Karpathy LLM Wiki pattern integrated into Obsidian agenic workflow | Hacker News Reader

Karpathy LLM Wiki pattern integrated into Obsidian agenic workflow

github.com

28 pointspssah49 comments

I am apparently someone who did not quite catch up to the AI lingo, so parts of the explanation are confusing to me (what is a "vault", why does this thing do stuff in a "loop", how is this more "agentic" than a cronjob).

Is my understanding correct that a normal LLM is stateless in the sense that when you talk to it today about frying pans, it does not remember that you spoke about fried rice yesterday? Is this solution effectively adding Markdown files as part of the prompt? Essentially writing into a file "whenever I talk about scripting, I explicitly mean the zsh"?

Vault is an Obsidian thing. This is just an "agent" (read: LLM harness in a loop) with access to it, a system prompt and a few markdown files. Nothing meaningful has changed in the LLM primitives, nor will it :)

this is great for when you want to feel like you have a lot of data and structure but dont want to validate it.

imagine wanting your own well graphed dataset wiki thats 70% reliable. well, now anyone can do this, regardless of effort!

I feel like the others didn't understand the part where you weren't into AI lingo. You're correct, this is like having a memory for an LLM, because an LLM is stateless.

When you chat with an LLM, there's a concept called 'conext'. In essence, context is feeding all previous messages into the LLM together with your latest message. Because context is essentially a finite resource (it requires system memory and increases processing time) the bigger AI providers use tricks to compress context.

These providers usually also have 'memory', which in essence is just parts of previous chats that are entered into the current context based on their relevance. I don't know exactly how this works, but I'd imagine that it does some search for related chats and then adds summaries of those.

In essence, this tool allows you to do those things locally. This allows you much more control of what history the LLM gets and therefore the 'context' it works with. This is important, because context can get dirty. You can notice this if you're chatting with an LLM, it goes completely the wrong way, you try to get it on the right track again, but it just won't. That's because it just tries to predict the next word based on the full context and it might end up consistently predicting the wrong next word.

I would happily pay for anything like this for Logseq.

As someone who uses Notion on the free plan, does Obsidian work better than Notion with local AI tools?

Obsidian is just a client that uses markdown files stored locally, so yes, since the files are typically on the same machine as the agent harness

And are exposed as local files stored in documented, stable, and sane paths.

Their sync is good too, in my limited experience.

Sync is good, but they also have a great git plugin. One big advantage of local text based notes is that you can just source control them and set up quick shortcuts to commit, push and pull; as well as setting up automations to do that automatically when you open the app.

> And are exposed as local files stored in documented, stable, and sane paths.

This is highly dependent on the user (Mine is a god damn mess) :D.

Reply on news.ycombinator.com