When a Karpathy Tweet Becomes a Tool
In early April, Andrej Karpathy posted a simple idea that hit the internet with unusual speed. Feed source documents into an LLM, have it turn the material into a personal Markdown wiki, then keep adding new documents so the wiki grows with the work. That was the gist. No glossy product launch, no keynote choreography, just a description of a workflow that sounded oddly practical for a piece of AI tech news.
Within hours, the idea had escaped the post and started mutating into tutorials, GitHub repos and how-to videos. That kind of spread usually belongs to novelty gadgets and weekend distractions. This one felt different. It landed in the same place as a messy desktop folder full of PDFs, half-read tabs, and notes you swear you’ll file later. If you work with research, reporting, policy docs, legal briefs, market notes, or any other pile of text that refuses to stay small, the pitch is immediately legible: stop relying on bookmarks, browser history, search engine luck and whatever context happens to survive in your head after a long afternoon.
The appeal isn’t that the wiki knows everything. It’s that it stops your own notes from forgetting things at the worst possible moment.
That, more than anything, explains why the idea traveled so quickly. It didn’t arrive as another abstract promise about artificial intelligence changing work. It showed up as a workaround for a very ordinary failure mode. People lose track of what they’ve read. They forget where a fact came from. They know they saw a useful line somewhere, then spend ten minutes trying to reconstruct the trail. A normal search box can sometimes help, but it also depends on the right keywords, the right phrasing, and a little luck. This wiki idea skips that scavenger hunt and tries to make the documents themselves do the sorting.
For desk workers, that’s the whole appeal. The promise isn’t magical. It’s almost rude in its simplicity. Put the material in one place, let the model extract people, companies, concepts and links between them, and keep expanding the notes as new text arrives. Instead of treating every document as a dead end, the system treats each one as a node that can connect to older material. A note about one company can point to a regulator, a product, a policy fight, a rival, or a story from months ago that suddenly matters again. That’s useful in tech policy, digital culture and power and politics work, where the same names keep reappearing under new labels.
There’s also something mildly dangerous about it, which is where the infohazard comparison comes in. It becomes very hard to ignore, once the idea crosses your feed. You start eyeing your own storage habits with suspicion. Why are there twelve places where this fact could live? Why is the good context trapped in a browser tab I opened on a Tuesday? Of all things, to remember a moving target?, why am I trusting bookmarks. The thought process isn’t graceful. It’s closer to an itch than a plan. But once the itch starts, building the thing feels less like a hobby project and more like self-defense.
That’s part of why the concept caught on beyond the usual AI crowd. It didn’t ask people to believe in a future where machines do everything. And it asked them to admit their current setup for remembering work’s clumsy. That’s a lower bar, and an easier sell. A lot of AI tools promise speed. This one promises recall. Different beast entirely.
The first wave of reaction made the idea feel half meme, half utility. It second wave made it feel like a workflow people might actually keep. What happens when someone stops talking about the concept and starts wiring it into their own archive’s where the story gets interesting, because the real test isn’t whether the idea sounds neat. It’s whether it can survive contact with the pile of material you already have.
Inside the Self-Updating Wiki
That viral idea turns out to be less mystical once you see the plumbing. The setup began with a Claude model writing the prompt for a Karpathy-style wiki, then the writer ran it from the terminal and watched the output land as a plain folder of Markdown files inside Obsidian. No glossy dashboard. No mystery box. Just files, folders, and a lot of text that could be opened, searched, and edited like any other newsroom note.
The first pass wasn’t built from scratch in some sterile vacuum, either. It was seeded with the writer’s own archive, which gave the model years of material to chew on. From that pile, it pulled out names, companies, products, concepts and recurring themes, then started threading them together. If a reporter had filed on Meta, OpenAI, TikTok, or a dozen smaller AI startups over time, the wiki could turn those mentions into linked pages instead of leaving them buried in old drafts and half-remembered tabs. That’s the whole trick, really. A personal knowledge base stops being a static dump and starts acting like a living index.
