Thanks for these thoughts and links Michael, and I’m glad you agree with the broad thrust of the post! You’ve given me a lot to think about and I’m finding my view on this is evolving.
I don’t think causes differ astronomically in the expected impact a reasonable EA should assign them after (let’s say) a thousand hours of learning and thinking about IBCs, using good resources
Thanks for this framing which is helpful. Reading through the comments and some of your links, I actually think that the specific claim I need to provide more of an argument for is this one:
There are astronomical differences in the expected value of different cause areas and people can uncover this through greater scrutiny of existing arguments and information.
I tentatively still hold this view, although I’m starting to think that it may not hold as broadly as I originally thought and that I got the cause area classification wrong. For example, it might not hold between near-term animal focused and near-term human focused areas. In other words, perhaps it just isn’t really possible, given the information that is currently available, to come to a somewhat confident conclusion that one of these areas is much better than the other in expected value terms. I have also realised that Maryam, in my hypothetical example, didn’t actually conclude that near-term animal areas were better than near-term human areas in terms of expected value. Instead she just concluded that near-term animal areas (which are mainstream EA) were better than aspecific near-term human area (mental health—which is EA but not mainstream). So I’m now starting to question whether the way I classified cause areas was helpful.
Having said all that I would like to return to longtermist areas. I actually do think that many prominent EAs, say Toby Ord, Will MacAskill and Hilary Greaves, would argue that longtermist areas are astronomically better than shorttermist areas in expected value terms. Greaves and MacAskill’s The Case for Strong Longtermism basically argues this. Ord’s The Precipice basically argues this, but specifically from an x-risk perspective. It might be that longtermism is the only case where prominent thinkers in the movement do think there is a clear argument to be made for astronomically different expected value.
Does this then mean it’s very important to educate people about ideas that help people prioritise between shorttermist areas and longtermist areas? Well I think if we adopt some epistemic humility and accept that it’s probably worth educating people about ideas where prominent EAs have claimed astronomical expected value without much disagreement from other prominent EAs, then the answer is yes. The idea here being that, the fact that these people hold these ideas without much prominent disagreement means there is a good possibility these ideas are correct and so, on average, people being aware of these ideas should make them reorient in ways that will allow them to do more good. This actually makes some sense to me, although I realise this argument is grounded in epistemic humility and deferring, which is quite different to my original argument. It’s not pure deferring of course as people can still come across the ideas and reject them, I’m just saying it’s important that they come across and understand the ideas in the first place.
So to sum up, I think my general idea might still work, but I need to rework my cause areas. A better classification might be: non-EA cause areas, shortermist EA cause areas, longtermist EA cause areas. These are the cause areas between which I still think there should be astronomical differences in expected value that can plausibly be uncovered by people. In light of this different cause classification I suspect some of my IBCs will drop—specifically those that help in deciding between interventions between which there is reasonable disagreement amongst prominent EAs (e.g. Ord and MacAskill disagree on how influential the present is, so that can probably be dropped).
Given this it might be that my list of IBCs reduce to:
Those that help people orient from non-EA to EA, so that they can make switches like Maryam did
Those that help people potentially orient from shorttermism to longtermism, so that they can make switches like Arjun did
I feel that I may have rambled a lot here and I don’t know if what I have said makes sense. I’d be interested to hear your thoughts on all of this.
I think you make a bunch of interesting points. I continue to agree with the general thrust of what you propose, though disagreeing on parts.
I actually do think that many prominent EAs, say Toby Ord, Will MacAskill and Hilary Greaves, would argue that longtermist areas are astronomically better than shorttermist areas in expected value terms. Greaves and MacAskill’s The Case for Strong Longtermism basically argues this. Ord’s The Precipice basically argues this, but specifically from an x-risk perspective. It might be that longtermism is the only case where prominent thinkers in the movement do think there is a clear argument to be made for astronomically different expected value.
I haven’t read that key paper from Greaves & MacAskill. I probably should. But some complexities that seem worth noting are that:
The longtermist interventions we think are best usually have a nontrivial chance of being net harmful
It seems plausible to me that, if “longtermism is correct”, then longtermist interventions that are actually net harmful will tend to be more harmful than neartermist interventions that are actually net harmful
This is basically because the “backfires” would be more connected to key domains, orgs, decisions, etc.
Which has various consequences, such as the chance of creating confusion on key questions, causing longtermism-motivated people to move towards career paths that are less good than those they would’ve gone towards, burning bridges (e.g., with non-Western governments), or creating reputational risks (seeming naive, annoying, etc.)
Even longtermist interventions that are typically very positive in expectation could be very negative in expectation if done in terrible ways by very ill-suited people
Neartermist interventions will also have some longtermist implications, and I’d guess that usually there’s a nontrivial chance that they have extremely good longtermist implications
E.g., the interventions probably have a nontrivial chance of meaningfully increasing or decreasing economic growth, technological progress, or moral circle expansion, which in turn is plausibly very good or very bad from a longtermist perspective
Related to the above point: In some cases, people might actually do very similar tasks whether they prioritise one cause area or another (specific) cause area
E.g., I think work towards moral circle expansion is plausibly a top priority from a longtermist perspective, and working on factory farming is plausibly a top priority way to advance that goal (though I think both claims are unlikely to be true). And I think that moral circle expansion and factory farming are also plausibly top priorities from a neartermist perspective.
