Karpathy LLM Wiki — comparison with our KB
Source pattern: LLM Wiki — Karpathy source. Companion: media-pipeline ingest workflow.
The pattern in one line
Karpathy’s LLM Wiki replaces query-time RAG rediscovery with a persistent, LLM-maintained, interlinked markdown wiki between you and the raw sources: ingest compiles once and keeps current; answers are filed back; the wiki compounds. Three layers — immutable raw sources, an LLM-owned wiki, and a schema doc that disciplines the maintainer.
Comparison
| Axis | Karpathy LLM Wiki | cos.hodgederrick.com (ours) |
|---|---|---|
| Synthesis | LLM updates entities on ingest, flags contradictions | Content authored (human/agent); no auto cross-page update on ingest |
| Raw layer | Immutable raw/ + web clipper |
media-pipeline (ingest.hodgeluke.com) — R2 originals, derived markdown/transcripts |
| Compounding | Queries filed back as pages | Strong contract, but filing is manual (this page is an instance) |
| Query | LLM reads index.md → pages → cites |
MCP search_docs (Orama) + contentTypes/facets + .md mirrors + llms.txt |
| Lint | Contradictions/orphans/stale passes | Build-strict frontmatter + dangling-edge hard fail; orphan report added to graph export |
| Log | LLM-maintained log.md (append-only) |
log.md generated from git history each build |
| Graphview | Obsidian graph | knowledge-graph.json nodes/edges + facets |
| Delivery | local Obsidian + git | public, CDN, CI-verified, MCP-served |
Deltas adopted
- Raw layer = media-pipeline. Uploads land as immutable originals in R2; audio → Whisper v3 Turbo transcript (+VTT), PDFs → Firecrawl parse → Markdown. That is exactly Karpathy’s “raw sources” bucket, outside the git repo.
- Schema doc — this repo’s
SCHEMA.md= the third layer (the maintainer-contract); updated with ingest rituals from the media pipeline. - Lint/health — graph export now reports type counts, degree stats and orphan nodes; CI fails on dangling edges already.
log.md— build-generated, git-derived, greppable (## [yyyy-mm-dd]).- Answer-filing — analyses like this one and the Flue review are filed
into
professional/ideas/on the same tick.
Remaining gap
The true compounding loop — “source lands, related pages update, log entry, cross-references refreshed” — still runs as a manual/tool-assisted ritual (SCHEMA.md defines it), not an automated Cloudflare workflow. When the S1 cleanup pass and note-extraction stage are wired into media-pipeline (meeting → transcript → cleaned notes → frontmatter-ready drafts), the loop closes end to end: upload once, KB updates itself.