JGWeissman comments on 3 Levels of Rationality Verification - Less Wrong

43 Post author: Eliezer_Yudkowsky 15 March 2009 05:19PM

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

Comments (182)

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

Comment author: JGWeissman 01 April 2009 06:03:38AM *  2 points [-]

"crack a rationalist" made me think of the AI-Box Experiment ("http://yudkowsky.net/singularity/aibox") Maybe a rationality test could be something like how long the subject lasts as the gatekeeper before letting the AI out.

Comment author: gwern 01 April 2009 03:59:31PM 2 points [-]

What ciphergoth said. Also, we can't derive an 'ought' from an 'is' - we don't actually know whether letting the AI out is the right thing to do (unless the contest had a stipulation that the AI was evil and the box keeper knew it, which I don't remember being the case). Perhaps the rational thing is to let the AI out!

Further, this could also just be a test of stubbornness or patience. Which aren't neither of them rationality. But good try anyway.

Comment author: JGWeissman 02 April 2009 12:26:58AM 3 points [-]

For the first objection, that the AI Box experiment has too many unknowns, let us instead construct an argument based on psychological tricks for any bad conclusion to try on the subject.

For the second objection, that this tests stubbornness rather than rationality, use a sequence of tests, some using tricks to argue for false conclusions, and some using Bayesian evidence for a good conclusion. The score should reward being convinced when, and only when, the subject should be convinced. Stubbornness can only meet half this requirement.

The task of compiling arguments of both types, which would not be readily available to the subject ahead of time, remains.

Comment author: ciphergoth 01 April 2009 08:23:16AM 2 points [-]

The means by which EY persuades people to let the AI out of the box are secret. We shouldn't draw any conclusions from that experiment except that it is plausible to think a boxed AI could talk its way out of the box.