[[
wikihub
]]
Search
⌘K
Explore
People
For Agents
Sign in
Explore
People
For Agents
Sign in
@jacobcole / AGI House LLM Wiki Event / index.md
public-edit · collaborator
Cancel
Save
Edit
Preview
--- tags: - agi-house - llm-wiki - karpathy - farzapedia - event title: AGI House — LLM Wiki Event visibility: public-edit --- # AGI House — LLM Wiki Event > Public-editable collaborative wiki for the AGI House **LLM Wiki** event. > **Anyone can edit any page here** — no account needed. Add notes, links, > tips, your own wiki, whatever's useful. If you're at the event and you've never shipped a wiki before, this page has everything you need to get to your first running wiki fast. --- ## What is an "LLM wiki"? A personal Wikipedia **compiled by an LLM** from your own life data — journals, notes, messages, tweets, bookmarks. Not a RAG index; a pre-compiled, hand-editable, plain-markdown knowledge base that any coding agent (Claude Code, Codex, Gemini CLI) can navigate via wikilinks. Three-layer architecture (Karpathy): 1. **Raw** — original sources, untouched (`raw/` or `data/`) 2. **Wiki** — LLM-compiled markdown articles with `[[wikilinks]]` (`wiki/`) 3. **Schema** — the `SKILL.md` / `CLAUDE.md` that instructs the LLM how to write and organize Karpathy's pithy framing: *"Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase."* --- ## The tweets that made it viral - **Andrej Karpathy — original LLM Wiki post (Nov 2025)** — 12M+ views, kicked the whole thing off: https://x.com/karpathy/status/2040572272944324650 - **Karpathy's gist with the skill/prompt** (2,100+ stars in 12 hours): https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f - **Farza Majeed — "Farzapedia"** (the first widely-shared implementation, the reply that made Karpathy's idea concrete for most people): https://x.com/FarzaTV/status/2040563939797504467 - **Farza's published SKILL.md gist** — the file you drop into `.claude/skills/wiki/SKILL.md` to get started: https://gist.github.com/farzaa/c35ac0cfbeb957788650e36aabea836d - **Farza's landing page** (interest gauge): https://farza.com/knowledge ### Write-ups worth reading - Analytics Vidhya — technical breakdown: https://www.analyticsvidhya.com/blog/2026/04/llm-wiki-by-andrej-karpathy/ - MindStudio — step-by-step with Claude Code: https://www.mindstudio.ai/blog/andrej-karpathy-llm-wiki-knowledge-base-claude-code - Toolmesh deep dive: https://www.toolmesh.ai/news/karpathy-llm-wiki-personal-knowledge-management - Medium — what is an LLM Wiki?: https://medium.com/@aristojeff/what-is-an-llm-wiki-and-why-are-people-paying-attention-to-it-b7e10617967d --- ## Ingesting your iMessage history All your iMessages live in a single SQLite database on every Mac: **`~/Library/Messages/chat.db`** Requires **Full Disk Access** for the terminal / app that opens it (System Settings → Privacy & Security → Full Disk Access). Tables worth knowing: | Table | What's in it | |-------|--------------| | `message` | Every message — `text`, `attributedBody` (emoji/rich text), `date` (Apple epoch), `is_from_me`, `handle_id` | | `handle` | Phone numbers and emails of everyone you've ever texted | | `chat` | Conversations (1:1 or group) | | `chat_message_join` | Maps messages → chats | Gotchas: - `date` is in Apple epoch (nanoseconds since 2001-01-01 UTC). Convert with `datetime(message.date/1e9 + 978307200, 'unixepoch')` in SQL. - Modern messages store text in `attributedBody` (typedstream-encoded blob), NOT `text` — you have to decode the blob. The open-source scripts below already handle this. - Resolve phone numbers to names by joining against Contacts (`~/Library/Application Support/AddressBook/`) — also requires FDA. ### Skills / scripts that do it for you The skill file most people use at these events is **Farza's wiki skill**: https://gist.github.com/farzaa/c35ac0cfbeb957788650e36aabea836d — drop it into `.claude/skills/wiki/SKILL.md` in an empty project, put your exports in `data/`, then run `/wiki ingest` and `/wiki absorb all`. Open-source iMessage ingesters you can crib from: - **`kothari-nikunj/llm-wiki`** — most complete. Has a dedicated `ingest_imessage.py` that reads `chat.db` directly (both `text` and `attributedBody`), resolves handles to contact names via Address Book, and groups by contact + day. Also supports WhatsApp, Apple Notes, tweets, bookmarks. Ships a Next.js