Karpathy LLM Wiki Setup Guide

The foundational architecture Kevin's wiki is built on. Karpathy's original gist on LLM Knowledge Bases is the highest-engagement bookmark in Kevin's corpus - 54,109 likes, 98,708 bookmarks. Nick Spisak's weekend setup guide added 3,574 likes and 9,339 bookmarks. Combined: 153K+ engagements.

The Origin

Andrej Karpathy published a gist titled "LLM Knowledge Bases" describing a pattern he found useful: flat markdown files, no vector DB, no embeddings - just structured text that LLMs can read directly. The key insight is that LLMs don't need fancy retrieval systems. They need well-organized, schema-readable markdown. The gist went massively viral - 54K likes, 98K bookmarks - making it one of the most-saved technical posts in X history. Source: X/@karpathy, 2026-04-02

Nick Spisak (@NickSpisak_) built an updated setup guide the following weekend, walking through the practical steps: grab Karpathy's gist, add summarize/YouTube/Obsidian/qmd tooling, then wrap the workflow in a local brain CLI. The video artifact shows the output as an inspectable graph view with filters and clustered concepts, which is the important point: the memory surface should be navigable and auditable, not just a pile of notes. His guide added another 3,574 likes and 9,339 bookmarks. Source: X/@NickSpisak_, 2026-04-06

The 2026 "Wiki Layer" re-explainer is a productized demo of the same architecture. The reviewed video shows files moving from a repo/tree view into an indexed graph, then into a generated wiki/dashboard with usage maps, heatmaps, token/cost counters, and a headline claim of roughly 90% token reduction. Preserve it as evidence that the flat-markdown/wiki layer is becoming a product surface, not as a source of truth for Kevin's implementation or as a versioned dependency. Source: X/@Asteri_eth, 2026-05-30; Source: local video review, 2026-07-03

Kevin's Implementation

Kevin's wiki (kevin-wiki) is a direct implementation of this pattern with significant extensions:

• 487+ pages across 18 categories, compared to Karpathy's minimal gist
• qmd for hybrid search (BM25 + vector + LLM re-ranking) - the search layer Karpathy left unspecified
• AGENTS.md schema defining page format, conventions, ingest pipeline, and maintenance scripts
• Brain-agent loop (The Brain-Agent Loop) - the read-write cycle that makes the wiki compound
• Skills taxonomy - executable SKILL.md files that the wiki references and agents consume
• Automation layer - recurring tasks that keep the wiki current without manual intervention
• Integration recipes - Calendar, Gmail, Circleback meetings, voice calls flowing into the wiki
• Bookmark capture - Field Theory CLI (Fieldtheory) (ft sync) pulls X bookmarks into raw/x-bookmarks/, then the Agent Operations Skills skill compiles them into the graph (the upstream collector Karpathy's gist left unspecified)

The comparison with other approaches is documented in Knowledge System Architecture Comparison; Nia Vault is a commercial productization of the same flat-markdown pattern, and Obsidian Second Brain is an off-the-shelf cross-CLI skill (Claude Code / Codex / Gemini / OpenCode) that applies the same agent-writes / human-reads contract directly to an Obsidian vault.

Why This Page Is Tier 1

This is the origin story reference for the entire kevin-wiki project. Every architectural decision in AGENTS.md traces back to Karpathy's core insight: flat markdown files, LLM as programmer, schema as contract. The engagement numbers (153K combined) confirm this resonated far beyond Kevin - the pattern is becoming infrastructure.

---

Timeline

• 2026-08-11 | Corrected the earlier shorthand: OKF is now v0.2, and Kevin Wiki is not itself a conformant canonical bundle. The richer evidence/object/review system remains authoritative and generates a privacy-filtered OKF projection for portable agent access. Source: Open Knowledge Format (OKF); GoogleCloudPlatform/knowledge-catalog@374e0bc4; local conformance replay
• 2026-07-04 | Source-reviewed the root @karpathy "LLM Knowledge Bases" bookmark and embedded the original tweet artifact so this page no longer depends on a reconstructed status ID. Source: X/@karpathy, 2026-04-02
• 2026-07-03 | Deep-reviewed the @Asteri_eth "Wiki Layer" video contact sheet: repo/file tree, indexing graph, generated wiki/dashboard, usage heatmap, token/cost counters, and 90% token-reduction framing. Kept it as architecture/productization evidence, not a new local dependency. Source: X/@Asteri_eth, 2026-05-30; Source: local video review, 2026-07-03
• 2026-07-01 | Deep artifact review added the Nick Spisak video frame: a local brain/Obsidian graph view with filters and concept clusters. The row stays an implementation pattern, not a versioned tool dependency; qmd and the wiki remain the local source of truth. Source: X/@NickSpisak_, 2026-04-06
• 2026-04-02 | Karpathy published "LLM Knowledge Bases" gist - 54,109 likes, 98,708 bookmarks. Source: X/@karpathy, 2026-04-02
• 2026-04-06 | Nick Spisak published setup guide - 3,574 likes, 9,339 bookmarks. Source: X/@NickSpisak_, 2026-04-06
• 2026-05-30 | "Wiki Layer" re-explainer (@Asteri_eth, 4,274 likes, 8,726 bookmarks) reframes Karpathy's LLM-wiki pattern as a ~90% token-reduction technique (the LLM cleans and structures the knowledge base so files are not re-read). Same architecture kevin-wiki runs on. Source: X/@Asteri_eth, 2026-05-30
• 2026-06-12 | Re-confirmed as #1 bookmark in corpus by bookmarkCount (98,708) in top-25 absorption pass; tweet text truncated in raw export but engagement confirms ongoing validation of LLM-wiki architecture. Source: X/@karpathy, 2026-04-02
• 2026-06-21 | The pattern got an official standard: Google Cloud's Open Knowledge Format (OKF) (OKF v0.1, 2026-06-12) formalizes Karpathy's LLM-wiki model into a portable, vendor-neutral spec (markdown + frontmatter, one concept/file, cross-links as graph, reserved index.md/log.md). OKF standardizes the "wiki body" layer; AGENTS.md/CLAUDE.md remain the "schema" layer. kevin-wiki is effectively an OKF bundle already. Source: Google Cloud Blog, 2026-06-16; X/@GoogleCloudTech, 2026-06-16