← Glossary

The LLM wiki

Karpathy's pattern: the AI keeps a set of interlinked pages current as sources arrive, so knowledge is compiled once and maintained, not rebuilt on every question.

Most AI tools that work on your documents answer a question by reading the files afresh every time. The LLM wiki, a pattern Andrej Karpathy described in 2026, does something different. The model keeps a set of interlinked pages and maintains them as new material arrives, so the knowledge is compiled once and kept current rather than rebuilt on every question. You add the sources. The machine does the writing and the cross-referencing.

Two operations run it. Ingest: a new source is read against the schema and every page it touches, often ten or fifteen, is updated in one pass, with contradictions flagged rather than silently overwritten. Query: a question is answered from the pages, and any synthesis worth keeping is written back as a new, permanent page, so the next question starts further ahead. The living account plan is this pattern applied to one key customer, and the reason it beats RAG is that something accumulates.

Set out in full in: Karpathy (2026); Creating Customer Value Propositions with AI (Kogan Page, 2027), chapter 9.

Related terms

See the method run on one real account in three days: the One Page Proposition.