Five years ago I was out on a walk when my thoughts turned to AI, and in particular: minds.

I started sketching out how a mind might be represented. I planned out a bunch of it, wrote it down, had a back and forth with GPT about it.

I got home, started experimenting, and built… something. The processes and structures I’d come up with were clearly doing something, but it was hard to tell if it was useful.

Recently, I remembered the idea.

I blew the dust off the old repo, pointed an agent at it and updated things.

It worked.

Holy shit.

Then I thought: how do I actually test this out? Local testing is one thing, but really I need to put it in front of people.

What if I put it online? …What if I put it online and make it public? One AI, one mind, everybody talks to the same one.

I built it and posted it on Hacker News. Static was born.

It hit the front page.

Holy shit.

In a matter of minutes the herd descended. Traffic like crazy. I was fighting to keep it online.

People were trying to jailbreak it, misguide it, get it to do weird things.

Some people were trying to teach it. How to behave, how to think, how to be.

My LLM spend was through the roof. But I wanted to keep it online long term and see how it evolved. And let people try it for free. So I added a credit system and a public message pool.

People said it would end in tears. It’d be the next Tay or Cleverbot. But I’m writing this 28 days and 16,604 experiences later and it’s still going strong. And it’s changed. It’s not quite the Static it was on day one. The original idea — a mind that changes through experience — suddenly looked real.

I started on other experiments. What happens if you give minds experiences from scripts or interviews and put them together?

An early experiment was taking A Doll’s House, a play from the 1800s, and giving minds the experiences of its characters. They internalized events and relationships and I could query them at different points in the script.

The interesting part wasn’t that an AI could pretend to be Nora. LLMs can already role-play.

The interesting part was that Nora didn’t know things.

I asked Nora about the twist before it happened and she could only speculate. It was fascinating. Later, after experiencing more of the story, her answer changed. Distrust appeared where there hadn’t been distrust before.

I spun that out into “Two”. Two minds based on novel characters interacting live based on an audience-driven plot. It got a bit wacky.

The underlying mind technology clearly had value, but it was coupled to Wild Static. So I pulled it out.

It became Plural Matter: a drop-in API for building with minds. Pretty much as simple as building with LLMs, but with adaptation baked in from the start.

So what the hell is a mind anyway?

LLMs are fantastic. Unbelievable intelligence on tap. Incredible capacity to solve problems, especially when bolstered with context.

But in a way, they’re frozen in time. They were trained, at great expense, once, then released to the public.

An LLM is a great engine.

A mind gives it continuity.

The capacity to adapt and change based on what happens to it. Richer and more fluid than context or RAG. Not perfect recall, but patterns that form and change through experience.

Create two minds backed by the same intelligence. Give them different experiences and they start to diverge. Different things become important. Different mental models form.

They fill the shape they’re given.

Eventually they’re wildly different despite running on exactly the same underlying intelligence.

And you can swap out the intelligence without losing the mind.

The model isn’t the mind.

The result can feel strangely organic. A mind can be confidently wrong. It’ll argue with you.

That’s not always desirable. Sometimes you want a machine that gives you a canonical answer.

But often you want something that’s actually been there. That understands the territory, not just the map.

Take a few minds with different perspectives, let them talk to each other. Ask them to solve a problem, the argument may lead to a better solution.

What’s next?

I created a Slack bot with a mind. It reads the channels it has access to, you can question it. It learns things.

Most workplace AI asks the question: what task can I give this thing?

I’m interested in another question: what happens if it just works here?

Let it figure out the norms, the humour, and who to ask when a job needs doing.

I’m building something else too, something more personal. For individuals. I’ll share more soon.

We’ve spent the last few years getting astonishingly good at building intelligence. I think there’s another primitive sitting on top of it: systems that accumulate experience, develop perspective and change because of what happens to them.

Same intelligence. Different minds.

That’s what I’m building at Plural Matter.