That said, I am strongly averse to EAs using ‘not neglected’ as evidence that some area isn’t worth supporting.
I strongly disagree that ‘not neglected’ isn’t good evidence. This is evidence depending only on pretty weak assumptions (returns which diminish faster than seeing more people working on a thing is evidence for the thing being good). For research-ish areas, I think you should probably have something like a log returns prior in which case this makes sense.
I’m somewhat sympathetic to “maybe someone should just do the serious prioritization research” in which case we don’t need rough except for prioritising prioritisation projects. ([Low confidence] But in practice, I think high level spot checking is necessary and often better than actual reports for longtermism IMO. For this heuristics are pretty important. It’s just pretty easy to have a lot of information/understanding which allows for beating carefully done prioritization research in the longermist space in many cases. Like research can be informative, but often not for the bottom line, but instead for answering various questions about fundamentals.)
For the definition of ITN that I use (but maybe people use others?), tractability is basically separable from neglectedness:
Recall the importance-tractability-neglectedness (ITN) framework for estimating cost-effectiveness:
Importance = utility gained / % of problem solved
Tractability = % of problem solved / % increase in resources
Neglectedness = % increase in resources / extra $
(Quote from the same post I linked above.)
it’s not even well defined (not neglected relative to overall amount of effort required, or in absolute terms)
In absolute terms ideally (as in definition above), though this might be somewhat poorly defined still.
At the end of the day, we just want to compute expected value with respect to various actions, but I still think that “are a bunch of people already trying to solve the problem and are they approaching it in a reasonable way” is a pretty good heuristic. (In research-ish fields, clearly some fields end up having pretty linear returns and we can usually predict this with some other simple heuristics.)
I don’t have time to respond to this in as much depth as I’d like, but maybe it’s worth a few cursory remarks, since much of what you’ve said touches on my frustrations with the concept.
returns which diminish faster than seeing more people working on a thing is evidence for the thing being good
I’m not sure how to parse this.
pretty weak assumptions… For research-ish areas, I think you should probably have something like a log returns prior in which case this makes sense.
This seems like an extremely strong assumption to me, especially since it’s basically the assertion I’m contesting:
‘research-ish’ is extremely underdefined and in practice not normally what we’re discussing. Eg the matter at hand: space colonisation is a confluence of a huge number of areas, including public opinion, funding, engineering, political will, etc. In fact, there’s very little academic research that needs to be done to make it happen. I think this is normally the case when EAs dismiss something as ‘not neglected’ (cf also climate change)
For technology-ish areas, public opinion-ish areas (ie almost any practical issue) I think you should have something much more like an S-curve prior, which requires you to make more assumptions than a log returns prior, and to do more empirical research to justify them. S-curves also have a much more complicated relationship with expected marginal value.
Some real world examples of problems and ambiguities with the concept that showcase these issues:
Climate change is often dismissed as ‘not neglected’, but at least some proportion of work that many metrics would consider as ‘climate change related’ is very ineffective or literally zero-sum (such as opposing vs advocating nuclear energy)
The EA movement began because Toby/Holden/Elie found a bunch of underutilised health-relevant research that was being ignored by a huge base of donors—if that work hadn’t been done, GWWC and Givewell would have been much less compelling sales pitches, and the movement would be able to boast of far fewer lives saved.
People often described AI safety research as ‘neglected’, but this is using what strikes me as a very tribal definition of AI safety—most tech workers spend their lives trying to ‘align’ computer programs with the goal they have in mind, but EAs tend to dismiss this irrelevant, implying that AI doom is highly likely if left to such people. I see no good reason for such a strong take, and if you even given some weight to the safety-relevance of such work, then AI safety becomes a far less neglected field, with ~25million people * whatever weighting you give them working on it.
A line from the recent EA Netherlands newsletter: ‘We have the core of a social movement in the Netherlands, but we don’t have much by way of a research field or a set of organisations putting said research into practice. This means that, while we have many people in the community who want to use their time and their talents to do good, we lack the organisations that can provide the work. There are opportunities abroad, but not enough, and the fact they are abroad limits their capacity to absorb talent located in the Netherlands.’ Another way of looking at this would be ‘EA field building and organisational development in the Netherlands is hugely neglected, and therefore and extremely high-leverage opportunity for doing such things’. Yet they seem to have (IMO sensibly) taken the opposite view.
It’s just pretty easy to have a lot of information/understanding which allows for beating carefully done prioritization research in the longermist space in many cases.
This seems like a very hard claim to justify, given both conceptually how difficult it is to measure outputs/value gains in the longtermist space and how few organisations in it are producing anything meaningfully measureable at all. I think many assumptions the space makes, going right back to the idea that longtermist work is better from a longtermist perspective than ‘short termism’ are justified on flimsy, self-supporting heuristics.
