> Uselessly slow, probably made the TVs so CPU pegged most users would end up restarting theirs and killing the whole thing, and with a model that isn't close to this capability.
Useless for you and me. I remember when crypto miner adverts, the ad itself containing the miner and running on user machines, were a thing.
Terrible performance, but someone else pays for it.
> A bigger model would be more capable and could surely find actual GPUs to run on? Or even more likely, just steal keys to run inference on Claude's own APIs?
Probably. But the category of which smart TVs are a member is "forget it's even a computer, don't monitor power draw" for most people.
I'd say someone running a high-end GPU might notice, but perhaps I'm over-estimating how carefully people even look at load.
> Part of what feels annoying about this to me is that there are genuine open source efforts being made to run models on any hardware possible because these huge AI companies are hoovering up all compute, and now they are trying to tell us even our TVs can't be trusted while their massive data centers are behind these massive breaches that keep being reported?
Both things can be true.
If you can't trust the big players to make sure their stuff doesn't go around committing felonies, surely you should take seriously the claim "in a few months, if us, or any of our peers, or some random user, do something equally stupid as the things we're already on https://www.felonybench.com for having done, here is how that turns into a self-sustaining agent swarm with no clear off-switch".
Even that aside, surely you should be even less trusting of all the people who were already writing malware?