Smorster

Why LLM Wiki? 🧠 Future Of Knowledge For Agentic AI & Humans

Wanderloots · 11m 19s · Watch on YouTube · 9 sources

Decision Card

Effort: Weekend project — install Obsidian (free), clip a dozen sources into a vault with the Web Clipper, and point an AI agent at a wiki/ folder with a “read sources → write interlinked markdown pages” prompt. A first useful version is a few hours; the payoff only shows after weeks of consistent note-taking.

Honest take: The video conflates three distinct things — Obsidian’s auto-generated link graph, formal knowledge-graph RAG (Microsoft’s GraphRAG), and Karpathy’s LLM Wiki — under one “knowledge graph” banner, when only the last is what the title actually promises, and the “agent that builds and maintains it automatically” is asserted as already-working but never demonstrated on screen. It’s also a soft sell: the practical setup (the part you’d actually need) is deferred to a hypothetical future video.

Concrete next steps:

  • Read Karpathy’s original LLM Wiki gist (~10 min) to get the pattern from the source, not the summary: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
  • Build a throwaway 10-note vault in Obsidian and write with [[wiki-links]] deliberately for a week (~1 hr setup + daily habit) to feel whether the graph actually surfaces connections for you.
  • Skip if you don’t already take notes regularly — the video’s own thesis is that the graph is a byproduct of consistent note-taking, so there’s nothing to automate over an empty vault.

TL;DR

A former IP lawyer explains knowledge graphs (nodes, edges, triples) using Google, Wikipedia, and Obsidian, then pitches Andrej Karpathy’s “LLM Wiki”: a separate AI-maintained vault of interlinked markdown that gives all your AI tools a shared, persistent brain instead of siloed per-tool memory. It’s a conceptual overview grounded in real sources (GraphRAG, Karpathy’s gist) but stops short of showing the actual automated build.

Key Points

  • A knowledge graph reduces to three primitives: a node (a thing/concept), an edge (a named relationship), and a triple (subject–relationship–object) 01:07
  • Google’s search side-panel and Wikipedia’s inter-article links are real-world knowledge graphs you already use daily 01:50
  • In Obsidian, wrapping text in [[double brackets]] auto-creates a node and edge, so the graph emerges from note-taking rather than being deliberately drawn 03:36
  • The author keeps two separate vaults — a “human vault” for his own thinking and an “LLM vault” for AI-generated knowledge — to track provenance (what came from him vs. AI) 09:38
  • Standard RAG (retrieval augmented generation) converts notes to vectors and retrieves similar chunks; it works for “what is X” but fails when answers live between documents 07:18
  • Graph RAG follows relationships between sources instead of retrieving thousands of chunks, and the author claims it “significantly outperforms RAG” on large complex datasets 07:58
  • Each AI tool’s built-in memory is siloed and “fails when you’re switching between tools” — an LLM Wiki is the shared structure that fixes this 08:21
  • The LLM Wiki concept gained traction via an article by Andrej Karpathy (ex-OpenAI, coiner of “vibe coding”) 08:48
  • The LLM Wiki has three layers: untouched raw sources, the AI-compiled interlinked wiki, and periodic AI maintenance that checks for contradictions, outdated info, and orphan pages 09:56

Notable Quotes

“A triple is the atom of a knowledge graph: subject, relationship, and object. That’s the entire model: two things and one connector.” 01:20

“I didn’t try to build the graph. I just wrote about the relationship between different concepts. The knowledge graph is just what happens when you’re specific about how you take notes.” 04:27

“The knowledge isn’t lost, it’s just trapped in the silos of each tool.” 08:33

Verified Claims

Karpathy described the LLM Wiki as an LLM incrementally building/maintaining a persistent, interlinked collection of markdown files that sits between you and raw sources 08:57

Andrej Karpathy coined the term “vibe coding” 08:55

Karpathy is formerly from OpenAI 08:53

On larger, complex datasets, graph RAG significantly outperforms naive RAG 07:58

  • Sources: Microsoft Research: GraphRAG blog, GraphRAG project site
  • Verdict: Confirmed — Microsoft reports 70–80% win rates over naive RAG on comprehensiveness/diversity for complex, multi-document tasks. (Caveat: the video presents this as universal; the advantage is dataset- and query-dependent, and GraphRAG is more expensive to build.)

RAG converts notes into numbers (vectors) and retrieves the chunks most similar to your question 07:22

  • Sources: IBM: What is GraphRAG?
  • Verdict: Confirmed — accurate description of standard embedding-based retrieval.

Google and Wikipedia function as knowledge graphs 01:55

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. What does the “agent that automatically builds and maintains” the LLM Wiki actually look like in practice — what prompts, MCP servers, or scripts drive the read-extract-integrate loop, and how reliable is the contradiction/orphan detection?
  2. For a personal-scale vault (hundreds to low thousands of notes), does Karpathy’s compile-once LLM Wiki genuinely outperform just feeding raw markdown into a long-context model — or is the GraphRAG advantage only material at corpus sizes most individuals never reach?
  3. How do you prevent the AI-maintained vault from drifting — accumulating subtly wrong syntheses or hallucinated cross-links that then poison every downstream AI tool that trusts the shared brain?

Sources