t just fetch answers—it authors, organizes, and improves them over time.

"I’ve been spending far more of my token budget manipulating knowledge than code," Karpathy wrote on X. "The LLM is no longer a glorified search engine. It’s a living archive."

How It Works

Karpathy’s system is deceptively simple yet revolutionary. Here’s how it breaks down:

1. The Raw/ Directory: Where Chaos Meets Order

All source material—research papers, articles, GitHub repos, datasets, even images—goes into a raw/ folder. Karpathy uses Obsidian, a note-taking app, to store everything in plain Markdown files, ensuring maximum compatibility with LLMs.

2. Incremental Compilation: AI as the Ultimate Editor

Instead of reprocessing data from scratch, the LLM ingests new documents one by one, categorizing them, summarizing them, and interlinking them with existing knowledge. Over time, this creates a web of interconnected notes, where each new addition strengthens the whole.

"It’s like a Wikipedia that writes itself—and gets smarter with every interaction," explained a researcher following Karpathy’s work.

3. The Index.md: Your AI’s Table of Contents

Central to the system is an index.md file—a catalog of every page in the wiki. When you ask the LLM a question, it first consults the index to find relevant entries, then drills down to synthesize an answer. No vector databases, no embedding searches—just human-readable, LLM-friendly structure.

4. The Schema: The Secret Sauce

Karpathy’s genius lies in the schema (a configuration file he calls CLAUDE.md or AGENTS.md). This file defines:

• How the LLM ingests new sources.
• How it answers queries.
• How it maintains consistency across pages.

It’s not just a tool—it’s a collaboration between human and machine, evolving as both learn what works best.

The Human Touch: Why This Isn’t Just Automation

Unlike traditional AI pipelines, Karpathy’s LLM Wiki isn’t about replacing human input—it’s about augmenting it. The human role shifts from manual curation to designing the rules of knowledge organization.

"The real bottleneck isn’t the AI’s ability to process data—it’s our ability to structure it meaningfully," Karpathy noted. "This system turns the LLM into a co-author, not just a tool."

The Big Picture: A New Era for AI Knowledge

Karpathy’s project isn’t just a personal experiment—it’s a blueprint for the next generation of AI systems. By treating knowledge as a compounding asset, LLM Wiki offers:

• Self-healing archives: If a fact changes, the LLM updates the wiki automatically.
• Auditable records: Every change is logged in log.md, a chronicle of the wiki’s evolution.
• Zero-cost maintenance: No need for expensive vector databases or RAG pipelines—just Markdown, structure, and an LLM.

What’s Next? The Future of AI Knowledge Bases

Karpathy isn’t stopping here. He’s exploring:

• Generating synthetic training data from the wiki to fine-tune LLMs, embedding knowledge directly into their weights.
• Expanding the system to handle larger datasets, potentially replacing enterprise RAG setups with Markdown-based wikis.
• Open-sourcing the architecture to let others build their own LLM-powered knowledge bases.

Sources

1. Startup Fortune – Andrej Karpathy Unveils LLM Wiki, a Living Archive for AI Ideas
2. Andrej Karpathy’s X Post – LLM Knowledge Bases
3. a2a-mcp – Andrej Karpathy LLM Knowledge Bases: Building AI-Powered Wikis in Obsidian
4. Antigravity Codes – Karpathy’s LLM Wiki: The Complete Guide
5. VentureBeat – Karpathy’s LLM Knowledge Base Architecture
6. DAIR.AI Academy – LLM Knowledge Bases