Greaves, MacAskill, and Ord might be partly presenting a line of argument that they give substantial credence to, without constantly caveating it for epistemic humility in the way that they might if actually making a decision
A question I think is useful is “Let’s say we have a random person of the kind who might be inclined towards EA. Let’s say we could assign them to work on a randomly chosen intervention out of the set of interventions that such a person might think is a good idea from a longtermist perspective, or to work on randomly chosen intervention out of the set of interventions that such a person might think is a good idea from a neartermist animal welfare perspective. We have no other info about this person, which intervention they’d work on, or how they’d approach it. On your all-things-considered moral and empirical views—not just your independent impressions - it >1,000,000 times as good to assign this person to the randomly chosen longtermist intervention than the randomly chosen neartermist animal welfare intervention?”
I’m 95+% confident they’d say “No” (at least if I made salient to them the above points and ensured they understood what I meant by the question).
(I think expected impact differing by factors of 10 or 100s is a lot more plausible. And I think larger differences in expected impact are more plausible once we fill in more details about a specific situation, like a person’s personal fit and what specific intervention they’re considering. But learning about IBCs doesn’t inform us on those details.)
I’d be interested in your thoughts on this take (including whether you think I’m just sort-of talking past your point, or that I really really should just read Greaves and MacAskill’s paper!).
Greaves and MacAskill don’t cover concerns about potential downsides of longtermist interventions in their paper. I think they implicitly make a few assumptions, such as that someone pursuing the interventions they mention would actually do them thoughtfully and carefully. I do agree that one can probably go into say DeepMind without really knowing their stuff and end up doing astronomical harm.
Overall I think your general point is fair. When it comes to allocating a specific person to a cause area, the difference in expected value across cause areas probably isn’t as large as I originally thought, for example due to considerations such as personal fit. Generally I think your comments have updated me away from my original claim that everyone should know all IBCs, but I do still feel fairly positive about more content being produced to improve understanding of some of these ideas and I’m quite excited about this possibility.
Slight aside about the Greaves and MacAskill paper—I personally found it a very useful paper that helped me understand the longtermism claim in a slightly more formal way than say an 80K blog post. It’s quite an accessible paper. I also found the (somewhat limited) discussion about the potential robustness of longtermism to different views very interesting. I’m sure Greaves and MacAskill will be strengthening that argument in the future. So overall I would recommend giving it a read!
I do still feel fairly positive about more content being produced to improve understanding of some of these ideas and I’m quite excited about this possibility.
Yeah, I’m definitely on the same page on those points!
So overall I would recommend giving it a read!
Ok, this has made it more likely that I’ll make time for reading the paper in the coming weeks. Thanks :)
Thanks for these thoughts and links Michael, and I’m glad you agree with the broad thrust of the post! You’ve given me a lot to think about and I’m finding my view on this is evolving.
Thanks for this framing which is helpful. Reading through the comments and some of your links, I actually think that the specific claim I need to provide more of an argument for is this one:
I tentatively still hold this view, although I’m starting to think that it may not hold as broadly as I originally thought and that I got the cause area classification wrong. For example, it might not hold between near-term animal focused and near-term human focused areas. In other words, perhaps it just isn’t really possible, given the information that is currently available, to come to a somewhat confident conclusion that one of these areas is much better than the other in expected value terms. I have also realised that Maryam, in my hypothetical example, didn’t actually conclude that near-term animal areas were better than near-term human areas in terms of expected value. Instead she just concluded that near-term animal areas (which are mainstream EA) were better than a specific near-term human area (mental health—which is EA but not mainstream). So I’m now starting to question whether the way I classified cause areas was helpful.
Having said all that I would like to return to longtermist areas. I actually do think that many prominent EAs, say Toby Ord, Will MacAskill and Hilary Greaves, would argue that longtermist areas are astronomically better than shorttermist areas in expected value terms. Greaves and MacAskill’s The Case for Strong Longtermism basically argues this. Ord’s The Precipice basically argues this, but specifically from an x-risk perspective. It might be that longtermism is the only case where prominent thinkers in the movement do think there is a clear argument to be made for astronomically different expected value.
Does this then mean it’s very important to educate people about ideas that help people prioritise between shorttermist areas and longtermist areas? Well I think if we adopt some epistemic humility and accept that it’s probably worth educating people about ideas where prominent EAs have claimed astronomical expected value without much disagreement from other prominent EAs, then the answer is yes. The idea here being that, the fact that these people hold these ideas without much prominent disagreement means there is a good possibility these ideas are correct and so, on average, people being aware of these ideas should make them reorient in ways that will allow them to do more good. This actually makes some sense to me, although I realise this argument is grounded in epistemic humility and deferring, which is quite different to my original argument. It’s not pure deferring of course as people can still come across the ideas and reject them, I’m just saying it’s important that they come across and understand the ideas in the first place.
