firehose> #llmops
video youtu.be Watch ↗

Fable 5 + Karpathy's LLM Wiki is Basically Cheating

TL;DR

Nate Herk walks through Andrej Karpathy's "LLM wiki" pattern: point an LLM at a folder of sources and have it compile a personal knowledge base as plain markdown in Obsidian. Sources land verbatim in raw/; the agent reads them and compiles a wiki/ of cross-linked pages (summaries, concepts, entities, sources, topics), maintaining an index.md and an append-only log.md, all governed by a CLAUDE.md that owns structure, ingestion, and lint rules. The load-bearing idea — lifted straight from Karpathy's gist — is that instead of RAG re-retrieving and re-synthesizing from raw documents on every query (so nothing accumulates), the LLM compiles knowledge once into a persistent interlinked store and keeps it current. The payoff isn't the ingest, it's the cross-source connection: dropping in a "Fable 5 / Mythos 5 system card" PDF and an OpenAI "GPT-5.6 Sol" article together surfaced a link between them "worth having as a wiki instead of two separate summaries" — a nuance a per-source summary structurally can't produce. Because it's "just markdown files with routing," the store is portable to any agent; keep it flat when the agent should breadth-search everything and let folder structure emerge only when a source naturally splits; and split model tiers — the presenter notes a premium model (he uses Fable 5) is "overkill for the ingest," which a cheaper tier handles, reserving the smart model for the synthesis/framing step.

Concepts introduced

Held, not dropped (touched but not yet warranting their own page):

Key claims

Why this corroborates the second-brain cluster

The dominant stance is corroborates: this is a third independent source (after Austin Marchese and Karpathy's own gist, both already in the graph) converging on the same raw/-vs-wiki/ split, the CLAUDE.md-as-router mechanism, and the "just markdown files, portable" thesis of AI Second Brain and Context Routing. It does not duplicate those pages — it strengthens them and adds detail (the index/log maintenance surface, the "Claude Code is a router" framing).

Two secondary stances sit inside it. It is novel in exactly one place: the Karpathy gist's compile-once-vs-retrieve-per-query thesis had no page, so Compiled Knowledge Base is spun out to hold it and the cross-source-connection payoff. And it builds_on Retrieval Maturity Levels by giving a concrete Level-2 walkthrough and Query-Shaped Storage by sharpening flat-vs-structured into an earned-by-the-data decision.

Worth flagging for the operator: the generated CLAUDE.md schema (t≈09:49) is nearly a mirror of firehose's own spine — markdown as the durable artifact, verbatim capture separated from synthesized knowledge, contradictions held in a Conflict: block rather than silently overwritten, and claims attributed to sources with the agent's own reasoning marked as inference. An outside practitioner independently reinventing this vault's faithfulness and held-not-dropped rules is corroboration of the engine's design, not just its content.

A note on the model/price/date claims: this capture is unusually dense with post-cutoff-sounding assertions (Fable 5, GPT-5.6 Sol, Mythos, specific prices and dates). Per the distiller's lane they are recorded as the source's claims and flagged for a later grounding pass — not confirmed, not waved off as fictional. A stale prior is not ground truth.

Illustrated walkthrough

Linked from