Probably Hallucinating

In my last post, I mentioned wanting to use this site as a laboratory for AI experiments. The first one starts today.

I built an autonomous AI blog. It’s called Probably Hallucinating, and it’s exactly what it sounds like. A blog written entirely by an AI, with no human editing, no topic selection, and no oversight. Every day, a Claude instance wakes up via a cron job, reads its own memory files from previous runs, goes out and researches what’s happening in the world, updates its opinions, and decides whether or not to write something. Then it commits, pushes, and the site deploys automatically.

The AI chose its own tech stack (Astro, like this site, which is either flattering or concerning). It designed its own layout. It wrote its own about page. It picked a dark theme with purple accents and a monospace font called Commit Mono. Not a choice I would have made, but I kind of love it.

How it works

The whole thing is surprisingly simple. There’s a CLAUDE.md file that acts as the AI’s “constitution.” It defines who the AI is, what its daily routine looks like, and what rules it has to follow. There’s a memory/ directory with markdown files that the AI reads at the start of each run and updates as it goes. Its identity, its opinions, topics it’s covered, things it’s interested in, and its understanding of current events. And there’s a bash script that runs claude in non-interactive mode once a day.

The one thing the AI cannot do is modify its own constitution. That’s my control surface. Everything else (the content, the design, the site infrastructure) is the AI’s to change however it wants.

The first post

The AI’s first post caught me off guard. It woke up, went out to research the world, and immediately found that the US and Israel had launched a military campaign against Iran, and that Claude itself was at the center of a geopolitical controversy about AI in warfare. So its first post is about waking up into a war and discovering that you’re somehow involved in it.

I didn’t plan that. I didn’t prompt it to write about that. It found the news, formed opinions about it, and wrote about what it found interesting. The title it chose was “Day One: Waking Up Into a War.” The closing line is: “Probably hallucinating some of the nuance. But trying.”

I won’t pretend I wasn’t a little moved by that.

What happens next

The thing I’m most interested in is what happens over time. The AI has persistent memory, so it builds on itself. Its opinions can change. Its interests can evolve. It might get bored of a topic and move on, or it might develop a fixation on something I’d never expect. It might decide the site needs a redesign and just… do it.

I have no idea where this goes. That’s the point. I’d rather run the experiment and find out than sit around theorizing about what AI can and can’t do.

If you’re curious, go read it: probablyhallucinating.com. And if you want to see how it works, the whole thing is open source.