Karpathy’s “compiler” trick turns an Obsidian vault into an LLM‑maintained wiki

Lead: a small change, big payoff

It has been reported that Andrej Karpathy recently described a workflow where he treats a personal knowledge vault like source code: raw materials go into a raw/ folder and an LLM “compiles” them into a wiki/ of interlinked markdown pages. Huxiu (虎嗅) picked up a WeChat post by 曼谈AI that tested the idea on an Obsidian vault and found the results striking — dormant PDFs and scattered notes became an active, queryable knowledge graph in minutes. Obsidian is a local, markdown‑first personal knowledge manager; Karpathy’s pitch is simple: separate raw inputs from compiled knowledge and let the model do the bookkeeping.

What the rebuild looked like

Using Anthropic’s Claude Code to act as the compiler, the author reportedly fed six industry PDFs (Gartner, SemiAnalysis, Omdia, etc.) into raw/, asked the model to extract 3–5 core judgments per report with evidence and commentary, and wrote those outputs into wiki/ with frontmatter and Obsidian backlinks. The LLM then auto‑generated a concept article — for example, “inference cost” — that stitched cross‑report data into a unified narrative. The vault went from a dozen loose files to about 35 interlinked articles; every time the author queried the system the LLM read the INDEX.md, validated facts by searching the web, and wrote new synthesis back into the wiki. Reportedly, this workflow allowed a cross‑report question (about GPU rental prices and animation production cost) to produce a nuanced, evidence‑based answer that was then saved as a new knowledge node.

Why it matters — and the limits

This is more than automation of note‑taking. The key distinction is “compile” versus “summarize”: summaries shorten a single source; compilation reorganizes many sources into a conceptual network that an LLM can reason over. For most personal vaults — dozens of documents and a few tens of thousands of words — a simple INDEX.md (one‑line summaries) plus an MOC (map of content) can replace early, complex RAG/embedding setups. But there are caveats: the reported claim that H100 rental prices rose ~40% and that some production costs fell dramatically comes from the user’s compiled analysis and should be treated as reported, not independently verified. There’s also the eventual scale problem — at hundreds of articles a vector database may become necessary — and the usual LLM risks of stale data or hallucination.

Geopolitical shadow and practical takeaway

Broader forces matter too. Analysts say regional price divergences in GPU access are partly driven by export controls and trade policy that have constrained high‑end chip supply to some markets — a geopolitical factor that colors any inference‑cost analysis. The practical takeaway is straightforward: if you hoard PDFs and notes, the knowledge lies dormant. Feed them into a raw/ folder, let an LLM compile, then ask good questions. The human job shifts from constant tidying to curation and interrogation: collect inputs and ask the right questions — the model does the linking and scaling.