The main fallacy here, one that destroys the entire premise, is that we cannot realistically assume the position of "beings at the end of time".
Everyone who has ever worked with data and modeling knows that extrapolations are inaccurate and unreliable; the further outside the domain of the actual data, the more unreliable.
Longtermism is basing your views (or worse, decisions) on decidedly unreliable assumptions of "what it is I think will be good in the hypothetical long term".
Theoretically this is flawed. Practically this is worse than flawed: very often it means pushing one's own dubious (or hideous) agenda by wrapping it in fancy paper. Had we heeded "longtermists" who predicted our ecological doom in, when? 1990? 2000? 2020? we would have been much worse off than we are now.
I agree that always thinking immediate gratification is flawed, too. We need a balance. But flawed as it is, immediate gratification is at least grounded in reality. Assuming you're at the end of time is not.
There's no substitute for making correct predictions: "if [people who were wrong] had followed [theory], they would have done something wrong" is not a strong argument against a theory.
It may be the case that, when humans consider actions, explicitly taking the long view leads to worse long-term consequences. I don't think you've made a strong argument for this either (is "not attempting to extrapolate at all" a better way to extrapolate? What is experience with data and modeling supposed to tell us about this question).
But none of these answer the question of, if we did know all the facts, what should we prioritize. At worst these would turn you into an esoteric longtermist, believing that longtermism is correct but convincing people to be longtermists is unhelpful. (This would be a substantial revision for BB at least, who would presumably stop writing articles on longtermism if this were his position.)
It is actually quite simple: do make predictions, but prioritize realistic, short-term predictions rather than inherently unreliable longer term predictions. It is not entirely without flaws but the error will be minimized this way.
As in optimization: the core of it is the steepest gradient descent. Granted, it needs to be refined to prevent getting stuck in a local minimum, but ditching it and walking uphill "where you believe you will find a global minimum" is a much worse approach.
As in traffic: if you are too intent on looking at all traffic lights half a mile ahead, you will likely end up running a red light light in front of you. Look at that one instead and let the further ones figure themselves out in their own turn.
As for the question "what would we prioritize had we known all the facts", it is simply irrelevant, along with the question about how many angels can dance on the head of a pin. You can even figure it out - but you won't be able to apply it to anything.
But. Having invested effort in figuring it out, you will have the urge to apply it wherever it remotely seems to fit, by force if necessary. And this is where the main danger of it lies. Much much woe has been brought upon the world by all versions of just this idea: let us inflict suffering for the sake of some greater good.
"it is simply irrelevant" -- this is a contentious and not at all obvious claim. It would seem like considering long-term future effects could plausibly have implications for what we do now. And indeed, most longtermists think it does make a difference. Global optimization and local optimization tend to find different solutions.
Which is why, even though it is not a situation we could ever find ourselves in, the "what would you recommend, looking back at history, knowing everything" question is relevant: it distinguishes between those who think effects on the far future matter and those who don't.
Your examples fail to engage with the question "if I actually wanted to make things go well, across all time and space, what would I do. Nobody will recommend looking at stoplights a mile away, because that doesn't advance the local *or* global goal.
Here's my attempt at an improved driving metaphor to capture the issue: Alice wants to drive to the West Coast (she's not picky about where). Bob wants to drive to the highest point on the continent.
There is a local rule Alice can follow that will probably work (at any intersection, take the road that points west), but she may do better by planning globally (go straight to the interstate, even if it means driving east to start). But there's no local rule Bob can follow to get where he wants to go -- it's a cursed optimization landscape for gradient ascent. In all cases, it matters what the goal is. In all cases, you should look at the traffic lights in front of you.
I get your points, and I kind of see where you are coming from.
What I can't agree with is the viewpoint of "even though we can never be in the situation of having the data to optimize globally, it nevertheless makes sense to consider what we would do if we did".
The danger here is that sooner or later we will trick ourselves into thinking that we have sufficient data to optimize globally, will act on this premise, and will mis-optimize, and will fail spectacularly, sometimes to tragic effects - history has more than a couple examples.
