If those models can run locally and according to the article they're the same ones that do everything else - that means their models are quite small and cheap, which would be suprising. Even a machine running a 30B quantized LLM (which is generally the lower bound of usefulness) at meaningful speeds costs tens of thousands, and that's just the silicon, never mind building an actual robot around it.

No one did. I was saying that those models would be small, dedicated models like those possibly developed by a robotics team. OpenAI's general purpose models are not small, dedicated models.

It isn't even optional. Anything with a motor should have sufficient executive authority on board to decide on how to safely power down in case of trouble. Loss of communications is a major impediment to deploying robotics to the point that I think it should be outlawed to have robots that do not have all of their thinking done locally.

But I suspect that this will be one of those regulations written in blood.

[dead]