I’m not sure this matters at all, but I tentatively think that percent of value of the best possible future, (we can call it ‘PVBPF’) which I’ve heard brought up by Will MacAskill a couple times now[1] (and maybe by others?) isn’t an ideal metric by which to measure or even mental/conceptual model by which to understand how good the world is.
My main concern is illustrated by the following:
Suppose we do a survey in 100 years[2] and conclude that the current world is, to our best guess, 50% as good as the best possible counterfactual[3]. So PVBPF = 50%
Then somehow we realize that actually there was actually ex ante a p=10−1,000,000 chance of a counterfactual outcome 5 times better than what we previously thought was the maximum of the distribution
What was originally PVBPF = 50% now becomes 50%∗(1/5)=1/10. So PVBPF = 10%
So the addition of an astronomically unlikely outcome- one that for all intents and purposes doesn’t change the ex ante the EV at all—radically effects our assessment of PVBPF.
Maybe this is fine—I’m not really claiming there’s a technical issue here (although maybe there is; more to think about).
But maybe there’s a vibes issue: 10% and 50% sound like really different numbers, and they are! But a 5x difference in PVBPF can correspond to either a 5x difference in actual moral value or the addition of an arbitrarily unlikely 5x-as-good-as-previously-believed-maximum counterfactual.[4]
So the concept of viatopia is that it’s a state of society that is on track to produce a near-best future — something that’s at least 90% as good as a future that we could have.
To get around the fact that I think the current world is net-negative. That’s a modeling inconvenience but one that I do think one that neither is nor reveals a fundamental weakness with the idea being gestured at.
I.e. as we make p=10−1,000,000 smaller and smaller, PVBPF doesn’t change but the delta between ex ante EVs represented by the original guess of the distribution of probabilities and moral values and the updated guess goes to zero
I’m not sure if this would just create other technical problems I haven’t thought of, but one solution could just be to replace “best possible” with, e.g., “ex ante 99.9th percentile”. Then you wouldn’t have this discontinuity from changing the max, but it captures basically the same intuition?
(One problem is that it’s not obvious what the threshold should be. But my guess is that when people are trying to figure out the relevant denominator for PVBPF they’re usually implicitly truncating the probability distribution over outcomes somewhere anyway, since it seems very hard to reason about the absolute maximum.)
Claude also suggested something like a conditional expectation of the upper tail, e.g. E[V|V≥q0.999], which seems interesting but I’ve not thought carefully about.
I’m not sure this matters at all, but I tentatively think that percent of value of the best possible future, (we can call it ‘PVBPF’) which I’ve heard brought up by Will MacAskill a couple times now[1] (and maybe by others?) isn’t an ideal metric by which to measure or even mental/conceptual model by which to understand how good the world is.
My main concern is illustrated by the following:
Suppose we do a survey in 100 years[2] and conclude that the current world is, to our best guess, 50% as good as the best possible counterfactual[3]. So PVBPF = 50%
Then somehow we realize that actually there was actually ex ante a p=10−1,000,000 chance of a counterfactual outcome 5 times better than what we previously thought was the maximum of the distribution
What was originally PVBPF = 50% now becomes 50%∗(1/5)=1/10. So PVBPF = 10%
So the addition of an astronomically unlikely outcome- one that for all intents and purposes doesn’t change the ex ante the EV at all—radically effects our assessment of PVBPF.
Maybe this is fine—I’m not really claiming there’s a technical issue here (although maybe there is; more to think about).
But maybe there’s a vibes issue: 10% and 50% sound like really different numbers, and they are! But a 5x difference in PVBPF can correspond to either a 5x difference in actual moral value or the addition of an arbitrarily unlikely 5x-as-good-as-previously-believed-maximum counterfactual.[4]
Eg on the latest 80k podcast, Will says:
To get around the fact that I think the current world is net-negative. That’s a modeling inconvenience but one that I do think one that neither is nor reveals a fundamental weakness with the idea being gestured at.
You might be able to make this more well-defined via the many-worlds interpretation of quantum mechanics
I.e. as we make p=10−1,000,000 smaller and smaller, PVBPF doesn’t change but the delta between ex ante EVs represented by the original guess of the distribution of probabilities and moral values and the updated guess goes to zero
I’m not sure if this would just create other technical problems I haven’t thought of, but one solution could just be to replace “best possible” with, e.g., “ex ante 99.9th percentile”. Then you wouldn’t have this discontinuity from changing the max, but it captures basically the same intuition?
(One problem is that it’s not obvious what the threshold should be. But my guess is that when people are trying to figure out the relevant denominator for PVBPF they’re usually implicitly truncating the probability distribution over outcomes somewhere anyway, since it seems very hard to reason about the absolute maximum.)
Claude also suggested something like a conditional expectation of the upper tail, e.g. E[V|V≥q0.999], which seems interesting but I’ve not thought carefully about.