That's why while considering longer-term goals and effects is a totally legitimate strategy, care must always be taken to check ourselves that we actually act on reason and common sense and wisdom, not on ideals or abstract theories or some one-sided vision of the future. And whenever we find ourselves in position that for the sake of whatever long term benefits we are undergoing (or worse, inflicting) short-term detriment, triple check again. If Alice, aiming for the west coast, finds herself going south to the point of seeing cactuses and people in sombreros, she needs to check her map before continuing; and if unable to do so, it might just be the right thing to follow the compass instead.
I agree with all your points of caution: one should not naively pursue a narrow target, and should definitely not drive to Mexico if they're trying to get to Oregon. I don't see any of this as an argument against longtermism, because longtermists agree on these points.
I'm still not sure what exactly you're arguing for though.
Consider also Charlie: Charlie, when he comes to intersections, likes to take the road going west, but does not have any continent-wide ambitions. He is completely indifferent to questions of arriving at the West Coast, and thinks it's silly to spend any time at all thinking about the West Coast while one is in Pennsylvania. Looking at national-scale maps is ridiculous, as they don't show any of the local roads he's been driving on -- not action guiding in the slightest.
Some possible positions:
(1) Longtermism is correct and at least sometimes action-guiding. (Bob, or Alice taking the interstate)
(2) Longtermism is correct as a description of what should be preferred / valued -- future generations matter -- but it is never action-guiding. Making the correct short-term decisions is the best way to improve the long-term future. (Alice on back roads, supposing this is optimal)
(3) Longtermism is incorrect in what it values. All that matters is doing good things now. (Charlie)
The reason the "what would we do if we did" question matters is because it's asking "where do we want to go?" The central question at issue here is philosophical, not (directly) practical.
Longtermism is correct *and* action-guiding in the ideal case that the question of "where do we want to go" (A) has an answer at all, and (B) has an answer that we can access and determine with at least approximate correctness. That is, if we know for a fact, or at least have reasonable grounds to assume, that we're on a continent where there's a West Coast, a road network, and a map thereof.
This assumption is a long shot. We cannot ascertain (A), let alone (B). We are, in the continent metaphor, on a continent where no one has ever been on the West Coast, the rumors about it are conflicting, roads are speculative, and maps are scarce and contradictory.
And if that is so, then (3) is always a possibility. Any long-term goal may end up being simply wrong: based on inaccurate extrapolation. This does not mean that it can never be action-guiding. It means we need to take those actions with a grain of salt, knowing they may be wrong, and being ready to retract. And refrain from driving off a cliff even if our GPS appears to tell us to.
The thought experiment of "what if we look back from the end of all time" does not help. First and foremost because it has already been attempted: we already tried to look at life from the perspective of "the only thing that matters is getting to eternity of heaven and everything else is subordinate": the ultimate form of longtermism. We know how trying to make it action-guiding ended.
In response to the other comments citing lack of consensus / tractability / difficulty of predicting how trends will continue - I feel like these are easily incorporated into a Strong Longtermist view by understanding the same is true for many presentist concerns and perspectives.
Isolating to a hypothetical where we are deciding who should be on a colony ship - all else equal it’s much better to have people who do not wish to propagate factory farming on their new colony world, thereby predictably influencing the future. Tractability and lack of consensus concerns can be folded into probabilities before deciding what action to take, much the same way they are in any other rationalist framework.
The thing I keep tripping on with strong Longtermism isn't the aggregation math and isn't the tractability question. It's underneath both.
Every longtermist actor holds a different picture of what the far future should look like, and what each considers important, what each hopes for, what each intervenes toward is not unified across the community. When multiple long-term interventions are running concurrently, which is already the situation we're in, they don't add up cleanly. They interfere. The further out you look, the more the aggregate looks like noise rather than signal.
So the essay's imperative, focus on making the far future go better, describes what a single actor plans. It doesn't describe what actually reaches the year 1 trillion AD. What reaches 1 trillion AD is whatever emerges from thousands of divergent plans grinding against each other, which is not the same object as any of the plans.
I've been circling this for months in my own writing, around how institutional systems process heterogeneous inputs into outcomes nobody chose. It shows up in psychiatric diagnosis, in ESG rating, in attention allocation. The Longtermist thesis has the same structural problem those systems have, but in better clothes.
