Karpathy LLM Wiki — Independent Convergence

This vault and Karpathy’s “LLM Wiki” pattern arrived at the same architecture independently. When two systems converge from different starting points, the shared structure is likely load-bearing.

Links: Obsidian Plugin Setup, Working With Claude, Cyborg Model, The Context Cache Hierarchy, The Three-Layer Method

Source

• Gist: karpathy/LLM Wiki (2026-04-02)
• Video walkthrough: Nate Herk — “Andrej Karpathy Just 10x’d Everyone’s Claude Code” (2026-04-05)
• Three-layer framing: Austin Marchese — “Stop Prompting Claude. Use Karpathy’s Method Instead.” (2026-06-09) — distills Karpathy at AISN 2026 into Spec → Verifier → Environment. The LLM-Wiki described on this page is the Environment layer; that video brackets it with the Spec and Verifier layers. See The Three-Layer Method.

The Shared Architecture

Both systems independently arrived at:

Where This Vault Goes Further

• Automated maintenance skills — /vault-sync updates INDEX, tags, cross-links in one pass. Karpathy’s workflow is manual.
• Domain-specific ingest — /debate-review is a structured pipeline for a specific source type. /vault-ingest dispatches by content type (article, YouTube, raw file). Karpathy’s ingest is generic.
• Cross-project transfer — /vault-to-project pushes vault research into project CLAUDE.md files. Karpathy doesn’t address how wiki knowledge flows into working codebases.
• Persistent memory layer — Claude Code’s memory/ system survives across conversations. Karpathy’s “hot cache” is a single file inside one project.
• Tag indexes as first-class objects — Tag files are pages with reverse-links, visible in graph view and queryable via Dataview’s FROM [[tags/tagname]].

Ideas Adopted From Karpathy

• raw/ folder — separation of immutable source material from processed wiki pages. Content was growing inside the vault structure; centralizing it in raw/ keeps the wiki clean and gives a clear ingest pipeline.
• Marp slide decks — markdown-to-slides for generating presentations from research content. CLI installed (marp), Obsidian plugin available.
• Web Clipper — browser extension for one-keystroke article capture to raw/. Manual path when automated fetch is bot-blocked.
• Dataview plugin — queryable vault via YAML frontmatter. Triggered migration of all 128 vault pages from bold-text metadata to YAML frontmatter.
• “File good answers back into the wiki” — explicitly naming the practice of turning query results and syntheses into vault pages, not just chat output.

Why Convergence Matters

Independent convergence from different starting points (Karpathy: research paper wikis; this vault: philosophy/economics/gaming research + debate analysis) suggests the shared patterns are structurally necessary, not stylistic choices:

1. Markdown files > vector databases — at current context window sizes, a well-indexed folder is a better knowledge representation than embeddings or RAG.
2. The LLM is the maintenance team — wikis die because humans won’t do bookkeeping. LLMs don’t get bored.
3. Index files > embedding search — at moderate scale (~100-500 pages), a maintained index is more reliable than semantic search.
4. The filesystem is the database — no infrastructure needed. Git gives you versioning, branching, and collaboration for free.

Karpathy frames this through Vannevar Bush’s Memex (1945) — a personal knowledge store with associative trails between documents. Bush’s vision was closer to this than to what the web became: private, actively curated, with the connections between documents as valuable as the documents themselves. The missing piece was who does the maintenance. The LLM handles that.

Tags

ai, epistemology