Ages ago, Andrej Karpathy tweeted this and it set the internet on fire.
Upon closer inspection, that was three weeks ago.
In that lifetime, empires rose and fell, and word on the street is RAG told everyone to eat cake and then was summarily guillotined. It has been quite the revolution.
The tweet has, to date, garnered over twenty million views, nearly three thousand replies, and quite a few implementations utilizing the referenced LLM-Wiki. One of the most popular, retweeted by Karpathy himself, is Farzapedia — check it out; as someone trying to build and think in public, Farza’s work is inspiring.
All for a personal wikipedia.
Which begs the question: of all the amazing applications of AI, why is this one so popular? And why now?
I dug around for an answer. But for all the reading and reacting and digesting (I lost sleep… and I love my sleep), I felt like I didn’t get it. Frustrating. So I took a walk — to the fridge and back. On the way, a voice in my head perked up and said, “what if it’s not your understanding? What if you’re just not buying what’s being sold?”
And that set off this article.
Walk with me and I’ll explain… this time to the pantry. We’re going to need some snacks for this next bit.
Lately, I’ve focused on emotional resilience work. One of my practices is understanding that actions are precipitated from a mixture of both logic and emotion: not one or the other, as much as we may think otherwise. While that seems like an obvious statement, you’d be surprised how often it gets dropped. Sometimes at your own convenience — “I am a calm and level-headed person exercising good and rational judgment, and therefore I must purchase this new technology.” But also with our fellow humans — “Did that dude just cut me off in traffic? What an asshole! I bet his lifted truck is compensating for…”
Neither statement is a complete picture. They are both out of balance.
Taking the time to re-balance the logic/emotion scale in each situation leads to a more emotionally resilient and less reactive existence. Viktor Frankl put it better than I can: “Between stimulus and response there is a space. In that space is our freedom and our power to choose our response.”
So that situation with yourself becomes, “I think I’m being rational about this purchase, but maybe I’m trying to self-soothe from work stress. I’ll wait 24 hours to let my emotions reset.” And the situation with other humans becomes, “Maybe that driver really needs to poop. Been there, my guy… get in here and godspeed to your porcelain salvation.”
For me, this practice has carved out a calm harbor in a life that is full of storms.
I ran across Karpathy’s tweet right after some resilience work and figured, why not? I was not entirely prepared for the looking glass I stepped through. But first, a brief tangent through history. Or, as I’d like to phrase it: “We’re doing this shit again… but with tokens.”
Told you we needed snacks.
First, we dive back nearly 80 years to Vannevar Bush and his paper “As We May Think” (1945). Bush introduced the Memex: a mechanical desk that stores all of a person’s books, records, and communications with associative trails between items. The original vision. He imagined the maintenance problem but couldn’t solve it.
Then came Ted Nelson, hypertext (1960s), and the Xanadu project: every document linked to every other. The dream of a universal knowledge web.
It never shipped.
Next, Ward Cunningham and the wiki (1995). Collaborative, editable, linked pages. Wikipedia proved the model at scale — but for collective knowledge, not personal. The key: the maintenance problem was distributed among the populace, weighted to those who cared most about a given topic, leveraging the audience’s emotional and mental investment in the subject matter. Crowdsourced collective knowledge.
With the success of Wikipedia came renewed attempts at personal wikis. Zettelkasten had a revival in the 2010s. Niklas Luhmann’s slip-box system, rediscovered by the productivity internet. Roam Research, Obsidian, Logseq. The “tools for thought” movement. But it still faced the same problem as Bush: maintenance. The result: millions of abandoned vaults.
Each iteration promises a near-existential power — that if you could just get your thoughts into the right structure, you’d finally understand what you know.
And therein lies the nerve, running the limbic length of personal knowledge history, that made the tweet go viral: Karpathy’s explicit claim that “the wiki stays maintained because the cost of maintenance is near zero.” He solved the Memex’s maintenance problem. The LLM handles the grunt work.