The system only works because it never waits for you to remember to remember.
Each morning, after journaling, the writer clips selected stories into Obsidian, converts them to Markdown and lets a local script decide where they belong. A story about a new policy fight might get folded into a company page, a concept page, and a topic cluster all at once. And a product announcement can update the page for the company behind it, the person who said it and the broader topic it sits inside. The script does the boring sorting, which is the kind of boring work that usually eats the day if nobody automates it.
That daily routine matters because it keeps the wiki current without asking for a heroic maintenance session. It’s not a one-time import that slowly decays into digital furniture. New material gets added, then the system moves it around. Pages grow. Links multiply. The home page refreshes every morning, so the front of the wiki reflects what has been added most recently rather than whatever happened to be on top last month. It feels closer to a newsroom bulletin board than to a polished knowledge product, and that seems to be the point.
The scale’s gotten a little absurd, in the best possible way. Which is enough to make a normal bookmarking system look underfed, given the wiki now holds well over fourteen hundred pages. Some entries are tiny, and others have swollen into small dossiers. The Meta page, for example, runs to roughly twelve thousand words and a four-figure web of links, which tells you something about how often that company has shown up in the writer’s beat and how much connective tissue the model can find once you let it loose on years of coverage. Nobody is pretending that’s tidy. It’s useful because it’s not tidy.
A plugin called Claudian sits on top of the whole thing and lets the writer ask the wiki direct questions. Instead of rummaging through old notes and hoping the right sentence turns up, the writer can query the archive itself. Which names keep surfacing in coverage of a certain policy fight? What companies were linked to a topic six months ago? Which stories mentioned a concept before it became part of the current conversation? That’s a different experience from searching a pile of documents by hand, and it feels especially natural for AI productivity work where context fades fast.
The broad idea isn’t far off from enterprise assistants like Microsoft 365 Copilot, Google’s work on agentic RAG systems, or the retrieval-heavy approach laid out in this recent paper on agentic retrieval. The difference here is scale and temperament. This version lives in Obsidian, runs from a terminal, and is shaped around one reporter’s archive rather than a company’s office suite.
The result is a personal knowledge base that acts less like storage and more like an argument with the past, asking, every morning, what already happened that might matter again today.
Why It Outran Search, Bookmarks, and Blips
Once the Karpathy-style wiki was running in Obsidian, the comparison stopped being theoretical. It wasn’t about whether knowledge management could be made elegant on a whiteboard. What stands out: it was about whether a tool could save time when a story, a source trail, or a half-remembered thread was already slipping away.
A Notion search agent had helped a bit. It could find things, which is more than you can say for a lot of digital furniture. But it lived on a different surface, and that mattered. Habit forms around proximity. If the thing you need sits in another app, behind another shortcut, with another way of asking for answers, you’ll use it less often than you think you will. It also tended to need extra nudging when citations mattered. That sounds minor until you’re trying to remember where a claim came from and the assistant hands you a confident answer with no receipts.
A search box is only useful if you already know how to ask the right question.
That’s the mismatch the wiki solves. It doesn’t wait for you to invent the correct prompt. And it keeps adding context in the same place where the context already lives. The system grew out of an older habit: making “blips” pages in Capacities for emerging themes. One of those pages tracked AI-driven job loss. The idea was simple enough. When a story started showing up in multiple places, give it a page, jot down a few notes and let future-you find it faster.
That worked for a while. Then the number of concept pages grew, and the whole thing got brittle. A random reminder system helped to prompt updates, but reminders have a way of becoming background noise. Maintenance turned into a small tax. New pages stopped feeling easy, then started feeling like chores. At that point, a system that depended on memory plus discipline was doing what these systems always do: quietly asking for more effort than advertised.
The wiki took that old blips setup and made it automatic. Instead of waiting for a person to notice a theme and create a page by hand, it generates concept pages for current topics as new material arrives. AI copyright gets a page. Youth social media bans get a page. Tokenmaxxing gets a page. AI and Congress gets a page. The point isn’t that each topic is grand or permanent. Most aren’t. The point’s that they recur, mutate, and drag other stories along with them.
