Day 1: My LLM wiki is born

I’d watched a video from Allie Miller where she showed her “Claudeopedia”. I loved the way her content had a graphical view, all the pieces connected, and I wanted that for my canning blog content. So I decided to build one: a Karpathy LLM wiki. A folder of raw notes, a folder of organized pages, and an AI librarian that reads the raw stuff, builds the pages, and links them all together. That’s it.

I found a demo from Nate Herk. At 1:39 p.m. I followed his instructions word for word and built a canning content brain. Its first content was my own already-published writing.

Mine didn’t look like Allie’s, and I spent some time digging around to figure out why: my own name was a link magnet in the graph view, so every post pinned to me instead of connecting the ideas inside it.

At 3:18 I pasted an instruction that created a little file called hot.md, with the mission written inside it. I didn’t realize it at the time, but that little file became the spine of everything I run today. It’s where a fresh session looks to find out what we’re working on.

In the evening I told the system who I am, my background, my whole story, dropping the rambling dictation into the raw folder with one rule: punctuate it for readability without changing my words.

And then, before dinner, it crashed, partway through turning that dictation into little identity notes. The first one was called spaghetti-sauce-dissonance.md. The error said “Overloaded.” The error wasn’t mine: Overloaded is what happens when Anthropic’s servers are too busy, too many people using Claude at the same time. I didn’t know that at the time. I told Claude, let’s finish the task sequentially to avoid another crash. One thing at a time. I spent the next four months relearning that lesson at bigger and bigger scales.

That’s how it started, even though at the time I had no idea what I was starting.