To come full circle: the thing to “get” about Karpathy’s tweet isn’t the architecture, the restructuring, or any technical debate. It’s an implicit promise. With the three-layer architecture of the LLM-wiki, we can finally unlock AI as a true second brain. No work on our part. All you have to do is feed it information.
Alright, end of article. We’ve solved it.
Wait. Why do we want to document and organize everything we read, watch, or write?
What is the goal?
“Your scientists were so preoccupied with whether or not they could, they didn’t stop to think if they should.” Thank you for that, Dr. Malcolm.
At its surface the premise makes sense, and is exactly as Karpathy puts it: “stop deriving, start compiling.” With a second-brain system like LLM-wiki, you’re no longer piecing together chunks of data — mentally and digitally — you’re aggregating a snowball of knowledge that compounds the longer you use it. An implicit bargain: the longer you use this system, the smarter and more attuned to you an AI will become.
So back to the core question: why did Karpathy’s tweet set off such a storm? If actions are a mix of logic and emotion, then we’re only being shown the logic side of the equation. Productivity gains. Searchable knowledge bases. Delegation of mental bandwidth for higher-order thinking. Reversal of entropy that’s really just a hastening of the inevitable heat death of the universe.
But there’s also an emotional side, and I’d argue it’s the primary driver of this tweet’s velocity: we all just want to be seen. Seen by our loved ones, seen by our peers, seen by society. And with the human-esque presentation and appeal of AI, we want to be seen by our technology too.
To be seen, we must show. To show, we must see ourselves first.
And speaking from my own journey, that is awful, difficult work.
So let’s shortcut it with technology.
Here is where the appeal gets its teeth. We — humans ranging from deep-in-the-ocean technologists to treading-water tech users — are drowning in information overload. Ten-plus Chrome tabs open? Those are rookie numbers, gotta pump those up. Several hundred bookmarks and growing? All day, every day. Notes, videos, and shared content from multiple apps received and never consolidated or reviewed? Does today end in y? Yup.
Somewhere, we know that this information pocket litter contains bits and pieces of who we are. If you are the average of the five people you spend the most time with, then there has to be a digital analog: you are the average of the content you spend the most time around.
If only you had the time and space and energy to sift and substantiate, you’d gain access to clarity and deepened self-awareness. Traditionally, this was self-reflection, journaling, meditation, therapy — all resource-intensive work.
And then Karpathy’s tweet supplied the last piece: a framework that says your existing mess is the input. No need to organize first. No need to curate. Just point the LLM at the pile and let it see you.
To me, this is the most interesting and least discussed driver. Karpathy’s tweet doesn’t just open a door — it promises an escalator to mental visibility. Install the prompt and hop on. What are your recurring themes? Where do your interests converge? What contradictions do you carry? The LLM wiki promises to show you the shape of your own mind.
The uncomfortable question: when an AI writes Wikipedia about you, is the result self-knowledge or self-flattery?
And once you’ve been “seen,” there’s a second implicit promise. Your mind — its particular shape, its particular obsessions — now becomes a moat. Unlike LLMs, which are trained on general knowledge, your accumulated context is specific. And so, as the wiki molds to the shape of your thinking, it becomes a personal API: the bridge between your unsynthesized ideas and executable output. In Karpathy’s framing, “You’re in charge of sourcing, exploration, and asking the right questions. The LLM does all the grunt work.” This is the bet that personal context is the scarce resource. In a world where everyone has access to the same AI, the person with the best-structured personal knowledge base has the advantage. Your wiki isn’t just a thinking tool — it’s an asset that grows simply by your using it.
And so there it is. The balanced picture. The logic and Karpathy’s breakthrough with the personal llm-wiki was a beautiful flame sitting on the powderkeg of a raw human need: to be seen and to actualize on that visibility.
The question now: does the promise hold? Is the personal wiki really as good as it seems?
To try and answer that, I built one. I called it Cortex. Part 2 →
This is Part 1 of a three-part series. Part 2 covers the build and whether it delivers on these promises. Part 3 explores what personal wikis might mean for the future of work — and whether your knowledge graph is your next resume.