Summary
A two-part essay arguing that Karpathy's large language model (LLM) Wiki pattern solves two long-standing knowledge management failures simultaneously: the statelessness of retrieval-augmented generation (RAG) — which re-derives the same synthesis from scratch on every query — and the maintenance collapse of human-curated wikis (where bookkeeping burden grows faster than value). The proposed architecture inserts an LLM-maintained, compounding layer of markdown files between raw sources and queries. Trey Ditto, running a public relations (PR) firm, demonstrates the pattern applied to client-specific institutional memory.
Key Claims
- RAG's structural flaw is statelessness. Every query retrieves and synthesizes from scratch. No insight accumulates between sessions. If a complex answer requires synthesizing five documents, that synthesis is performed anew every time — and discarded.
- Human-maintained wikis fail from scaling mismatch. The bookkeeping burden of cross-referencing, tagging, and noting contradictions grows faster than the value produced. Users abandon them.
- The LLM Wiki inserts a compounding pre-compilation layer. Instead of retrieving at query time, the LLM pre-compiles source material into interlinked concept pages. Each ingest enriches the structure permanently. Knowledge compounds across sessions rather than evaporating.
- Karpathy's mental model: "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase." The user curates inputs and reviews outputs; the LLM performs all structural maintenance.
- The schema file (CLAUDE.md) is the operating manual. It defines page formats, linking conventions, ingestion workflow, and how the agent should answer questions. It is where the user's intelligence is imprinted on the system.
- Ingest updates 10–15 related pages per source. The value of an ingest is not the per-source summary but the cross-linking — enriching existing concept pages with new connections, adding backlinks, and logging the action.
- Queries can generate new pages. If a new connection is discovered during a query session, the LLM files it back into the wiki as a permanent page. The query operation is not read-only.
- Lint is a health-check operation: hunting for broken links, stale claims, contradictions, and orphan pages. Run periodically rather than per-ingest.
- Auditability is a core advantage over RAG. "The knowledge lives in local, human-readable files. You can open any page, verify claims, and improve the system rules without ever touching the underlying sources or getting locked into a proprietary application. You're building a system you own, not renting a service."
- Ditto's institutional memory application. Client-specific wikis compiled from market reports, competitor analyses, and media coverage. New team members receive curated, interlinked understanding rather than raw document dumps. The defensibility claim: "Models generalize. Businesses specialize. Defensibility lives in that gap."
- The pattern marks a maturity shift. "We are moving away from treating LLMs purely as search engines or text generators, and finally starting to use them as tireless librarians and system maintainers."
Concepts Introduced or Extended
Entities Mentioned
Quotes
"Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase."
"The knowledge lives in local, human-readable files. You can open any page, verify claims, and improve the system rules without ever touching the underlying sources or getting locked into a proprietary application. You're building a system you own, not renting a service."
"Models generalize. Businesses specialize. Defensibility lives in that gap."
"You don't avoid competition by being better. You avoid it by being hard to replicate."
"We are moving away from treating LLMs purely as search engines or text generators, and finally starting to use them as tireless librarians and system maintainers."
My Take
The article is advocacy writing with no empirical claims, but the core architectural insight is sound and directly operational — this wiki is itself an instance of the pattern described. The strongest argument is the auditability one: a human-readable, locally-owned markdown wiki is inspectable and correctable in a way that a vector database and retrieval pipeline simply isn't. The weakest part is the Ditto section, which is more testimonial than analytical. The article does not address cost, latency, token budget, multi-user collaboration, or how to handle sources that update over time — all real operational considerations. It also does not cover the failure modes of the LLM Wiki itself (model hallucination baked into permanent pages, schema drift, page proliferation without pruning).