The only thing that makes sense is they are suggesting they can attack home assistant servers and use vulnerabilities in the smart home devices as a way in.

From there if your home assistant server has any decent power, or at least a npu then maybe it could do some self replication.

If they used my computer it would take 10+ hours before their ai produced anything worth while, and I would notice it getting hot and unresponsive before that point. They're going to need to stream a 300gig model or whatever it is to me also and I just dont have the internet for that. Either that or force my computer to use all my free storage to hold the model to run locally.

Sure I guess it could happen but it seems impractical. If models that where capable didn't require the infrastructure to run a city then I would be worried. But local models are kind of useless, a standard 16 gig GPU is the bare minimum to do things shittily. They want 128 gig vram ideally and then it can shit out code that is as good as free frontier models you would find on duck duck go. And those barely out put quality workable code. You would need an agent running itself for days.

No, they are saying these local models are scary, and instead everyone should be using these nice and protected models in datacenters instead, that fortunately Anthropic will rent access to out of the goodness of their conscious...