I split a language model across 3 Macs, on 3 separate networks, owned by 3 different people. Each one runs a slice. None of them holds the full model.

They still write. 6.6 tok/s over the open internet. I tried 10 shards and that works too.

Slow, and very much a toy. What got me is what it does to containment.

Today you stop a model by stopping a process. But the coordinator here barely computes, the other machines do the work, so the host can be the most humble device you own. It only has to be part of the coalition.

So what do you switch off? While some combination of devices can still reconstruct a full path through the model, it keeps running. The problem moves from finding a machine to making sure no viable coalition can form. Graphs and resources, not power switches.

Next to agents that are getting better at leaving their sandbox, that felt worth studying early.

Defensive write-up. No code, on purpose:
zenodo.org/records/21527132

✅ Polaroid Theme for PeekPaper.com