That matters in daily reporting because the useful context is rarely the obvious context. A new policy announcement makes more sense if you can immediately see the earlier arguments around copyright. A podcast idea gets sharper when you can see which recent links keep landing under the same topic instead of treating each one as an isolated blip. The wiki clusters those links under living topics, so story leads and audio angles don’t have to be rediscovered from scratch every morning.
The structure also helps when a subject suddenly starts breathing on its own. A topic page can collect fresh links, notes and related pages without requiring a new manual category every time the news shifts a little. That’s a small thing until you’ve spent years watching the same issue mutate across platforms, hearings, product launches and the occasional public panic. Then it feels less like a clever feature and more like a decent filing system that finally learned to pay attention.
Then again, the rest of the stack still matters, of course. Raycast does the fast utility work: launching apps, quick AI searches, math, moving windows around, and pulling clipboard history when a stray sentence vanishes into the ether. Capacities still holds the daily journal, which is a different job from the wiki’s broader memory. Recall remains handy for near-instant YouTube summaries, which is a fine example of a tool doing one annoying thing fast enough that you stop resenting it.
That split is part of the appeal. The wiki isn’t trying to replace every piece of software on the desktop. It’s taking over the part where older systems were weakest: surfacing the right context at the right moment, without making you go spelunking for it. In the background, a few familiar tools keep doing the smaller jobs. The wiki keeps turning recent material into something you can actually use before the next deadline shows up, in the foreground.
The Maintenance Tax—and Why It Still Wins
That usefulness comes with a tax. The wiki doesn’t sit there politely waiting for the next query. Pages swell, some turn into small novels and the whole thing has to be compacted now and then so the useful material doesn’t get buried under its own enthusiasm. Scripts break too. Not every morning, maybe, but often enough to remind you that this is a hand-built system, not a polished app with a glossy settings page and a support team waiting in the wings.
The first draft of the system had another problem: the writing was almost too compressed to read comfortably. LLMs can do that. Give them a pile of context and they’ll cheerfully produce prose that feels like it was assembled by someone trying to fit a filing cabinet into a carry-on bag. So the whole setup had to be rewritten with another model. This time to sound closer to AP style, with cleaner sentences and less of the dense shorthand that makes machine output look clever for about five seconds and then annoying for the next five minutes.
The best automation quietly removes busywork without pretending the mess has vanished.
That trade-off shows up most clearly when the news is moving fast and the details matter. During the OpenAI and Hugging Face agent breach story, the wiki gave the writer a fast way to refresh context, check source material and sort out what was confirmed before sitting down to write or record a podcast. That’s a pretty narrow use case, sure, but it’s a real one. For tech news work, where a story can change shape while you’re still opening tabs, the difference between vague memory and a current, searchable archive isn’t abstract. It’s the difference between sounding prepared and sounding like you just skimmed the group chat.
The catch’s that this setup is deeply personal. It depends on one person’s archive, one person’s tagging habits, one person’s patience for fixing scripts before coffee. A newsroom, a sales team, or a random startup would probably find it clumsy in short order. Pages would grow in odd directions. The maintenance would annoy everybody. Someone would ask why the thing keeps making pages about subjects nobody asked for. Fair question. It’s too idiosyncratic, too dependent on constant tending, and too attached to its creator’s workflow to pass as a universal product.
Still, the shape of it’s hard to shrug off. When workplace software starts organizing itself around what people actually read, save and ask about, a lot of old friction may start to look unnecessary. Not glamorous, just useful. A tool that knows where the context lives, keeps it current and gets out of the way when it doesn’t have anything to add. That sounds modest until you remember how much of office life is spent hunting for the right note, the right page, or the one sentence you swear you saved somewhere. If the future of knowledge work arrives in a folder full of Markdown files and a few cranky scripts, that would be very on brand.