For the definition of ITN that I use (but maybe people use others?),
I don’t believe in practice (m?)any people use that version. In order to get a neatly cancellable equation, it quietly replaces ‘tractability’ - a relatively intuitive concept that people can often do useful work on—with something like elasticity of tractability, which I’ve never heard anyone opine on, directly or indirectly in any other context.
I strongly disagree that ‘not neglected’ isn’t good evidence. This is evidence depending only on pretty weak assumptions (returns which diminish faster than seeing more people working on a thing is evidence for the thing being good). For research-ish areas, I think you should probably have something like a log returns prior in which case this makes sense.
I’m somewhat sympathetic to “maybe someone should just do the serious prioritization research” in which case we don’t need rough except for prioritising prioritisation projects. ([Low confidence] But in practice, I think high level spot checking is necessary and often better than actual reports for longtermism IMO. For this heuristics are pretty important. It’s just pretty easy to have a lot of information/understanding which allows for beating carefully done prioritization research in the longermist space in many cases. Like research can be informative, but often not for the bottom line, but instead for answering various questions about fundamentals.)
See also Most problems fall within a 100x tractability range.
For the definition of ITN that I use (but maybe people use others?), tractability is basically separable from neglectedness:
(Quote from the same post I linked above.)
In absolute terms ideally (as in definition above), though this might be somewhat poorly defined still.
At the end of the day, we just want to compute expected value with respect to various actions, but I still think that “are a bunch of people already trying to solve the problem and are they approaching it in a reasonable way” is a pretty good heuristic. (In research-ish fields, clearly some fields end up having pretty linear returns and we can usually predict this with some other simple heuristics.)
I don’t have time to respond to this in as much depth as I’d like, but maybe it’s worth a few cursory remarks, since much of what you’ve said touches on my frustrations with the concept.
I’m not sure how to parse this.
This seems like an extremely strong assumption to me, especially since it’s basically the assertion I’m contesting:
‘research-ish’ is extremely underdefined and in practice not normally what we’re discussing. Eg the matter at hand: space colonisation is a confluence of a huge number of areas, including public opinion, funding, engineering, political will, etc. In fact, there’s very little academic research that needs to be done to make it happen. I think this is normally the case when EAs dismiss something as ‘not neglected’ (cf also climate change)
For technology-ish areas, public opinion-ish areas (ie almost any practical issue) I think you should have something much more like an S-curve prior, which requires you to make more assumptions than a log returns prior, and to do more empirical research to justify them. S-curves also have a much more complicated relationship with expected marginal value.
Some real world examples of problems and ambiguities with the concept that showcase these issues:
Climate change is often dismissed as ‘not neglected’, but at least some proportion of work that many metrics would consider as ‘climate change related’ is very ineffective or literally zero-sum (such as opposing vs advocating nuclear energy)
The EA movement began because Toby/Holden/Elie found a bunch of underutilised health-relevant research that was being ignored by a huge base of donors—if that work hadn’t been done, GWWC and Givewell would have been much less compelling sales pitches, and the movement would be able to boast of far fewer lives saved.
People often described AI safety research as ‘neglected’, but this is using what strikes me as a very tribal definition of AI safety—most tech workers spend their lives trying to ‘align’ computer programs with the goal they have in mind, but EAs tend to dismiss this irrelevant, implying that AI doom is highly likely if left to such people. I see no good reason for such a strong take, and if you even given some weight to the safety-relevance of such work, then AI safety becomes a far less neglected field, with ~25million people * whatever weighting you give them working on it.
A line from the recent EA Netherlands newsletter: ‘We have the core of a social movement in the Netherlands, but we don’t have much by way of a research field or a set of organisations putting said research into practice. This means that, while we have many people in the community who want to use their time and their talents to do good, we lack the organisations that can provide the work. There are opportunities abroad, but not enough, and the fact they are abroad limits their capacity to absorb talent located in the Netherlands.’ Another way of looking at this would be ‘EA field building and organisational development in the Netherlands is hugely neglected, and therefore and extremely high-leverage opportunity for doing such things’. Yet they seem to have (IMO sensibly) taken the opposite view.
This seems like a very hard claim to justify, given both conceptually how difficult it is to measure outputs/value gains in the longtermist space and how few organisations in it are producing anything meaningfully measureable at all. I think many assumptions the space makes, going right back to the idea that longtermist work is better from a longtermist perspective than ‘short termism’ are justified on flimsy, self-supporting heuristics.
I don’t believe in practice (m?)any people use that version. In order to get a neatly cancellable equation, it quietly replaces ‘tractability’ - a relatively intuitive concept that people can often do useful work on—with something like elasticity of tractability, which I’ve never heard anyone opine on, directly or indirectly in any other context.