Schema

This wiki uses the Karpathy LLM Wiki pattern: three layers that separate raw evidence, compiled knowledge, and structural metadata.

Three Layers

daily/ — Raw Dated Entries

Immutable source documents, organized chronologically. Each file is named YYYY-MM-DD-topic-slug.md. These are Jeremy's original thinking entries — idea lists, reading notes, experiments, reflections.

269 entries spanning 2014–2020.

Browse chronologically via log.

wiki/ — LLM-Maintained Synthesis Pages

Compiled topic pages that synthesize across many daily entries. The LLM owns this layer — creates pages, updates them when new sources arrive, maintains cross-references.

Current synthesis pages:

• Machine Intelligence
• Thinking Frameworks
• Learning and Education
• Creativity and Ideas
• Focus and Productivity
• Career and Ambition
• Human Nature
• AI Safety and Alignment
• Research and Science
• Decision Making
• Books and Reading
• Social and Community
• Startup and Business
• Writing and Communication
• Memetics and Ideas

Root — Topic Pages

~190 standalone topic pages live at the root level. These are deep dives into specific subjects (e.g., abstraction, decision-making, meta-modeling). They sit between daily entries and wiki synthesis pages in depth.

schema.md (this file) — Structural Metadata

Describes how the wiki is organized, what conventions to follow, and how the layers interact.

Page Format

Every wiki page has two layers:

• Compiled truth (above the --- divider): current best understanding, rewritten freely
• Timeline (below the --- divider): append-only evidence trail with dates

Frontmatter

---
title: Page Title
type: topic | person | project | idea | source-summary
visibility: public
tags: [area1, area2]
sources: [daily/2017-08-07-deep-problems.md]
---

Wikilinks

Inline wikilinks are what make this a knowledge graph:

• Weave links into prose: [[meta-modeling|Meta-modeling]] is central to [[thinking-frameworks|thinking frameworks]]
• Link on first mention only per section
• Cross-category links are the most valuable

Operations

Ingest

When processing a new source: read it, identify entities, update or create wiki pages, add wikilinks, update index and log.

Query

Read index to find relevant pages, synthesize answer, file valuable answers back as wiki pages.

Lint

Periodically check for contradictions, stale claims, orphan pages, missing cross-references.