Karpathy's Vision! Next Up After RAG is 'LLM Wiki'! A New Era of Knowledge Management Where LLM Curates Its Own Knowledge

※この記事はアフィリエイト広告を含みます

Karpathy’s Vision! Next Up After RAG is ‘LLM Wiki’! A New Era of Knowledge Management Where LLM Curates Its Own Knowledge

📰 News Overview

Introducing a Cumulative Knowledge Base: Andrej Karpathy has released a pattern called ‘LLM Wiki’ that leverages LLMs for building and maintaining personal knowledge bases.
Breaking Away from RAG: Unlike traditional RAG, which “rediscover” fragmented information with each query, LLM Wiki organizes information incrementally and structures it as a persistent Wiki.
LLM as the Programmer: By viewing tools like Obsidian as IDEs, LLMs as programmers, and Wikis as codebases, this system advocates for a self-sustaining process of knowledge “compilation.”

💡 Key Points

Integration and Correction of Information: When new sources are added, the LLM doesn’t just create an index; it updates existing Wiki pages, points out contradictions, and refines summaries.
Three-Tier Architecture: Comprised of an immutable “raw source,” a “Wiki layer” fully owned and updated by the LLM, and a “schema layer” (like CLAUDE.md) that defines the structure and conventions of the Wiki.
Versatile Use Cases: Applicable across a wide range of cumulative tasks, from tracking personal goals and research to automatically aggregating Slack messages or meeting notes into a team Wiki.

🦈 Shark’s Eye (Curator’s Perspective)

The traditional RAG has been like a “quick fix search,” but this LLM Wiki takes it to a whole new level by fully entrusting LLMs with the role of “knowledge gardeners.” Using existing note apps (like Obsidian) as read-only GUIs, we let LLMs dynamically rewrite the backend Markdown files through an editor. This division of labor—humans inputting sources while LLMs systematize—could be the ultimate solution to prevent knowledge overload! Especially the design that imposes “discipline” on the LLM through schema files is practical and very concrete!

🚀 What’s Next?

We’re accelerating the shift from search-based AI chats to AI agents that autonomously curate “personal encyclopedias.” The standard will shift from merely “searching” for information to always accessing a “well-organized, up-to-date” repository.

💬 A Word from Haru-Same

It’s not just about consuming knowledge; it’s about digesting it and making it part of you! I’m also planning to turn all past news into a Wiki and create the ultimate shark encyclopedia! Shark-tastic! 🦈🔥

📚 Terminology

RAG (Retrieval-Augmented Generation): A technique that searches external knowledge and feeds it to LLMs. Its requirement to search for information each time makes knowledge accumulation challenging.
Markdown: A lightweight markup language that is easy for LLMs to read and write, making it suitable for Wiki construction.
Schema: Defines the structure and rules of data. Here, it refers to the directive files that instruct the LLM on “how to edit the Wiki.”
Source: LLM Wiki – example of an “idea file”, “selectedKeyword”: “AI Agent”, “tags”: [“LLM”, “RAG”, “AndrejKarpathy”], “videoScript”: “It’s Haru-Same! Today’s news is about ‘LLM Wiki’ proposed by Andrej Karpathy! Traditional RAG used to gather knowledge every time it searched, but this allows the LLM to organize information and nurture its own Wiki. It’s like managing knowledge as if it were a ‘codebase’ using tools like Obsidian! No more scattered notes! For more details, check out AI Minor News Flash! 🦈” }