> Rather, it’s the purely procedural criticism of their making a confident assertion as such. The thought seems to be that observers can recognize this as epistemically out of bounds without themselves knowing anything about the issue at all.
A month ago I made a bet against Roman Yampolskiy that…
> Rather, it’s the purely procedural criticism of their making a confident assertion as such. The thought seems to be that observers can recognize this as epistemically out of bounds without themselves knowing anything about the issue at all.
A month ago I made a bet against Roman Yampolskiy that his credence of existential catastrophe from AI is too high. I claim that a layperson without specific knowledge of AI risk can know that his credence is too high. In this case, I think "confidence-policing" by this person would be valid. Do you agree? Is there something different going on here making this not confidence-policing?
"I bet Roman Yampolskiy that his p(doom from AI), which is currently equal to 99.999999%, will drop below 99.99% by January 1st, 2030. If I am right, Roman owes me $100. If I am wrong, and his credence does not drop below 99.99% by 2030, then I owe him $1.
""Bet accepted. To summarize the terms: If your estimate of existential risk due to AI drops from your current credence of 99.999999% to less than 99.99% before 2030, you owe me $100. If it does not drop below 99.99%, I will owe you $1 on January 1st, 2030.
Your current credence, p(doom)=99.999999%, implies that there is at most a 0.01% chance that your credence will ever drop to 99.99%, which is why you risking your $100 to my $1 seems profitable to you in expectation.
On the other hand, this bet seems profitable to me because I think your p(doom)=99.999999% is irrationally high and think that there is >1% chance that you will recognize this, say ""oops"", and throw the number out in favor of a much more reasonable p(doom)<99.99%."""
Yeah, I think that's different: you're making an alternative claim (perhaps based on general background knowledge) about what substantive credence is most reasonable. That seems very different from the kind of purely procedural confidence-policing I have in mind. For the latter, the policers would generally be unwilling to take any bet precisely because that would be *committal* in a way that they are opposed to.
Confidence-policing is a new concept to me:
> Rather, it’s the purely procedural criticism of their making a confident assertion as such. The thought seems to be that observers can recognize this as epistemically out of bounds without themselves knowing anything about the issue at all.
A month ago I made a bet against Roman Yampolskiy that his credence of existential catastrophe from AI is too high. I claim that a layperson without specific knowledge of AI risk can know that his credence is too high. In this case, I think "confidence-policing" by this person would be valid. Do you agree? Is there something different going on here making this not confidence-policing?
My bet (archive.is/izuQ2):
"I bet Roman Yampolskiy that his p(doom from AI), which is currently equal to 99.999999%, will drop below 99.99% by January 1st, 2030. If I am right, Roman owes me $100. If I am wrong, and his credence does not drop below 99.99% by 2030, then I owe him $1.
""Bet accepted. To summarize the terms: If your estimate of existential risk due to AI drops from your current credence of 99.999999% to less than 99.99% before 2030, you owe me $100. If it does not drop below 99.99%, I will owe you $1 on January 1st, 2030.
Your current credence, p(doom)=99.999999%, implies that there is at most a 0.01% chance that your credence will ever drop to 99.99%, which is why you risking your $100 to my $1 seems profitable to you in expectation.
On the other hand, this bet seems profitable to me because I think your p(doom)=99.999999% is irrationally high and think that there is >1% chance that you will recognize this, say ""oops"", and throw the number out in favor of a much more reasonable p(doom)<99.99%."""
Yeah, I think that's different: you're making an alternative claim (perhaps based on general background knowledge) about what substantive credence is most reasonable. That seems very different from the kind of purely procedural confidence-policing I have in mind. For the latter, the policers would generally be unwilling to take any bet precisely because that would be *committal* in a way that they are opposed to.