LLM Wiki

Appearance

Built 26/04/17 09:31commit 4c9ce40

Karpathy LLM Wiki Gist ​

中文 | English

Summary ​

This source defines the core LLM Wiki pattern: keep raw sources immutable, let the model maintain a persistent markdown wiki, and use a schema file to constrain ingest, query, and maintenance behavior.

Source ​

Key Contributions ​

  • Reframes the system from query-time retrieval to a maintained synthesis layer that compounds over time.
  • Defines a three-layer architecture: immutable raw sources, LLM-maintained wiki pages, and a schema file that governs behavior.
  • Treats index.md and log.md as operational primitives rather than incidental files.
  • Makes ingest, query, and lint explicit maintenance loops instead of ad hoc chat behaviors.
  • Positions the human as curator and question-asker while the LLM handles bookkeeping, cross-links, and consistency work.

Strongest Claims ​

  • The wiki should sit between the user and the raw documents, so knowledge is compiled once and then maintained.
  • Good query outputs should be filed back into the wiki instead of disappearing into chat history.
  • Maintenance burden is the bottleneck for personal and team knowledge bases; LLMs lower that burden enough to keep the system current.
  • Small helper tools are optional and should appear only when the wiki outgrows simple index-first navigation.

Practical Implications For This Vault ​

  • Prefer integrating new source material into existing concept and topic pages before creating new pages.
  • Keep raw/ read-only and use source pages to capture what each source changes in the durable synthesis.
  • Read index.md first during query work so page discovery stays cheap before introducing heavier search.
  • Treat lint as editorial maintenance: contradictions, stale claims, and missing pages matter more than pure link hygiene.