@homsiT@karpathy Right exactly. What feels compelling to me isn’t graph view by itself, but graph view behind the output: you work with the LLM on something in canvas, and as it evolves, you can see the knowledge map it’s pulling from and connecting through.
I wrote more about this below, but I still don’t see graph view itself as the real day-to-day value. Visualizing connections is cool to play with. The real unlock is when an LLM can use the underlying KG, traverse those relationships, and turn them into useful output🤔
https://t.co/E3OKXELrLm
@FarzaTV arrived at the exact same conclusion. The wiki is not for me, it's for the agent.
The piece that made it really compound for me was forcing every cross-link to carry a rationale, not just "these are related" but how and why. Turns the wiki into a lightweight KG the agent traverses without inferring connections from scratch.
Full set of design decisions: https://t.co/E3OKXELrLm
@aakashgupta this is exactly what happened to me. Took the idea file, had my agent build it in an afternoon, and then the real work started. The design decisions Karpathy intentionally left vague.
I've found they matter a lot more than the folder structure. Wrote them up here: https://t.co/E3OKXEKTVO
@ansubkhan Sick UI! That knowledge graph view and the neighborhood mini-graphs is very cool.
I went the opposite direction, optimized the wiki entirely for LLM traversal, so my main UX is a chat that queries the wiki to produce outputs, not browsing the wiki itself.
Would be sick to have something like your KG visual on the side as the chat progresses though. Best of both worlds.
The design decisions behind my approach: https://t.co/E3OKXEKTVO
@MushtaqBilalPhD@MushtaqBilalPhD I built it this way first, and realized I was never going to read the actual wiki myself. So I optimized it for the LLM like a lightweight KG, so I can use it with the LLM to generate outputs.
The full set of design decisions: https://t.co/EuSaCEblkA
@heynavtoor the "compiles once, compounds forever" framing is def the way to think about it.
What I found is that the compounding only kicks in when cross-links carry explicit rationale, not just "these are related" but why and how. That's what turns it from a linked wiki into a knowledge graph the LLM can actually traverse.
Wrote up the full set of design decisions here: https://t.co/E3OKXEKTVO
@milesdeutscher@milesdeutscher Loved this. Claude Code + Obsidian is exactly how I built mine too.
The missing piece for me was turning plain wikilinks into a knowledge graph with rationale on every cross-connection. More on that here: https://t.co/EuSaCEblkA
@NickSpisak_ your article was the push I needed to actually build the thing instead, thanks!
Once I had it running though, a whole new set of questions came up and I ended up taking a different direction. The wiki became a lightweight knowledge graph optimized for the LLM to traverse, not for me to read. The design decisions that got me there: https://t.co/E3OKXEKTVO
@karpathy your gist started something big, to say the least.
I took it in a slightly different direction: the wiki is optimized for the LLM to traverse (dense, structured, edge-labeled cross-links), while the outputs are what's optimized for me. The wiki is scaffolding. A lightweight KG.
Here are the design decisions that made it compound: https://t.co/E3OKXEKTVO
The Wiki Nobody Reads
I built @karpathy 's LLM knowledge base... then realized I was never going to read it.
So I turned it into a real lightweight knowledge graph (with meaningful edge labels on every cross-link).
The design decisions that actually make it compound:
https://t.co/E3OKXEKTVO