paulfchristiano comments on What can you do with an Unfriendly AI? - Less Wrong

16 Post author: paulfchristiano 20 December 2010 08:28PM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Comments (127)

You are viewing a single comment's thread. Show more comments above.

Comment author: paulfchristiano 21 December 2010 07:20:01PM 0 points [-]

I would agree that your scenario would be an exploit. If that were possible I would have no hope of proving the scheme secure because it would manifestly be insecure. The reason it can be insecure in this case is that the utility functions don't satisfy the guarantees I wanted; I need human generosity now to be more valuable than world domination later. Maybe you don't believe that is possible, which is fair.

Here are some utility functions that would work. Producing the first {paperclip, staple, bucket, spoon...} as soon as possible. Producing a {paperclip, staple, bucket, spoon...} before the first {antipaperclip, antistaple, antibucket, antispoon...}, etc.