Using AI /
Building a second brain the model keeps tidy
The same method works for one person. Karpathy's own example was not a company. It was his own research. He has been indexing papers, articles, repos and images into a raw folder, and having a model compile them into a personal wiki he can then ask questions of.
I do a version of this already, and I think it is one of the more quietly useful things you can set up for yourself. Not because it makes you more productive in the LinkedIn sense. Because it lets your curiosity finally add up.
A pile in your head does not compound
Everything you have ever read closely, every rabbit hole you fell down, every problem you solved at two in the morning, is worth something. The trouble is where it lives.
It lives in your memory, which is leaky, and in forty open browser tabs, and in a notes app you stopped trusting. Naval Ravikant calls the good version of this specific knowledge. The kind you cannot be trained for, found by following real curiosity rather than whatever is fashionable. It is your edge precisely because nobody else has the same pile.
But a pile in your head does not compound, and it cannot be queried. You re-read the same article a year later because you forgot you ever read it. You half-remember a fact and cannot find where it came from. The knowledge is real and almost entirely inert.
What the setup actually looks like
It is less elaborate than it sounds. There are three parts.
- A raw folder. Everything you read goes in as plain text or markdown. Karpathy uses a browser clipper to turn web articles into local files, and pulls the images down too so the model can see them. The point is to own a local copy, not a link that rots.
- A compiled wiki. You point a model at the raw folder and have it do the work I described for a company: summarise each source, write articles for the concepts that keep coming up, and link them together. You read the wiki. You barely write it.
- A place to view it. Karpathy uses Obsidian as the front end, because it renders markdown and the links between notes as something you can walk through. The tool barely matters. A folder of markdown is the durable thing.
That is the whole machine. A folder you feed, a model that compiles, and a reader you browse.
Then you can ask yourself questions
Once the wiki is a few hundred notes deep, it stops being storage and starts being something you can interrogate.
You can ask what you actually think about a topic, drawing on everything you have read about it. You can ask the model to find the connection between two rabbit holes you never realised were related. You can have it render the answer as a short brief, a set of slides, or a chart, and then, as with a company brain, file that output back in so the next question starts from it.
Your explorations add up. That is the whole point. A year of reading stops being a year of forgetting and becomes a body of thought you can stand on.
It is allowed to be for fun
I want to be honest about the register here, because the productivity framing misses the better reason.
You do not have to build this to ship more or earn more. The plainest reason to keep a second brain is that it is satisfying to watch your own curiosity accumulate into something. To ask a question and have a year of your own reading answer it. To follow a thread on a Sunday and know it will still be there, connected to everything else, the next time you reach for it.
The compounding is real and useful. But the joy of it is reason enough on its own.
The rule
Stop trusting your memory to hold what you learn. It leaks, and the best things you know are the easiest to lose.
Keep one folder of what you read. Let a model compile it into something you can browse and question. Then go back to being curious, and let the machine make sure it all adds up.