A personal knowledge base for building deep, compounding expertise in AI and technology. Built on the LLM Wiki pattern by Andrej Karpathy.

Raw sources go in. A structured, interlinked wiki comes out. Knowledge accumulates permanently instead of being rediscovered from scratch.

The core idea is simple: instead of asking an AI questions and forgetting the answers, the AI writes and maintains a wiki on your behalf.

You drop articles into raw/
        ↓
Claude reads, extracts, and structures knowledge
        ↓
Wiki pages are created in wiki/ with full wikilinks
        ↓
You browse, query, and build on it in Obsidian
        ↓
Everything is version-controlled on GitHub

No RAG. No vector databases. Just markdown files that compound over time.

vault/
├── .claude/
│   └── commands/           # Slash commands for Claude Code
│       ├── ingest.md       # /ingest
│       ├── lint.md         # /lint
│       ├── query.md        # /query
│       ├── report.md       # /report
│       └── update.md       # /update
├── raw/                    # Drop source articles here
│   ├── archive/            # Processed sources moved here automatically
│   └── assets/             # Images and attachments
├── wiki/
│   ├── concepts/           # Algorithms, architectures, techniques
│   ├── entities/           # People, labs, models, papers
│   ├── synthesis/          # Cross-topic analyses and open questions
│   ├── index.md            # Master catalog of all pages
│   └── log.md              # Chronological record of all operations
├── output/                 # Generated reports and briefings
├── CLAUDE.md               # Agent config and operating rules
└── README.md               # This file

Core:

• Machine learning fundamentals (transformers, diffusion, RL, optimisation)
• Large language models (architecture, training, alignment, inference)
• AI agents and tool use (scaffolding, memory, multi-agent systems)
• AI infrastructure (compute, distributed training, serving, quantisation)
• Frontier labs (Anthropic, OpenAI, DeepMind, Meta AI, Mistral)
• Emerging techniques (MoE, RLHF, DPO, RAG, speculative decoding)

Adjacent:

• Hardware as it relates to AI capability
• AI safety and alignment research
• Regulation and policy

• Obsidian — for browsing the wiki and graph view
• Claude Code — the AI agent that runs the wiki
• Node.js — required for Claude Code
• A GitHub account — for version control

1. Clone this repo

git clone https://github.com/yeewliang/AI-LLM-Wiki.git
cd AI-LLM-Wiki

2. Install Claude Code

npm install -g @anthropic/claude-code

3. Open the vault in Obsidian File �� Open Vault → select this folder

4. Open Claude Code in the vault directory

claude

Claude reads CLAUDE.md automatically on every session.

Drop any article, PDF, or note into raw/. Name it with a date prefix:

raw/2026-05-02-attention-is-all-you-need.md

Use the Obsidian Web Clipper browser extension to clip articles directly into raw/ with automatic date-prefixed filenames.

/ingest

Claude processes all new files in raw/, builds wiki pages, updates the index, archives the sources, and commits to GitHub.

/query what are the trade-offs between RLHF and DPO
/query what do my sources say about speculative decoding

Answers come only from your own ingested knowledge, with citations.

/report transformer inference optimisation
/report state of AI agents 2026

Saved to output/ as a timestamped markdown file.

/lint

Run monthly. Fixes broken wikilinks, orphan pages, missing frontmatter, and stale entity pages.

Open Obsidian and use:

• Graph View — see all concepts and entities as a connected graph. Clusters reveal the shape of your knowledge. Gaps appear as thin or isolated nodes.
• Wikilinks — follow [[concept]] links between pages like a personal web.
• Search — full-text search across everything you've ever ingested.
• wiki/index.md — the master catalog, organised by type.

• Andrej Karpathy's LLM Wiki gist
• second-brain by NicholasSpisak
• Vannevar Bush's Memex (1945) — the original vision for associative knowledge