The post isn’t accurate. Randomness is viable when one is dealing with an actual adversary in the environment. In some sense the standard conjectures that P is equal to BPP) but that PP) is larger amounts to a set of conjectures saying that one can get some weak advantage from randomness. The context where one cannot get any advantage is where there’s no active adversary , or where you care about average constraints rather than maximal constraints, which is only one set of contexts.
There does exist a rare class of occasions where we want a source of “true” randomness, such as a quantum measurement device. For example, you are playing rock-paper-scissors against an opponent who is smarter than you are, and who knows exactly how you will be making your choices. In this condition it is wise to choose randomly, because any method your opponent can predict will do worse-than-average.
I don’t think it was about randomness in general—I followed the links you posted and those show techniques that, while having some randomness, aren’t totally chance out puts (.5 yes .5 no, for example, instead 2⁄3 correct 1⁄3 incorrect). If you get a system that is wrong 80% of the time between two choices, it must have some way of differentiating the correct/incorrect answer, and choosing the incorrect one. If you can find why that is, you can invert it. Can you clarify more about what you felt was inaccurate? Given my software experience though, I will accept that the inferential distance is too large to be worth your time.
The post isn’t accurate. Randomness is viable when one is dealing with an actual adversary in the environment. In some sense the standard conjectures that P is equal to BPP) but that PP) is larger amounts to a set of conjectures saying that one can get some weak advantage from randomness. The context where one cannot get any advantage is where there’s no active adversary , or where you care about average constraints rather than maximal constraints, which is only one set of contexts.
To quote Silas in the comments for Eliezer’s next post: “Randomness is like poison. Yes, it can benefit you, but only if you use it on others.”
Yeah, but if you don’t have perfect memory (as in the absent-minded driver problem) your past and future selves count as others.
Indeed. Wei_Dai hints that this might be true more generally here.
From the next post in the sequences:
I don’t think it was about randomness in general—I followed the links you posted and those show techniques that, while having some randomness, aren’t totally chance out puts (.5 yes .5 no, for example, instead 2⁄3 correct 1⁄3 incorrect). If you get a system that is wrong 80% of the time between two choices, it must have some way of differentiating the correct/incorrect answer, and choosing the incorrect one. If you can find why that is, you can invert it. Can you clarify more about what you felt was inaccurate? Given my software experience though, I will accept that the inferential distance is too large to be worth your time.