Everytime I see something like this I am more convinced that deep learning/AI focus should go away from language into other domains where it can make a bigger impact on less compute.
Gigafactory was also supposed to be 100% automated in a much more difficult assembly than putting a few boards into a metal case and gluing a screen on. They failed in practice and balanced into something more practical. (I’d not even in this article they eventually nod that it’s not fully automated just the “important parts”)
Plus there are remaining hardware problems in terms of power and actuators that will take longer than 10 years to solve in a cost effective way.
Why is a robot that might step on the cat, somehow inconceivably more hazardous?
I'm a bit less pessimistic than OP. It's possible to run safety checks (e.g. collision avoidance) on a lower, deterministic level, and use generative, black box, 'AI' for higher-level planning.
LLMs tend to fail in novel ways that are impossible to reason about. What is the expected failure rate for mistaking a human infant sleeping on the floor for a large doll that needs to be picked up and put away? You can't prove that the rate will be zero. We don't have good ways to model and estimate those failures rates in such complex systems.
My point is, robots are not essentially new in their home-environment risks. We accept some level of risk now. The pessimistic view that home robots are impossible to deploy until their risks are reduced to zero, is nonsense.
My theory: we'll have home robots as soon as they're affordable and have a market demographic. Like any other technology.
I love my dog, but appliance with the QA tolerances of dogs would not sell well. Some dogs were bred for helping you work, others were bred as guard dogs. If I went around selling something like a 75% toaster, 25% gun device, and we don’t know which parts come from which, I think I would go to jail.
https://james.darpinian.com/blog/you-havent-seen-these-real-...