This sounds fantastic, and I wish I could attend. (Weekend conferences / meetings don’t work well when you have kids at home on the weekends.) In lieu of being able to participate, I’ll briefly note a couple things I’d want to point out about well-being and metrics of progress.
First, for well-being, I think it is clear that “well being” isn’t uniform across people, and there are deep problems with any sort of preference aggregation. The issues here are complex, and I am sure they will be discussed—but there are a pair of conclusions which might be inferred, but such inference is unjustified.
1) Not being able to measure something does not imply it’s not real. Measurement is always reductive, and that’s fine—approximate progress is better than verifiable and concrete lack thereof.
2) The inability to agree on universally applicable wellbeing measures, measurement methods, and ways to aggregate disparate measures does not imply that progress itself is difficult, just that we’re not going to be able to fully justify any concept of well-being, and therefore for progress.
Second, the limitations of metrics in general apply; they are reductive, cannot capture our “actual” not-fully-understood goals, and will lose fidelity to those goals if they are pursued. (cf. Campbell’s law/ Goodhart’s law.) But a key point about metric maximization is that even in the presence of (non-superintelligent) misalignment, maximizing the metric will usually greatly improve the actual goal, albeit not as much as if the metric were better designed.
The question about whether utility can be aggregated across populations is an interesting and important one, and one that we’ll be discussing at this event.
This sounds fantastic, and I wish I could attend. (Weekend conferences / meetings don’t work well when you have kids at home on the weekends.) In lieu of being able to participate, I’ll briefly note a couple things I’d want to point out about well-being and metrics of progress.
First, for well-being, I think it is clear that “well being” isn’t uniform across people, and there are deep problems with any sort of preference aggregation. The issues here are complex, and I am sure they will be discussed—but there are a pair of conclusions which might be inferred, but such inference is unjustified.
1) Not being able to measure something does not imply it’s not real. Measurement is always reductive, and that’s fine—approximate progress is better than verifiable and concrete lack thereof.
2) The inability to agree on universally applicable wellbeing measures, measurement methods, and ways to aggregate disparate measures does not imply that progress itself is difficult, just that we’re not going to be able to fully justify any concept of well-being, and therefore for progress.
Second, the limitations of metrics in general apply; they are reductive, cannot capture our “actual” not-fully-understood goals, and will lose fidelity to those goals if they are pursued. (cf. Campbell’s law/ Goodhart’s law.) But a key point about metric maximization is that even in the presence of (non-superintelligent) misalignment, maximizing the metric will usually greatly improve the actual goal, albeit not as much as if the metric were better designed.
Good points. I agree about metrics.
The question about whether utility can be aggregated across populations is an interesting and important one, and one that we’ll be discussing at this event.