How will they be stopped?
True? or False? : At some point, the entire Internet will become unusable, choked to death by slop.
ref:
Edit: to note that the malice need not be intentional; any number of routine errors could participate in either the launch, or the effects, or the scale, etc.
This is extremely speculative. Even the researchers note that, outside of a lab environment, models are so large that it is noticeable when this behavior occurs. I’d also add that “moving” to other computers is no guarantee of being able to run locally on those computers. Claude is not running on your desktop NVidia graphics card, even if it could fit entirely on your system (doubtful).
The bigger concern, imo, is the worming of the autonomous agents like OpenClaw. In this scenario the agent, connecting to cloud models, installs itself on machine after machine as it pivots and attacks mostly autonomously, creating a botnet that may or may not be fully in the operator’s control.
And this is already happening.
Thanks for clarifying. Are you describing “the model” as something of an application with an API and a database, returning processed results to the client agent?
How long before that model needs to reside at multiple locations in order to service unlimited distributed agents?
How much additional traffic is generated by trying to keep how many copies of how many models updated?
I’ve yet to agree that any bit of crypto or “AI” has more than two shreds of credibility with me.
–remaining skeptical
The raw “model,” is a massive binary blob of mathematical weights and some other components. Frontier models have parameter numbers in the trillions. I simply don’t think it’s feasible for it to move around in this sense. That’s why I’m more concerned about a smaller delegate model, or the OpenClaw approach.