Background: I'm thinking through anthropics on my own to form my own views, here are some working notes
There's been disputes about how to define a reference class in anthropic reasoning. I think you can skip those disputes by letting the choice of reference class be a degree of freedom.
How would you do that?
Part of that involves an intuition that there are some observers that cannot be other observers. For example, observers who only see red cannot be observers who only see blue. Another example is observers who don't choose to get drunk cannot be observers who did choose to get drunk. By deciding what you're saying you know about the observers, you can group them up in what I'll call patches.
The patches are defined by filters which I call conditional gates. Those are conditions which, if satisfied, constitute a transformation from an observer of one patch to an observer of another patch. You can't be a random sample across different patches because different patches don't permit you to be any of them - you must be one or the other.
The choice of conditional gates used to carve out a patch of observers is entirely up to you, so long as you satisfy them! If you satisfy the constraints to be in the patch, then you should expect yourself to be a typical member of those in that patch.
The fact that you can choose your filter is interesting to me, because it means you can expect to be a typical member under some filter, and also a typical member of some other filter, and so on - and that reasoning about different filters can lead to different hypotheses about reality which can all be true.
Some ways of choosing a filter can be very uninformative, like if your filter only includes yourself - then you're a typical member of a patch of just yourself, which is trivial.
These patches also allow for dynamic references classes - you can move to another reference class by passing through a conditional gate, becoming a different kind of observer, and arriving on a new patch. There's choice involved.
In SSA you're only sampling from actual observers. If that's the case then SSA can only make local claims using Anthropics - claims about patches. Whenever counterfactuals occur, observers become some other kind of observer, and so a conditional gate needs to be drawn.
In the Doomsday Argument, you assume your birth rank should be a typical sample over all birth ranks for minds that'll be born. There are many counterfactual ways the future could go and you do not know them in advance - if you're allowed to sample yourself from observers as if there's one timeline - it must be because you've chosen conditional gates that every possible branching future can satisfy, under my frame. The more stringent the definition of observer, the more likely it cannot be satisfied by all possible counterfactual futures. This means the inclusion criteria for the observer patch must be very broad and likely uninformative about what we care about. As such, under my frame, the Doomsday Argument wouldn't actually be saying doom is soon in a way we might care about. Rather, it'd be saying "you won't last long trying to find something common across every single possible future the deeper into the future tree you go".
I'll note that observers within a patch do not need to be alive at the same time - the conditional gating can logically organize observers all sorts of ways. A peculiar property is that you can allow inanimate or abstract things to be 'observers' in your observer patch, so long as you also satisfy the constraints to belong in it. I actually find that fine - it doesn't privilege some particular notion of observer. It can treat anything as having a "vantage point" and thus being an observer.
There's plenty of open questions like:
What does it mean to be a typical observer given some patch defined by the filters chosen?
Are there any plausible examples where this frame seems to generate productive, nontrivial beliefs about reality that are harder with SIA or SSA?
How do you aggregate information about reality informed by anthropic reasoning under different choices of filters?
How do you pick filters in an 'informative' way?
Are there any trade offs you make when you add more or less filters when defining your patch?
If an observer's choices can count as a conditional gate, could you have a conditional gate based on the choice of filter made by an observer?
How does it compare to SSA and SIA?
...
I might address these in future posts. There's also a lot of tangents I wound up going on when thinking about this, so I'd like to address those as well.
Background: I'm thinking through anthropics on my own to form my own views, here are some working notes
There's been disputes about how to define a reference class in anthropic reasoning. I think you can skip those disputes by letting the choice of reference class be a degree of freedom.
How would you do that?
Part of that involves an intuition that there are some observers that cannot be other observers. For example, observers who only see red cannot be observers who only see blue. Another example is observers who don't choose to get drunk cannot be observers who did choose to get drunk. By deciding what you're saying you know about the observers, you can group them up in what I'll call patches.
The patches are defined by filters which I call conditional gates. Those are conditions which, if satisfied, constitute a transformation from an observer of one patch to an observer of another patch. You can't be a random sample across different patches because different patches don't permit you to be any of them - you must be one or the other.
The choice of conditional gates used to carve out a patch of observers is entirely up to you, so long as you satisfy them! If you satisfy the constraints to be in the patch, then you should expect yourself to be a typical member of those in that patch.
The fact that you can choose your filter is interesting to me, because it means you can expect to be a typical member under some filter, and also a typical member of some other filter, and so on - and that reasoning about different filters can lead to different hypotheses about reality which can all be true.
Some ways of choosing a filter can be very uninformative, like if your filter only includes yourself - then you're a typical member of a patch of just yourself, which is trivial.
These patches also allow for dynamic references classes - you can move to another reference class by passing through a conditional gate, becoming a different kind of observer, and arriving on a new patch. There's choice involved.
In SSA you're only sampling from actual observers. If that's the case then SSA can only make local claims using Anthropics - claims about patches. Whenever counterfactuals occur, observers become some other kind of observer, and so a conditional gate needs to be drawn.
In the Doomsday Argument, you assume your birth rank should be a typical sample over all birth ranks for minds that'll be born. There are many counterfactual ways the future could go and you do not know them in advance - if you're allowed to sample yourself from observers as if there's one timeline - it must be because you've chosen conditional gates that every possible branching future can satisfy, under my frame. The more stringent the definition of observer, the more likely it cannot be satisfied by all possible counterfactual futures. This means the inclusion criteria for the observer patch must be very broad and likely uninformative about what we care about. As such, under my frame, the Doomsday Argument wouldn't actually be saying doom is soon in a way we might care about. Rather, it'd be saying "you won't last long trying to find something common across every single possible future the deeper into the future tree you go".
I'll note that observers within a patch do not need to be alive at the same time - the conditional gating can logically organize observers all sorts of ways. A peculiar property is that you can allow inanimate or abstract things to be 'observers' in your observer patch, so long as you also satisfy the constraints to belong in it. I actually find that fine - it doesn't privilege some particular notion of observer. It can treat anything as having a "vantage point" and thus being an observer.
There's plenty of open questions like:
I might address these in future posts. There's also a lot of tangents I wound up going on when thinking about this, so I'd like to address those as well.