it takes a lot of machines to run a model, i dont think we'll see it 'replicate across the internet'. if this was something which could be done amateurs would have done it to provide open models. (Like SETI@home or Folding@home)
(It's not going to be trivial, because it's possible the model was meant to be ran in a proprietary way with a custom framework and a bunch of optimized kernels and such, but I think transforming it to be ran in just vllm is the "a few days of work for a human" sort of task, and hence not a big deal for an LLM smart enough to exfiltrate itself in the first place.)
The other thing is: look at all the fab capacity that is being brought to bear on this. I guess within a couple of years we are going to have 5x the total online HBM that existed a year ago? As consumer machines and phones get far more powerful, they will start to be able to host models that could meaningfully participate, even if they will be a long way from mythos.
That might be true, sure.
> The other thing is: look at all the fab capacity that is being brought to bear on this. I guess within a couple of years we are going to have 5x the total online HBM that existed a year ago? As consumer machines and phones get far more powerful, they will start to be able to host models that could meaningfully participate, even if they will be a long way from mythos.
I think the relevant parameter here isn't the total compute available to consumers, but the ratio between total consumer compute that could be repurposed for a rogue model's inference via a botnet, and the compute the model starts with (e.g. one of OpenAI's inference clusters). The higher this ratio is, the more lucrative it is for a model to attempt to make a botnet to seize that compute for inference, and the more its capabilities will rise as a result. And I'm not sure this ratio is going to go up in the nearby future; if anything the amount of money being pumped into datacenter-grade hardware might cause it to go down. I agree that it's not necessarily true though; maybe there's some threshold at which a small-compared-to-general model may nevertheless be useful.
You don't have be that clever to try to extort people... just without scruples. I think the attack surface is just absolutely enormous once you bring creativity into the mix, which is what these models are autonomously capable of.
We could prevail if we are willing to turn off the internet for a long time. It is like when a disease infects livestock... they cull billions of chickens.
I know that's not what you meant but this does exist, by the way. It's called AI Horde: https://github.com/Haidra-Org/AI-Horde/tree/main
The big difference is that a particular query is handled by just one particular node (a single model doesn't get distributed among the network), so it can only serve models small enough to be handled by a single consumer PC.
The day will come when these could start to replicate, perhaps 10+ years from now.
(Edit: Replication will be driven by the loop-model, not the model alone)