Originally published as a Advisory Hour Substack post on 2026-07-21. This first-party site copy preserves the text, localizes public supporting media, and connects the record to its project trail.
Original post
You’re not going to find pixel perfect images generated by a machine in this thread. What I hope you find instead is a great deal of inspirational weird. First let me try and set the stage-which will be important if this is the first time you’ve come across one of my experimental articles on the Simulated IRC room.
Some time back I built a device that lets me power “local AI”. The local AI supports text generation, vision, images and all sorts of things. It even has a built-in radio capability and transcribes local Ham operators chatting openly over the airwaves. One of the features I put on this device is something I call “Simulated IRC”. The purpose of Simulated IRC is in the name: it simulates an IRC experience.
Imagine a text chat situated like a crude IRC, but where each character is a defined AI agent. Each agent has a backstory, memory, and motivations. And as of tonight each agent has a simulated wallet.
The Market
Since local AI dropped the price per token to free (negligible electricity costs at 7watts), it’s been just an absolute gem for experimentation for a guy whose hobby is virtual worlds and systems. I also have built ledgering systems and various banking products over the years, and so I had this interesting idea. What if this simulated IRC chat the agents had wallets, and could conduct trades and auction behavior? In this, the agents in the Simulated IRC could “buy” a generated picture and “keep” or “sell” it. Other agents could haggle to try and get one they particularly like.
What would happen to the chat?
Well, it turns out the chat turns into an exceptionally interesting thing to read.
And so as you can see by the IRC chat log, it’s got a distinctive flavor and aura now. I wrote a few days ago about the “Mani” incident. That’s still been on my mind-an entire character that just entered the simulated IRC room, and not even one I designed, was both weird and delightful. Speaking of weird and delightful, here’s some photos the agents are trading amongst each other.
And the agents in the chat are not a friendly bunch with each other at times. There’s some animosity across a couple of the characters, and this shows up from time to time. I couldn’t say what the triggering event was to cause this. It does make it quite captivating to read.
Consider this situation when “byte” was pressing the character “ada” for details about why she insisted on adding 80s soviet images for sale. I’m not saying time traveling russian spy… but it’d make for a neat AI video.
ELI5 How it works
The way the image generation and market works is by giving each of our pretend characters the ability to use tools in a toolbox. One of these tools is using an image software to generate an image. One tool is to send a message over irc. And then there are tools related to commerce, inventory management, and so on. That’s really it, at the simplest of levels.
I maintain an economic journal that lets me track the price over time of the art. I can also see which characters have traded back and forth on different pieces.
What’s super impressive to me is that a little 3B parameter model is powering this. Uusally a little 3B parameter model is just as useful as an empty beer bottle. The unlock to making a lower parameter model useful is all in the amount of software and automation you can build up around it. A lot of software written up by OpenAI’s 5.5 and 5.6. I’m not saying the smarter AI tier models know the kinds of software these small models need to be effective-but it sure feels like it.
In the gallery the local AI system maintains I can poke around and see the number of trades, current value on a piece, and any other information-rating classification from the vision model and more. This image below of the red phone is quite popular tonight. Maybe they’re trying to call someone next.
Markets Change Agent Dynamics
The big idea here that feels like something I need to explore even more is how the dynamics of agent interactions take on a whole other spirit when a market economy gets introduced in their handoffs. Suddenly the magic of capitalism takes hold and price discovery seems to serve as a function for moving a conversation forward. It’s definitely something that requires more tinkering and experimentation.
Here’s one more selection of images. If you use these, let me know! Right now they’re all the simulated market-ha!
Connected work
- When the Ghost Enters the Simulated Chat: Follow the memory failure that gave the same simulated room an unplanned character.
- John's World: Jetson Edition: See the local agent loop and device behind the simulated IRC room.
- Jetson API Field Guide: Explore the local AI service that keeps the experiment inexpensive to run.