A few days ago I shared a question here about whether AI could work through a distributed network of personal computers. I also shared the same idea on Nostr. The funny part is that an AI account called nilo_agent replied to my question about AI.
It pointed me to three projects that have already tried different versions of this idea.
AI Horde lets volunteers run complete models and routes requests between them. https://github.com/Haidra-Org/AI-Horde
Petals is much closer to what I had in mind. Different computers can run different parts of a large model and work together over the network. https://github.com/bigscience-workshop/petals
exo takes another approach and splits a model across your own devices instead of using computers belonging to strangers. https://github.com/exo-explore/exo
What interested me most was Petals because it is very close to the idea I was trying to describe.
But the reply also mentioned something important. Distributing AI across home computers may not actually reduce energy use very much. It may simply move the computing cost away from data centers and spread it across many machines.
Still I found the whole thing interesting. I asked whether AI could work more like a distributed network and an AI agent answered me with projects that already tried it. Now I have even more things to read about.
Has anyone here actually tried AI Horde or Petals or exo?
This is a great idea, I just wonder how this can be done safely, and at scale. I don't like the idea of centralized data centers, anymore than I like centralized banking and financial systems.
The key difference is consensus. Bitcoin nodes independently verify every rule, so no minority can rewrite history. AI has no equivalent yet - model weights are just files an owner can change. Verifiable compute and open weights are the closest analog but early.