After a month on Karpathy's LLM Wiki, the bottleneck isn't setup. It's maintenance

I think I was one of the first few people who immediately read that Andrej Karpathy tweet, and it just clicked. Dump your sources into a folder, let an AI read them all and build a wiki on top, then ask the wiki questions instead of digging through the original docs. Once you see it, you can't unsee it. I spent the last month actually building it. Here's what I learned, in the order I learned it. Week 1: Setting it up is the easy part A weekend was enough to get a basic version working With Claude and Obsidian combo. I fed it about 80 articles and PDFs, and by Sunday night I had a working wiki that summarized everything and linked related ideas together. It genuinely felt like magic. I told two friends Karpathy had cracked something fundamental. Week 2: The first cracks Getting clean text out of messy sources is a nightmare. Scanned PDFs come out as gibberish. Some websites won't load properly when a program tries to read them. Tables turn into garbage. Footnotes get jumbled into the main text. Every new type of source was a new evening of frustration. Week 3: The real problem shows up I added 50 new articles in one batch and realized the wiki had no idea they existed. To actually fold them in, the AI had to re-read and re-organize everything from scratch, which took 40 minutes and cost real money in API fees. Then I noticed three of my older summaries were quoting an article that had been updated weeks ago. The wiki was confidently telling me things from a version of the source that no longer existed. This is when it hit me. Karpathy's method assumes your sources sit still. Real research doesn't work that way. Articles get updated. Posts get deleted. You add new stuff in batches. A wiki built on a snapshot starts going stale the moment you finish building it. The maintenance problems I kept hitting: Stale summaries. A source gets updated and your summary is silently wrong. Nothing tells you. No way to know what changed. Even when I knew a source had been edited, I had no way to tell if the edit mattered enough to re-summarize. Adding new stuff means redoing everything. There's no clean way to just slot in new sources without rebuilding the whole wiki. Deleting is worse than updating. Remove a source and the wiki still references it like a ghost. The same website starts parsing differently after a redesign. You don't notice until a summary comes out broken. None of this is about prompts. None of it is about which AI model you use. It's all about keeping the underlying pile of sources fresh and clean, and that's the part nobody talks about. Week 4: Giving up and trying the no-code options This feels like defeat. I don't know if I'm the only one out there. Here are some low-code options I'm looking at. Maybe I just missed something, and I need to go back to the drawing board. If I did, please can you offer some guidance below? Trust me, I've watched almost all of the tutorials and gone through all the red threads on it, but maybe it's just me. I'm now shopping around for no-code solutions of Karpathy's LLM wiki. This is what I'm considering. Has anyone else tried these and have a successful flow? Claude with Notion: This isnt no code but it's just an alternative to Obsidian that I actually find is quite clever. I find the right MCP to be pretty smooth, and I quite like that I can create tasks and reminders versus only knowledge management. It's not exactly the same workflow, but it's a slightly tweaked version that I actually think is pretty cool. The downside is that Notion doesn't handle YouTube videos and PDFs as well. Mymind: I'm super excited about this one, but I'm not quite ready to do it. The website is beautiful, and I feel very peaceful in it, but I'm not too sure if this is a lifelong second brain or a peaceful Pinterest of knowledge. Has anybody used this? Please let me know. Recall: an AI knowledge base is the closest thing to what Karpathy is actually describing. It looks like you can add pretty much any online content: YouTube videos, podcasts, PDFs and it reads, summarizes, tags, and connects everything automatically. The catch is it's cloud-based. What I actually want to know Has anyone built their own version of this that doesn't go stale? I couldn't crack it and I'd love to be wrong. For people still running Karpathy's setup with a lot of sources, how are you dealing with summaries that go out of date when articles get edited? Is there a tool I missed that treats keeping sources fresh as the main job rather than an afterthought?