FIELD STATION experiment #001, revised: a roomful of computers running AI together
We're switching FIELD STATION experiment #001 from co/core to SwarmLLM: one model, running together on computers we're already using.
Most people using AI tools every day are, often without thinking about it, sending their questions and tasks to a handful of huge companies' data centres.
In July, we published FIELD STATION experiment #001: running our own slice of an AI co-op. We planned to use co/core, an AI inference co-operative, to test whether a group of people could run AI work for one another on hardware they already own.
Having spent more time with it, we have decided that co/core is not the right tool for the experiment we want to run.
The question remains:
How might we run AI in a different way?
What changed
The original experiment was about exchange. One person's computer could do AI work for someone else in the group. Participants could be both users and providers. The work could move between machines, with a public record of completed jobs.
There is still something worth testing in that arrangement. It asks whether AI infrastructure might be organised around reciprocity, rather than the familiar pattern of subscription, platform dependence, and remote service provision.
For this first FIELD STATION session, though, we want something smaller, more immediate, and easier to experience together.
Instead of asking people to exchange AI work across a network, we want to ask what happens when a group runs one model together, across the computers already switched on in the room.
One model, shared by the room
We are going to use SwarmLLM.
SwarmLLM lets a group create a shared room, connect devices with a short code, and divide a model across those devices. Each device takes a slice of the model, with more capable machines taking more of the work. The group then sees the output together.
It is a little like BitTorrent, though the comparison only goes so far. Rather than sending everything to one remote machine, the task is distributed across a group of connected computers.
SwarmLLM describes this as "one model, shared by the room". It does not require accounts, and it says that participants' words do not leave the room.
This does not mean that a group of laptops becomes equivalent to a hyperscale data centre. It does not make AI free of material costs, energy use, supply chains, or the damage associated with the production and disposal of computing equipment.
The narrower proposition is more interesting.
If our computers are already on, and a group wants to use AI together for a bounded piece of work, can those machines become the place where the inference happens?
That changes some relationships.
The model runs on devices that people in the group own or control. The people involved can see the immediate infrastructure. The group can inspect the arrangement, experience its limits, and decide whether it is useful.
From exchange to a room
co/core led us to think about a network of people exchanging AI work. SwarmLLM leads us to think about a room, including an online room.
That makes it a better fit for what FIELD STATION wants to do next.
We are interested in practical encounters with different technological arrangements, not only arguments about them. We want to put people in contact with the machinery, the friction, and the trade-offs usually hidden behind an API, a chat box, or a monthly bill.
A short group session will not settle the question of digital sovereignty. It will give us something to examine together.
- Did it work?
- Was it slow?
- What did it require from people and their devices?
- What did it feel like to use a model assembled from the computers in the group?
- What kinds of work suit this arrangement?
- What kinds of work do not?
- Does this make a meaningful difference to the relationship between people, AI, and infrastructure?
Join the session
We are convening an online getting-started session for people who want to try this with us.
How might we run AI in a different way?
Tuesday 15 September, 14:00 BST Online Around 60 minutes Free
This will be a small, practical session. We will introduce the original co/core experiment and explain why we are changing direction. We will set up a shared SwarmLLM room, connect participants' devices, run some prompts together, and discuss what happened.
You do not need to arrive as an AI engineer. You will need:
- A computer you can use during the session
- A reasonably stable internet connection
- A willingness to try something that may be a little rough around the edges
We will cover:
- The original co/core experiment and the reason for the change
- Distributed inference and the idea of a model shared between devices
- What it means to use computing devices that participants already own
- Privacy, ownership, energy, infrastructure, and digital sovereignty
- The practical friction involved in making this work
- What we should test next
Places are limited by the practicalities of running an experiment together, rather than by any desire to create scarcity.
What we will publish
As with the original experiment, we will write up what happens in public.
That includes what works, what breaks, what creates friction, and what we learn from trying to run AI in this way. We are not looking for a polished demonstration of a finished alternative. We are looking for a careful, public account of a small experiment.
The question remains the same:
How might we run AI in a different way?
This time, we are starting with the computers we're already using.