So to sum up, I think my general idea might still work, but I need to rework my cause areas. A better classification might be: non-EA cause areas, shortermist EA cause areas, longtermist EA cause areas. These are the cause areas between which I still think there should be astronomical differences in expected value that can plausibly be uncovered by people. In light of this different cause classification I suspect some of my IBCs will drop—specifically those that help in deciding between interventions between which there is reasonable disagreement amongst prominent EAs (e.g. Ord and MacAskill disagree on how influential the present is, so that can probably be dropped).
Given this it might be that my list of IBCs reduce to:
Those that help people orient from non-EA to EA, so that they can make switches like Maryam did
Those that help people potentially orient from shorttermism to longtermism, so that they can make switches like Arjun did
I feel that I may have rambled a lot here and I don’t know if what I have said makes sense. I’d be interested to hear your thoughts on all of this.
I think you make a bunch of interesting points. I continue to agree with the general thrust of what you propose, though disagreeing on parts.
I haven’t read that key paper from Greaves & MacAskill. I probably should. But some complexities that seem worth noting are that:
The longtermist interventions we think are best usually have a nontrivial chance of being net harmful
It seems plausible to me that, if “longtermism is correct”, then longtermist interventions that are actually net harmful will tend to be more harmful than neartermist interventions that are actually net harmful
This is basically because the “backfires” would be more connected to key domains, orgs, decisions, etc.
Which has various consequences, such as the chance of creating confusion on key questions, causing longtermism-motivated people to move towards career paths that are less good than those they would’ve gone towards, burning bridges (e.g., with non-Western governments), or creating reputational risks (seeming naive, annoying, etc.)
See also Cotton-Barratt’s statements about “Safeguarding against naive utilitarianism”
Even longtermist interventions that are typically very positive in expectation could be very negative in expectation if done in terrible ways by very ill-suited people
Neartermist interventions will also have some longtermist implications, and I’d guess that usually there’s a nontrivial chance that they have extremely good longtermist implications
E.g., the interventions probably have a nontrivial chance of meaningfully increasing or decreasing economic growth, technological progress, or moral circle expansion, which in turn is plausibly very good or very bad from a longtermist perspective
Related to the above point: In some cases, people might actually do very similar tasks whether they prioritise one cause area or another (specific) cause area
E.g., I think work towards moral circle expansion is plausibly a top priority from a longtermist perspective, and working on factory farming is plausibly a top priority way to advance that goal (though I think both claims are unlikely to be true). And I think that moral circle expansion and factory farming are also plausibly top priorities from a neartermist perspective.
Greaves, MacAskill, and Ord might be partly presenting a line of argument that they give substantial credence to, without constantly caveating it for epistemic humility in the way that they might if actually making a decision
See also the idea of sharing independent impressions
A question I think is useful is “Let’s say we have a random person of the kind who might be inclined towards EA. Let’s say we could assign them to work on a randomly chosen intervention out of the set of interventions that such a person might think is a good idea from a longtermist perspective, or to work on randomly chosen intervention out of the set of interventions that such a person might think is a good idea from a neartermist animal welfare perspective. We have no other info about this person, which intervention they’d work on, or how they’d approach it. On your all-things-considered moral and empirical views—not just your independent impressions - it >1,000,000 times as good to assign this person to the randomly chosen longtermist intervention than the randomly chosen neartermist animal welfare intervention?”
I’m 95+% confident they’d say “No” (at least if I made salient to them the above points and ensured they understood what I meant by the question).
(I think expected impact differing by factors of 10 or 100s is a lot more plausible. And I think larger differences in expected impact are more plausible once we fill in more details about a specific situation, like a person’s personal fit and what specific intervention they’re considering. But learning about IBCs doesn’t inform us on those details.)
I’d be interested in your thoughts on this take (including whether you think I’m just sort-of talking past your point, or that I really really should just read Greaves and MacAskill’s paper!).
Thanks for this.
Greaves and MacAskill don’t cover concerns about potential downsides of longtermist interventions in their paper. I think they implicitly make a few assumptions, such as that someone pursuing the interventions they mention would actually do them thoughtfully and carefully. I do agree that one can probably go into say DeepMind without really knowing their stuff and end up doing astronomical harm.
Overall I think your general point is fair. When it comes to allocating a specific person to a cause area, the difference in expected value across cause areas probably isn’t as large as I originally thought, for example due to considerations such as personal fit. Generally I think your comments have updated me away from my original claim that everyone should know all IBCs, but I do still feel fairly positive about more content being produced to improve understanding of some of these ideas and I’m quite excited about this possibility.
Slight aside about the Greaves and MacAskill paper—I personally found it a very useful paper that helped me understand the longtermism claim in a slightly more formal way than say an 80K blog post. It’s quite an accessible paper. I also found the (somewhat limited) discussion about the potential robustness of longtermism to different views very interesting. I’m sure Greaves and MacAskill will be strengthening that argument in the future. So overall I would recommend giving it a read!
Yeah, I’m definitely on the same page on those points!
Ok, this has made it more likely that I’ll make time for reading the paper in the coming weeks. Thanks :)