pull down to refresh

Given how trivial it would be to tie an LLM in to a bot-to-bot IRC network, I'm surprised it hasn't been done. Add to that a layer using something like Lightning to manage resource allocation (funding individual nodes) and we could have a massive distributed LLM to act as a super-mind.
I could see it working like this: a request comes in with some number of sats along with the request. That request gets propagated through the network with different nodes handling different layers of the interpretation and processing. When the funding gets used up, the network returns the request to the originating node as a response.
It doesn't need to be IRC, either. Using Matrix would allow us to branch out into multimodal models for processing non-text data and the like.
You can simulate this with turn-based interaction in a simple python script. I've done some experiments with this and in general you want to have extremely low error rates and extremely high error recovery, just like with agentic solutions, which honestly, we do not have today, even though it came along a bit since I first tried (with llama2.)
I've for example experimented with making LLMs use the socratic method to drill down a problem to its core between different agentic roles, but it is hard to get consistency in the face of high error rates; with llama2 the success rate was under 50% in a 5-turn script, similar with qwen3 though the smaller distills for Qwen are much worse than for other models for some reason.
There are perhaps some optimizations to make, but I'm shy to introduce additional authoritarian roles, and that wouldn't help your IRC idea either.
reply
I really like the idea of getting models to dial down on requests with the Socratic method. Did you publish your experiment anywhere?
reply
Don't think I did. I'll prepare a gist after the next round of playing with it.
reply