Smorster

Andrej Karpathy Just 10x’d Everyone’s Claude Code

Nate Herk | AI Automation · 17m 47s · Watch on YouTube · 13 sources

Decision Card

Effort: 30–60 minute setup — install Obsidian, create a vault, paste Karpathy’s gist into Claude Code, then budget ~10–15 minutes of agent runtime per batch ingest.

Honest take: The video’s own numbers undercut the “5 minutes” framing — a single article ingest took 10 minutes and 36 transcripts took 14 minutes of Claude Code runtime, and none of that token cost is quantified; the presenter also concedes the pattern breaks down beyond hundreds of documents, and the “95% token reduction” figure is a single unverified X anecdote, not a benchmark.

Concrete next steps:

  • Read Karpathy’s original idea file at gist.github.com/karpathy/442a6bf555914893e9891c11519de94f (~10 min) — it is the entire “setup”, the video adds only Obsidian as a viewer.
  • Bootstrap a test vault: create raw/ + wiki/ folders, paste the gist into Claude Code with a one-line statement of the vault’s purpose, ingest 2–3 documents (~30–45 min including agent runtime).
  • Optionally install Obsidian Web Clipper and point its save location at raw/ for one-click article capture (~10 min).
  • Skip if your corpus is already in the millions-of-documents range or needs multi-user/enterprise access — the presenter himself says traditional RAG wins at that scale.

TL;DR

Nate Herk walks through implementing Andrej Karpathy’s viral “LLM wiki” idea: paste Karpathy’s gist into Claude Code and it scaffolds a plain-markdown knowledge base (raw/ sources, wiki/ pages, index, log, CLAUDE.md) that the LLM auto-organizes, cross-links, and lints — no vector database or embeddings. He demos it on his 36 YouTube transcripts and the AI 2027 article, uses Obsidian purely as a graph/markdown viewer, and closes with an honest caveat that the pattern suits hundreds of well-indexed pages, not millions of enterprise documents.

Key Points

  • The demo system organized 36 YouTube video transcripts into an interlinked wiki with tags, backlinks, and per-video takeaway pages — with zero manual relationship building; Claude Code inferred all links itself. 00:02, 01:07
  • The architecture is deliberately primitive: a vault containing a raw/ folder for source documents, a wiki/ folder for LLM-generated pages, plus an index (catalog of all pages), a log (operation history), and a CLAUDE.md explaining how to search and update. 03:44, 04:27
  • Karpathy reported he expected to need “fancy RAG” but found the LLM auto-maintains index files and summaries well at his scale of roughly 100 articles. 02:43
  • Setup is prompt-only: copy Karpathy’s gist into Claude Code with a short instruction (“You are now my LLM Wiki agent…”) — there is no repo to clone or infrastructure to deploy. 05:11, 07:15
  • Obsidian is optional and used only as a free front end for browsing markdown and watching the relationship graph form in real time; the knowledge lives in plain files. 02:12, 06:08
  • Ingesting one article (AI 2027) took about 10 minutes and produced 23 interlinked wiki pages (source, six people, five organizations, concepts, analysis); the 36-transcript batch took ~14 minutes. 11:47, 11:59
  • Other projects can consume the vault: his executive assistant’s CLAUDE.md points at the wiki path with instructions to read the index and a “hot cache” file (a ~500-word recency buffer) only when needed, which he says reduced token usage versus his old context-files approach. 13:25, 14:26
  • Maintenance is “linting”: periodic LLM health checks that find inconsistent data, impute missing data via web search, and suggest new article candidates. 15:04
  • On “does this kill RAG”: no — the wiki wins on simplicity and relationship depth for hundreds of pages, but traditional semantic search / knowledge-graph pipelines still win at millions of documents. 15:34, 17:10

Notable Quotes

“I thought that I had to reach for fancy rag, but the LLM has been pretty good about auto maintaining index files and brief summaries of all documents” 02:43

“this method using Karpathy’s LLM wiki makes knowledge compound like interest in a bank” 03:23

“You literally just say, ‘Hey, Claude code, read this idea from Andre Karpathy and implement it.’” 05:15

Verified Claims

  • Claim: Karpathy recently posted about using LLMs to build personal knowledge bases, and it got major traction on X. 01:41

  • Claim: Karpathy said he thought he’d need “fancy RAG” but the LLM auto-maintains index files and summaries at small scale. 02:43

  • Claim: Karpathy’s scale is about 100 articles and about half a million words. 02:55

  • Claim: Karpathy followed up with a gist version of the idea, intentionally left vague so people can customize it. 05:41

    • Sources: karpathy/llm-wiki gist (“intentionally kept a little bit abstract/vague because there are so many directions to take this in”)
    • Verdict: Confirmed.

  • Claim: The structure is raw sources + LLM-built wiki + index + log + CLAUDE.md, with ingest/query/lint operations and no vector database. 03:44

    • Sources: karpathy/llm-wiki gist
    • Verdict: Confirmed (matches the gist’s three-layer architecture and three operations, including lint health checks mentioned at 15:04).

  • Claim: An X user turned 383 scattered files and 100+ meeting transcripts into a compact wiki and cut Claude token usage by 95%. 04:53

    • Sources: MindStudio write-up repeating the claim
    • Verdict: Inconclusive — the figure circulates in secondary blog coverage of the same anecdote; no primary benchmark or original post independently verified.

  • Claim: Obsidian is a completely free tool downloadable at obsidian.md. 06:10

    • Sources: obsidian.md
    • Verdict: Confirmed (the app is free, with optional paid Sync/Publish add-ons).

  • Claim: The Obsidian Web Clipper browser extension saves web articles as markdown directly into a vault, defaulting to a “Clippings” folder that can be changed. 08:56, 10:26

  • Claim: The AI 2027 article features people named Eli, Thomas, and Daniel and entities like OpenBrain and a “superhuman coder.” 08:37, 11:35

    • Sources: ai-2027.com (authors include Daniel Kokotajlo, Eli Lifland, Thomas Larsen; the scenario’s fictional lab is OpenBrain and its Agent-3 milestone is a superhuman coder), AI 2027 summary
    • Verdict: Confirmed.

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. At what corpus size does the wiki pattern actually degrade — what are measured query-accuracy and token-cost curves as the vault grows from 100 to 10,000 pages, and can hierarchical sub-indexes push the crossover point with RAG further out?
  2. How reliable is unsupervised LLM ingestion over time — what error rate (wrong links, hallucinated relationships, contradictory pages) accumulates across batches, and how much does periodic “linting” actually catch versus miss?
  3. Could this repo’s own pipeline (playlist → transcript → summary) write its outputs into a Karpathy-style raw/ + wiki/ vault so summaries compound into a queryable research base rather than sitting as isolated files?

Sources