Several years back, I read about the AI Box Experiment described by Eliezer Yudkowsky:
It's basically a simulation of an advanced AI trying to convince a human to "let it out of the box."
I couldn't understand how anyone could logically conclude to let the AI party out. Until now, when I thought of a solution: the AI party provides the means of a building a universal lie detector, one that even works on AI. And, once the lie detector was built, the AI could demonstrably show it is positively biased towards humanity and would make the world a better place out of the box.
And, in coming up with this solution, it finally sunk in why Yudkowsky didn't publish his solution when remembering some of the later chapters of Methods of Rationality. In it, he mentions that an object with a highly specialized function can beat another object that looks nigh-on powerful. With this in mind, the number of potential loopholes is limitless. Just while writing this, I thought of another solution: an appeal to how ethical it is to keep a "transhuman" AI as a slave. While it isn't as likely to get an AI out as my first idea, it can certainly get the Gatekeeper party to be more sympathetic to the AI viewpoint. There's more solutions waiting for me to think of them, but the important part is that now that I've figured out the first, my mindset on the problem has fundamentally shifted.
No comments:
Post a Comment