Really nice post! But I wonder how a "panic switch" for AI should be implemented.
Let's develop a quick example:
If you would build a robot controlled by an AI (of any stage) would have a button mounted on his back in order to stop it the button could to two things.

Option #1: It would "kill" the AI. In this case, the robot would do anything to prevent you as the operator from pushing.

Option #2: It would reward the AI, therefore it thinks it's done. In this case, the robot would try to push the button by himself or get you to push the button since its the easiest way to get the reward.

I guess the same thing would happen if you would replace the button with software triggers. And if you would limit an AIs software, would it still be intelligent?

Artificial intelligence is a really interesting topic and you post gives a great overview, thank you for sharing!