gist.github.com
llm-wikillm-wiki. GitHub Gist: instantly share code, notes, and snippets.
April 04, 2026 ·
Last month I wondered in one of these posts about the difference between information and knowledge. I heard David Deutsch say on a podcast the other day that knowledge is information with causal power. I really liked that. He said it was a working, inevitably incomplete and informal definition, and that all definitions ultimately contain things that cannot be fully pinned down or transferred in their entirety. I also liked that. Sounds similar to Karl Popper saying it's impossible to speak in such a way that you cannot be misunderstood. That for some reason is comforting for me to hear.
Then the other day I noticed Andrej Karpathy made this mechanism to have an LLM start create something akin to a personal Wikipedia via tools like Obsidian. This knowledge graph business has been ongoing for a while. I was an early user of Roam Research, the flashpoint out of which Obsidian was born. I've been manually building knowledge graphs for years at this point, so I know a thing or two about how they work. There was a lot of hype when they first came by, popularized as second brains and such, and there was the just-as-expected mockery of the whole thing. I have started noticing more and more in the last few years that technological progress is simultaneously overhyped by certain crowds and as a reaction downplayed and ridiculed by some. The truth tends to be somewhere in the middle.
I will link Karpathy's idea-gist below, which I thought was a neat trick here; that what's being shared is not code, but purposefully open-ended instructions for an LLM to operationalize in its local environment. That, is cool. However, I do find myself wondering what changes if instead of me manually dropping notes, links, pictures and media, etc. in my Roam Research or Obsidian notebook...it's an LLM doing it?
I can imagine the knowledge graph's volume being a few magnitudes higher than mine as an LLM can read and write at an astonishing pace. But I'll ask the classic question I've been asking for similar new tech updates recently.
Okay...so what?
I'll have to actually do this to truly find out, to be fair. But I think the overall reality of the situation remains somewhat the same is my guess. As in, I find having knowledge graphs extremely useful and I use mine multiple times a day, every day. This little blogging experiment I started is quite simply my knowledge graph externally exposed to a website. I write in my knowledge graphs exactly how I write here. And I sometimes look up things I had put in their previously. When I first got into it, I think I overdid it and put too many things in there that made it really messy really quickly. Tweets, links, pictures, PDFs and what not that I never referenced again, and things seemed like one of those hoarding reality TV shows in a few months. Why would an LLM doing the same thing not yield the same result?
The best analogy I've heard on knowledge graphs over time is that of gardening. Because if you don't tend to it and maintain it, it's not a place you want to spend much time in after too long. LLMs are like tractors or green revolution type technology upgrades in that sense. I won't be surprised if they enable huge knowledge graphs, and maybe some people need that. But the upgrade is in how automated one is able to do these tasks, not in the ultimate responsibility one has in owning and maintaining the garden or graph.
I should give it a shot. We'll see how it works out.
Title inspired by etymologies of knowledge and base.
comments section