The main fallacy here, one that destroys the entire premise, is that we cannot realistically assume the position of "beings at the end of time".
Everyone who has ever worked with data and modeling knows that extrapolations are inaccurate and unreliable; the further outside the domain of the actual data, the more unreliable.
Longtermism is basing your views (or worse, decisions) on decidedly unreliable assumptions of "what it is I think will be good in the hypothetical long term".
Theoretically this is flawed. Practically this is worse than flawed: very often it means pushing one's own dubious (or hideous) agenda by wrapping it in fancy paper. Had we heeded "longtermists" who predicted our ecological doom in, when? 1990? 2000? 2020? we would have been much worse off than we are now.
I agree that always thinking immediate gratification is flawed, too. We need a balance. But flawed as it is, immediate gratification is at least grounded in reality. Assuming you're at the end of time is not.
There's no substitute for making correct predictions: "if [people who were wrong] had followed [theory], they would have done something wrong" is not a strong argument against a theory.
It may be the case that, when humans consider actions, explicitly taking the long view leads to worse long-term consequences. I don't think you've made a strong argument for this either (is "not attempting to extrapolate at all" a better way to extrapolate? What is experience with data and modeling supposed to tell us about this question).
But none of these answer the question of, if we did know all the facts, what should we prioritize. At worst these would turn you into an esoteric longtermist, believing that longtermism is correct but convincing people to be longtermists is unhelpful. (This would be a substantial revision for BB at least, who would presumably stop writing articles on longtermism if this were his position.)
It is actually quite simple: do make predictions, but prioritize realistic, short-term predictions rather than inherently unreliable longer term predictions. It is not entirely without flaws but the error will be minimized this way.
As in optimization: the core of it is the steepest gradient descent. Granted, it needs to be refined to prevent getting stuck in a local minimum, but ditching it and walking uphill "where you believe you will find a global minimum" is a much worse approach.
As in traffic: if you are too intent on looking at all traffic lights half a mile ahead, you will likely end up running a red light light in front of you. Look at that one instead and let the further ones figure themselves out in their own turn.
As for the question "what would we prioritize had we known all the facts", it is simply irrelevant, along with the question about how many angels can dance on the head of a pin. You can even figure it out - but you won't be able to apply it to anything.
But. Having invested effort in figuring it out, you will have the urge to apply it wherever it remotely seems to fit, by force if necessary. And this is where the main danger of it lies. Much much woe has been brought upon the world by all versions of just this idea: let us inflict suffering for the sake of some greater good.
"it is simply irrelevant" -- this is a contentious and not at all obvious claim. It would seem like considering long-term future effects could plausibly have implications for what we do now. And indeed, most longtermists think it does make a difference. Global optimization and local optimization tend to find different solutions.
Which is why, even though it is not a situation we could ever find ourselves in, the "what would you recommend, looking back at history, knowing everything" question is relevant: it distinguishes between those who think effects on the far future matter and those who don't.
Your examples fail to engage with the question "if I actually wanted to make things go well, across all time and space, what would I do. Nobody will recommend looking at stoplights a mile away, because that doesn't advance the local *or* global goal.
Here's my attempt at an improved driving metaphor to capture the issue: Alice wants to drive to the West Coast (she's not picky about where). Bob wants to drive to the highest point on the continent.
There is a local rule Alice can follow that will probably work (at any intersection, take the road that points west), but she may do better by planning globally (go straight to the interstate, even if it means driving east to start). But there's no local rule Bob can follow to get where he wants to go -- it's a cursed optimization landscape for gradient ascent. In all cases, it matters what the goal is. In all cases, you should look at the traffic lights in front of you.
I get your points, and I kind of see where you are coming from.
What I can't agree with is the viewpoint of "even though we can never be in the situation of having the data to optimize globally, it nevertheless makes sense to consider what we would do if we did".
The danger here is that sooner or later we will trick ourselves into thinking that we have sufficient data to optimize globally, will act on this premise, and will mis-optimize, and will fail spectacularly, sometimes to tragic effects - history has more than a couple examples.
