I am not sure where is this question coming from. I am not suggesting any particular studies or ways of conducting them.
Maybe it’s worth going back to the post from which this subthread originated. Acty wrote:
If we set a benchmark that would satisfy our values … then which policy is likely to better satisfy that benchmark...? But, of course, this is a factual question. We could resolve this by doing an experiment, maybe a survey of some kind.
First, Acty is mistaken in thinking that a survey will settle the question of which policy will actually satisfy the value benchmark. We’re talking about real consequences of a policy and you don’t find out what they are by conducting a public poll.
And second, if you do want to find the real consequences of a policy, you do need to run an intervention (aka an experiment) -- implement the policy in some limited fashion and see what happens.
Oh, I guess I misunderstood. I read it as “We should survey to determine whether terminal values differ (e.g. ‘The tradeoff is not worth it’) or whether factual beliefs differ (e.g. ‘There is no tradeoff’)”
But if we’re talking about seeing whether policies actually work as intended, then yes, probably that would involve some kind of intervention. Then again, that kind of thing is done all the time, and properly run, can be low-impact and extremely informative.
I am not sure where is this question coming from. I am not suggesting any particular studies or ways of conducting them.
Maybe it’s worth going back to the post from which this subthread originated. Acty wrote:
First, Acty is mistaken in thinking that a survey will settle the question of which policy will actually satisfy the value benchmark. We’re talking about real consequences of a policy and you don’t find out what they are by conducting a public poll.
And second, if you do want to find the real consequences of a policy, you do need to run an intervention (aka an experiment) -- implement the policy in some limited fashion and see what happens.
Oh, I guess I misunderstood. I read it as “We should survey to determine whether terminal values differ (e.g. ‘The tradeoff is not worth it’) or whether factual beliefs differ (e.g. ‘There is no tradeoff’)”
But if we’re talking about seeing whether policies actually work as intended, then yes, probably that would involve some kind of intervention. Then again, that kind of thing is done all the time, and properly run, can be low-impact and extremely informative.
--
Yep :-) That’s why GlaDOS made an appearance in this thread :-D