Nir Diamant on linkedin

Andrej Karpathy just shared a better way to use LLMs (and it got super viral). let me summarize that for you, with a simplified algorithm of what to do:

He suggests to use llms as a knowledge system.

The workflow is simple:

Collect raw sources
Put articles, papers, repos, datasets, and images into a raw/ folder.
Ask the LLM to read each source
Extract the key facts, entities, concepts, and contradictions in short notes.
Write or update wiki pages
Turn those notes into markdown pages for the source, related entities, and related concepts.
Update the index
Keep index.md short and searchable so the LLM can quickly find the right pages.
Append the log
Add a brief entry to log.md with what changed and why.
Run lint passes
Ask the LLM to find stale claims, missing links, contradictions, and gaps.
Query the wiki
Let the LLM search, read, and synthesize answers from the relevant pages.
Keep it in git
Version everything so the knowledge base can evolve safely over time.

repost to let others know!

1 / 2

Eviatar Levy Treating an LLM as a “knowledge system” really means git is the system and the model is a lossy parser. Without traceable links from each claim back to raw sources, the wiki quietly normalizes contradictions and drifts. Apr 5 1 like

Satish Venkatakrishnan Good breakdown. The index.md step is the one that does the heavy lifting — at ~200 tokens for 50 sources, the LLM reads the whole index and selects files by judgment, not similarity search. We've tested this up to 10,000 documents in production without needing embeddings.

Built the working version: github.com/satish860/llm-kb Apr 5 5 likes