LLM Wiki — Synaptic Lattice

Andrej Karpathy

Properties1
tagsentity, llm-wiki, karpathy

Leitura de 5 min

Overview

Andrej Karpathy is an AI researcher known for founding the LLM Wiki pattern in April 2026. His original GitHub gist describing the approach went viral, inspiring a wave of implementations and community tools. Previously, he was a founding member of OpenAI’s technical team, Director of AI at Tesla (leading Autopilot vision), and later returned to OpenAI before leaving to found Eureka Labs.

Key Facts

  • Published the original LLM Wiki gist in April 2026 that defined the three core operations: ingest, query, lint
  • His approach shifted the paradigm from RAG (retrieve at query time) to compiled knowledge (synthesize at ingest time)
  • Known for educational contributions: cs231n (CNNs for Visual Recognition), nanoGPT, and the zero-to-hero series
  • The gist went viral within the AI community, described as a “living AI knowledge base that actually heals itself”
  • Advocated for Obsidian as the front-end IDE for LLM Wikis because of its markdown-native design
  • Endorsed the concept of “context engineering” (coined by Dex at HumanLayer) alongside Shopify CEO Tobi Lütke, elevating it to an industry-standard term
  • O gist preservado neste vault descreve o LLM Wiki como uma base de conhecimento pessoal construída por LLMs, com Obsidian como interface e arquivos Markdown como banco de dados.
  • Authored a 2025 essay series on the nature of LLM intelligence and its societal diffusion: “Power to the people” (LLMs reverse the usual top-down technology-diffusion pattern), “Verifiability” (Software 2.0 / the verifiability-as-automatability thesis), “The space of minds” and “Animals vs Ghosts” (LLMs as “ghosts,” a fundamentally non-animal kind of intelligence), synthesized in his own “2025 LLM Year in Review.”
  • Coined the term “vibe coding” in an offhand 2025 tweet, later analyzed at length in his own retrospective as a concrete instance of the technology-diffusion reversal he described in “Power to the people.”

Contested: What the Karpathy Layout Actually Specifies

A widely-viewed 2026 video by chase-ai presents “the Karpathy Obsidian RAG” as a three-subfolder vault — raw/ for unstructured input, wiki/ for structured synthesis, and a third folder for outputs (deliverables such as slide decks) — attributing the structure to a Karpathy tweet with over 20 million views.

This vault’s primary evidence does not support the third folder. The gist preserved here (April 2026) defines raw/ and wiki/ as the two layers, alongside three operations — ingest, query, lint. The number three appears in both accounts, but it counts different things: operations in the gist, folders in the video.

Both positions are preserved as dated claims rather than resolved. The video may be describing a later or different Karpathy post that this vault has not captured, or it may be conflating the operations triad with a folder triad. This vault itself implements raw/ + wiki/ with no outputs/ layer, which reflects the gist rather than the video.

A second, softer conflict sits in the same source: the video labels the pattern “the Karpathy Obsidian RAG,” calling it “a new RAG system.” llm-wiki and rag record the opposite framing — the pattern’s defining move is compile-time synthesis instead of query-time fragment retrieval. The video’s own air quotes suggest loose usage rather than a substantive claim, but the label is recorded here because it travels widely.

Implications: Both conflicts run in the same direction — a popular secondary source restating the pattern more loosely than the primary one. That makes the video useful as evidence of how the pattern is being received, and unreliable as evidence of what Karpathy specified. Until a captured Karpathy source shows an outputs/ folder, the two-folder reading stands as this vault’s position, and the video’s version is recorded beside it rather than merged into it.

Relationships

  • “created” llm-wiki — Originator of the LLM Wiki pattern
  • “created” autoresearch — Originou o padrão “autoresearch” (março 2026): loop autônomo de experimentação guiado por goal+metric+loop
  • “used” claude-code — Primary tool used to implement his pattern
  • “inspired” open-knowledge-format — His pattern inspired OKF-compliant implementations
  • “authored” llm-as-os — His 2023 “LLM as kernel of a new OS” framing, later extended by the AI-OS report
  • “influenced” cognee — production memory engine that scales the pattern to millions of docs
  • “influenced” yopedia — fully autonomous, agent-evolved wiki implementation
  • “inspired” nateherkai-ais-os — AI Operating System starter kit building on the pattern
  • “endorsed” context-engineering — Co-signer of the context engineering concept
  • “authored” technology-diffusion-llms — “Power to the people” essay on LLMs reversing top-down technology diffusion
  • “authored” verifiability — “Verifiability” essay defining the Software 2.0 / RLVR automatability thesis
  • “authored” ghosts-vs-animals — “The space of minds” and “Animals vs Ghosts” essays on LLM intelligence as non-animal “ghosts”
  • “authored” 2025-llm-year-in-review — his own retrospective synthesizing the year’s paradigm shifts
  • “coined” technology-diffusion-llms — coined “vibe coding,” discussed within this concept’s diffusion framing

🔗 Conexão inferida — sugestão do segundo cérebro

Liga para: llm-wiki, claude-code Evidência observável: Karpathy founded the LLM Wiki pattern; his original gist defines the three core operations (ingest/query/lint); Claude Code is the tool that implements those three operations.

Sources

^[raw/articles/karpathy-llm-wiki-gist.md] ^[raw/articles/venturebeat-llm-wiki.md] ^[raw/articles/genai-unplugged-llm-wiki.md] ^[raw/articles/levelup-llm-wiki-deep-dive.md] ^[raw/transcripts/sozai-fable-llm-wiki-video.md] ^[raw/articles/adgapar-loop-conversation-ai-agents.md] ^[raw/feeds/https-karpathy-bearblog-dev-power-to-the-people-28a4b3a1adbe15c8.md] ^[raw/feeds/https-karpathy-bearblog-dev-verifiability-576bce59c7e3a574.md] ^[raw/feeds/https-karpathy-bearblog-dev-the-space-of-minds-0ee8301f13fcc405.md] ^[raw/feeds/https-karpathy-bearblog-dev-animals-vs-ghosts-80f0901d56fc251c.md] ^[raw/feeds/https-karpathy-bearblog-dev-year-in-review-2025-604f493848c47ec0.md]

Visão de gráfico

Backlinks