That's why while considering longer-term goals and effects is a totally legitimate strategy, care must always be taken to check ourselves that we actually act on reason and common sense and wisdom, not on ideals or abstract theories or some one-sided vision of the future. And whenever we find ourselves in position that for the sake of whatever long term benefits we are undergoing (or worse, inflicting) short-term detriment, triple check again. If Alice, aiming for the west coast, finds herself going south to the point of seeing cactuses and people in sombreros, she needs to check her map before continuing; and if unable to do so, it might just be the right thing to follow the compass instead.
I agree with all your points of caution: one should not naively pursue a narrow target, and should definitely not drive to Mexico if they're trying to get to Oregon. I don't see any of this as an argument against longtermism, because longtermists agree on these points.
I'm still not sure what exactly you're arguing for though.
Consider also Charlie: Charlie, when he comes to intersections, likes to take the road going west, but does not have any continent-wide ambitions. He is completely indifferent to questions of arriving at the West Coast, and thinks it's silly to spend any time at all thinking about the West Coast while one is in Pennsylvania. Looking at national-scale maps is ridiculous, as they don't show any of the local roads he's been driving on -- not action guiding in the slightest.
Some possible positions:
(1) Longtermism is correct and at least sometimes action-guiding. (Bob, or Alice taking the interstate)
(2) Longtermism is correct as a description of what should be preferred / valued -- future generations matter -- but it is never action-guiding. Making the correct short-term decisions is the best way to improve the long-term future. (Alice on back roads, supposing this is optimal)
(3) Longtermism is incorrect in what it values. All that matters is doing good things now. (Charlie)
The reason the "what would we do if we did" question matters is because it's asking "where do we want to go?" The central question at issue here is philosophical, not (directly) practical.
It's tricky. It's a mix of all three.
Longtermism is correct *and* action-guiding in the ideal case that the question of "where do we want to go" (A) has an answer at all, and (B) has an answer that we can access and determine with at least approximate correctness. That is, if we know for a fact, or at least have reasonable grounds to assume, that we're on a continent where there's a West Coast, a road network, and a map thereof.
This assumption is a long shot. We cannot ascertain (A), let alone (B). We are, in the continent metaphor, on a continent where no one has ever been on the West Coast, the rumors about it are conflicting, roads are speculative, and maps are scarce and contradictory.
And if that is so, then (3) is always a possibility. Any long-term goal may end up being simply wrong: based on inaccurate extrapolation. This does not mean that it can never be action-guiding. It means we need to take those actions with a grain of salt, knowing they may be wrong, and being ready to retract. And refrain from driving off a cliff even if our GPS appears to tell us to.
The thought experiment of "what if we look back from the end of all time" does not help. First and foremost because it has already been attempted: we already tried to look at life from the perspective of "the only thing that matters is getting to eternity of heaven and everything else is subordinate": the ultimate form of longtermism. We know how trying to make it action-guiding ended.
In response to the other comments citing lack of consensus / tractability / difficulty of predicting how trends will continue - I feel like these are easily incorporated into a Strong Longtermist view by understanding the same is true for many presentist concerns and perspectives.
Isolating to a hypothetical where we are deciding who should be on a colony ship - all else equal it’s much better to have people who do not wish to propagate factory farming on their new colony world, thereby predictably influencing the future. Tractability and lack of consensus concerns can be folded into probabilities before deciding what action to take, much the same way they are in any other rationalist framework.
The thing I keep tripping on with strong Longtermism isn't the aggregation math and isn't the tractability question. It's underneath both.
Every longtermist actor holds a different picture of what the far future should look like, and what each considers important, what each hopes for, what each intervenes toward is not unified across the community. When multiple long-term interventions are running concurrently, which is already the situation we're in, they don't add up cleanly. They interfere. The further out you look, the more the aggregate looks like noise rather than signal.
So the essay's imperative, focus on making the far future go better, describes what a single actor plans. It doesn't describe what actually reaches the year 1 trillion AD. What reaches 1 trillion AD is whatever emerges from thousands of divergent plans grinding against each other, which is not the same object as any of the plans.
I've been circling this for months in my own writing, around how institutional systems process heterogeneous inputs into outcomes nobody chose. It shows up in psychiatric diagnosis, in ESG rating, in attention allocation. The Longtermist thesis has the same structural problem those systems have, but in better clothes.