What are the best arguments for/against some of MIRI’s core positions.
We should worry about tool AI.
-Tool AI and oracle AI are different. Oracles are agents in a box. Tools are not agents, so they can’t take actions in the world or optimize an unfriendly utility function any more google maps optimizes a utility function. Why not just tell the AI to figure out physics/math/CS?
If it is an agent, why doesn’t Intelligence imply benevolence?
-emotions (like happiness/sadness) are vague concepts in the same way that objects are fuzzy concepts (think of invariant representations of faces). So, if an agent is intelligent enough to recognize fuzzy objects, shouldn’t it also accurately recognize fuzzy emotions (and realize when it’s doing something stupid like making ‘happy’ paperclips).
Thanks! I’d already read the first link, and remember thinking that it needed to be argued better.
Mainly, I still think people conflate tools with agents in a box. It seems obvious (in principle) that you could build an AI that doesn’t do anything but Math/CS/Physics, and doesn’t even know humans exist.
I’m planning on writing up my disagreements more formally. But first, I’m waiting on getting a copy of Nick Bostrom’s new book, so that I can be certain that I’m critiquing the strongest arguments.
I hadn’t seen the second link. I’ll definitely have to give it a more thorough read-through later.
What are the best arguments for/against some of MIRI’s core positions.
We should worry about tool AI.
-Tool AI and oracle AI are different. Oracles are agents in a box. Tools are not agents, so they can’t take actions in the world or optimize an unfriendly utility function any more google maps optimizes a utility function. Why not just tell the AI to figure out physics/math/CS?
If it is an agent, why doesn’t Intelligence imply benevolence?
-emotions (like happiness/sadness) are vague concepts in the same way that objects are fuzzy concepts (think of invariant representations of faces). So, if an agent is intelligent enough to recognize fuzzy objects, shouldn’t it also accurately recognize fuzzy emotions (and realize when it’s doing something stupid like making ‘happy’ paperclips).
Your first point was discussed in detail here. Your second point was discussed in many places on LW, most recently here, I think.
Thanks! I’d already read the first link, and remember thinking that it needed to be argued better. Mainly, I still think people conflate tools with agents in a box. It seems obvious (in principle) that you could build an AI that doesn’t do anything but Math/CS/Physics, and doesn’t even know humans exist.
I’m planning on writing up my disagreements more formally. But first, I’m waiting on getting a copy of Nick Bostrom’s new book, so that I can be certain that I’m critiquing the strongest arguments.
I hadn’t seen the second link. I’ll definitely have to give it a more thorough read-through later.