Benja comments on An angle of attack on Open Problem #1 - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (84)
I played around a bit with quining solutions, and came up with the following, which solves this toy problem fairly well:
AI_Q2(3) should double down on AI_Q2(4) as well as AI_Q2(4^^^^4). (As well as variants that are provably equivalent to itself like speed/size-optimized versions of itself.) I sent this to Benja last night and he responded with (in part) "You've succeeded in making me uncertain whether quining approaches could actually be directly useful in solving the real problem (though it still doesn't seem likely)."
I agree it doesn't seem likely the real problem can be solved with quining approaches, but I'm post this solution here in case anyone can extend the idea further. At the very least it should be interesting to understand why quining approaches don't work on the real problem. What relevant aspect of the real problem isn't being captured by this toy problem?
I should be explicit that I don't assign strong confidence to either of the following intuitions, but FWIW:
One point where intuitively this may be more likely to fail than my trick is when you try to replace yourself by something that's expected to be better than you at maximizing its expected utility with bounded computational resources. If you want something better than yourself, it seems rather limiting to require that it won't do anything you wouldn't do. (That's an argument against quining approaches being the right thing, not for my approach being that.) -- On the other hand, an intuition in the other direction would be that if your original AI has a parameter n such that for n -> infinity, it converges to an expected utility maximizer, then your approach might still lead to an increasing sequence of better and better EU maximizers.