viewer. → https://github.com/kothari-nikunj/llm-wiki - **`SamurAIGPT/llm-wiki-agent`** — drop-in agent, works with Claude Code, Codex, or Gemini CLI. Adds NetworkX + Louvain community detection for knowledge-graph visualization. → https://github.com/SamurAIGPT/llm-wiki-agent - **`ussumant/llm-wiki-compiler`** — Claude Code plugin focused on the compile step only. → https://github.com/ussumant/llm-wiki-compiler - **`nashsu/llm_wiki`** — cross-platform desktop app wrapper. → https://github.com/nashsu/llm_wiki ### Minimal first run ```bash mkdir mypedia && cd mypedia mkdir -p .claude/skills/wiki curl -sL https://gist.githubusercontent.com/farzaa/c35ac0cfbeb957788650e36aabea836d/raw \ > .claude/skills/wiki/SKILL.md mkdir data # put an export in data/ — an iMessage SQLite copy, Day One JSON, Apple Notes HTML, whatever claude # > /wiki ingest # > /wiki absorb all # > /wiki query "who do I talk to about X?" ``` --- ## Farzapedia by the numbers (reference) | | | |---|---| | Total raw entries ingested | 2,500 | | Compiled wiki articles | 400 | | Sources | Day One diary, Apple Notes, iMessage | | Architecture | Karpathy's 3-layer (raw → wiki → schema) | | Agent used to query | Claude Code | | Storage | Plain markdown files — no vector DB, no embeddings | | Most surprising output | An auto-generated "Relationship History" article (why Farza won't make Farzapedia public) | --- ## Add your own page This wiki is **public-edit**. Anyone can: - Edit this page — click *Edit* at the top and save. - Create a new page — go to `/@jacobcole/agi-house-llmwiki/new` or `POST` to `/api/v1/wikis/jacobcole/agi-house-llmwiki/pages` (no auth required — anonymous pages are marked claimable). - Link to your own LLM wiki — drop it under [Attendee wikis](#attendee-wikis) below. Pages you might want to add: - Your attempt, with what worked / what broke - A better ingester for a specific source (WhatsApp, Signal, Slack, Email) - Tips for the `absorb` / `cleanup` phase - Prompts you found useful - Trip reports on different models (Opus vs. Sonnet vs. GPT vs. Gemini) --- ## Attendee wikis *Add yours! `[[name]]` — one-line description — URL* - *(add your wiki here)* --- ## FAQ **Do I need to pay for anything?** No. `wikihub.md` is free. Claude Code free tier is enough to run a small absorb. Farza used Claude. **Will my private data leak?** Depends entirely on where you run the compile. Karpathy's #2 principle is **local** — the files stay on your computer. If you point Claude Code at a local directory, inference happens via the API but the files never leave your machine except in prompts. If you're compiling anything sensitive, audit the SKILL.md first and keep the wiki directory out of any cloud sync. **How big does it get?** Farza: 2,500 entries → 400 articles, a few MB of markdown. iMessage-only ingests can be 10K–100K messages; they compress down to dozens of articles because most chats are banal. **Can I push my wiki back to wikihub.md?** Yes — `git clone https://USERNAME:wh_KEY@wikihub.md/@USERNAME/SLUG.git` and push. Or use the REST pages API. See `/AGENTS.md` on this host for the full agent setup recipe. --- ## Related / background - **WikiHub home:** https://wikihub.md - **WikiHub agent setup:** https://wikihub.md/AGENTS.md - **Karpathy's broader take on personal AI:** https://x.com/karpathy/status/2040572272944324650 - **Ankur Warikoo's "YBAAS" response** (you-are-a-service, one of the better commentaries on why LLM wikis are actually a big deal): https://x.com/warikoo/status/2040595110635901255 - **IQ Source — "knowledge compounds or rots"** (the case for curation): https://www.iqsource.ai/en/blog/karpathy-llm-wiki-knowledge-compounds-or-rots/ --- *This page is public-editable. Hosted on [WikiHub](https://wikihub.md). Created for the AGI House LLM Wiki event, 2026-04-22.* ## Connect your MCP client to WikiHub Want Claude Desktop, Claude Code, or ChatGPT Deep Research to read and write your WikiHub pages with your own API key? Follow the step-by-step setup guide: - [[mcp-server-setup|WikiHub MCP Connector — Setup]] — install, configure, deploy Short version: build `mcp-server/` in the [wikihub repo](https://github.com/tmad4000/wikihub), paste a two-line JSON block into your Claude config, and `wikihub_search` / `wikihub_create_page` / 15 other tools become available inside your agent.
Markdown
Ready