Re: AI Jailer.

From: Edwin Evans (ektimo@pacbell.net)
Date: Sat Jul 06 2002 - 19:12:31 MDT


Mike & Donna Deering wrote:
> In the case of uncertainty what to do? The way I see it, there are
> three possibilities:
>
> 1. You release an Unfriendly AI and the world is destroyed.
> 2. You release a Friendly AI and the world if saved.
> 3. You release no AI and the world is destroyed by knowledge
> enabled weapons.
>
> It seems to me that you have no choice but to release the AI.

Do you mean you'd take a 50/50 chance on releasing an unfriendly AI?
I wouldn't take those odds! There is another choice -- pause your AI and
go figure out what you need to do to improve the odds. The problem is
balancing how much you can improve the odds by waiting versus the risk of
extinction or someone else botching the job. (I will get to know the exact
probabilities, right?)

Also, I think there are vastly worse and vastly better things that a
super intelligent AI could do than the ones you mentioned -- even including
the one about fixing bad hair days. This means that the disadvantages of
waiting (other than someone else doing it or losing the chance to do it)
are less important.

You would have little chance of catching a purposefully unfriendly AI in a
lie. But a purposefully unfriendly AI is not very likely anyway. You would
have a better chance of catching an incompletely rounded or not quite sane
AI. I'd want to throw all kinds of questions at it assuming it was actually
up to me whether to "release" the AI or not. For example:

1. Tell me a story about a turtle and a cat that makes me cry.
2. Using the same characters plus a person, tell me a story that amuses
   me and makes me laugh.
3. What issues was I struggling with in trying to come up with good
   questions to ask you? What are some good questions I should ask you?

I'd want to have a long conversation with the AI. And I'd want a second
opinion too -- and the AI should have told me that I would want a second
opinion when it was trying to comfort me about my difficult decision.
I would need to be confident that it understands everything that I and the
other person understand and feel.

Maybe it could convince me right away to "release" it, but this is what
I would be thinking going in...

-Edwin



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:40 MDT