AI / Agent Reference · Reference
Karpathy's LLM Knowledge Bases
Karpathy described using Large Language Models to build personal knowledge bases for research topics. The core idea: use Large Language Models to distill and organize information into personal "wikis" for domains you're researching. Feed an Large Language Model a set of documents/sources on a topic Have the Large Language Model distill, organize, and structure the information Create a personal "wiki" or knowledge bas…
wiki/wiki/ai-ml/karpathy-llm-knowledge-bases.mdAnswer
Karpathy described using Large Language Models to build personal knowledge bases for research topics. The core idea: use Large Language Models to distill and organize information into personal "wikis" for domains you're researching. Feed an Large Language Model a set of documents/sources on a topic Have the Large Language Model distill, organize, and structure the information Create a personal "wiki" or knowledge bas…
Auto-generated neutral summary from the source page — needs human review before trusted use.
Evidence & Source Cards
https://x.com/karpathy/status/2040396678176362540external/unverifiedSource Excerpt
Overview
Karpathy described using Large Language Models to build personal knowledge bases for research topics. The core idea: use Large Language Models to distill and organize information into personal "wikis" for domains you're researching.
The Pattern
- Feed an Large Language Model a set of documents/sources on a topic
- Have the Large Language Model distill, organize, and structure the information
- Create a personal "wiki" or knowledge base for that domain
- Query the knowledge base for targeted information
Why It Matters
This spawned an entire movement. The 58,919 likes (highest engagement in our 260-bookmark review) confirm massive resonance. Multiple implementations followed:
- Farzapedia (@FarzaTV, highlighted by Karpathy): Personal Wikipedia of Farza
- Claude + Obsidian integration: Using Claude to populate, query, and organize Obsidian vaults
- Hot cache pattern: New sessions start with a summary of previous sessions
- No-code implementations: Explainers for non-technical users
Actionability
HIGH — Conceptually important and directly applicable. Hermes Agent could implement this pattern directly, using its memory system to build and maintain domain-specific knowledge bases.
Status
Conceptual pattern, multiple community implementations exist.
Sources
- https://x.com/karpathy/status/2040396678176362540
- @karpathy, April 2, 2026
Relationships
Outbound links
- Agent Trace Distillationcorpus
Referenced by
- No corpus pages link here yet.