Would I like a humanoid in my house?

Context from the virtual world

• My US trip: asked ChatGPT to visualize everything for the plan we had planned together. Came back after half an hour and there it was: the whole website, with the map, day-by-day plan, everything.
• It's so much better than the Markdown plan I had before.
• We talked about different paradigms for AI. I think we're there. If you're doing anything complex, just ask AI to create a website for it. So much easier to see, and it can do that on demand, so quickly.
• We could even have replies be mostly HTML files instead of Markdown.
• I didn't spend any cognitive energy. I said "hey, visualize the plan," came back half an hour later, done.
• I want that for the physical world.

What I want in the physical world

• Hey, wash my stuff, change the bedding, clean up the dishes, ... and then I focus on my own work and reading or whatever, and the robot is just doing that.
• Capability: we're not there yet, but it could be anytime soon.
• The other thing is whether I would actually put that in my home. And that is where security comes in. I think that's actually so crucial.

How they should exist without being very scary

Not the cloud. You don't want the thing controlled from the cloud, because if someone hacks that, you have a very big problem. It can do basically everything.

Not the machine either. If the machine is complete in itself, has everything it needs in itself, that's a risk too. You'd have to trust that the model is well aligned and that all the safeguards are in there, so that it doesn't do anything from the dystopian novels.

So both local and cloud are bad. The middle ground is something else:

• A server room in your house that isn't connected to the outside internet. That's the robot's brain, and the robot works from there.
• Server room automatically shuts off as soon as there's any attempt to attack it, physically or in a cyber way.
• The robot would have to be so smart that it finds a way to access the thing without it noticing. That's pretty hard.
• You would not give the robot as much intelligence as the thing that was used to build the software. Yeah, the robot can use a computer, use the smartest AI, whatever—but it's much more difficult.
• No cables connecting to it, so it's only wireless. It doesn't connect to the internet.
• So it would be quite difficult for the robot to do any harm there. Of course it still needs to be refined, but that would be significantly safer.
• A module in the robot: if the robot does anything evil, it shuts it down.
• In the server room: ablation, mechanistic interpretability. It cannot hide its thoughts. That's a thing humans can do, until now—question is until how long, but for now we can do it and the robot can't. So immediately, if there is something like that, it gets shut off. No chance at all.

Then safeguards for the motors:

• Say we fixed alignment of the big thinking model. The model could still just by accident do something bad. That is a mechanical issue, we need really some sophisticated safeguards.
• Probably we can learn a lot from autonomous driving, because autonomous driving needs a lot of safeguards to be approved. If they can manage to work it out, I think there's a good chance we can manage to work it out for robots as well.
• And we will have so much intelligence in the coming years, so it should be much easier.
• I really think it's a soluble thing.

Why I want them to exist

• It's amazing how much resources this frees up for literally everything else.
• Combined with guardian angels (more on this here). That would free up even more. You could send it to interviews, or shopping for you. It knows what you like, it knows who you are. Agents meeting agents.
• Then you would really say no to everything, or almost everything, in person. Your agents and your robots do it, and you can really focus on the things you want to do.
• There's this principle, hell yeah or no. Either it's a complete "yeah, I want to do it," or it's a no. Because otherwise you won't find the time to do everything you really want to do.
• It's a beautiful principle, but right now it's only possible for rich people. You can't say no to everything—if you say no to shopping or cooking, then you don't eat.
• So it can't work for everyone without the robots. That's why I want them to exist.
• And it will be possible in that age. That's what I'm looking forward to.