This post outlines an idea for a research project. I think someone should do something like this, and there’s a ~50% chance I will at some point. I’m publicly sharing this idea in order to:
- Highlight this potentially important question/crux
- Get feedback
- See if there’s anyone else who might want to do a project along these lines[1]
- Make coordination, and avoiding duplication of work, easier[2]
Feel very free to comment here, message me, and/or schedule a call with me.
Some experts and EAs appear to think there’s a nontrivial chance of civilizational collapse. Some further argue that reducing the risk of collapse, or increasing the chance of recovery, should be a top priority. Others argue that that wouldn’t be worth prioritising even if there’s a nontrivial chance of collapse, based on the premise that the chance of recovery is already high. The latter group might suggest instead prioritising reducing risks of extinction or dystopia, or prioritising things unrelated to existential risks. Indeed, views on the odds of recovery from civilizational collapse seem to be pushing large (and growing) pools of money and talent either towards or away from work aimed at reducing collapse risks or increasing chances of recovery.
Advancing our thinking on that matter thus appears highly valuable. Furthermore, I see a way to do that that seems tractable and neglected. Specifically, I propose modelling the likelihood of various types of recovery from various types of collapse scenarios, following roughly the following steps:
- Think about how to carve up the possible causes of collapse (e.g. impact winter, pandemics), types of collapse (e.g. loss of population, industry, both), and types of recovery (e.g. recovery of GWP, political institutions, values), and think about what to focus on.
- “Likelihood of recovery from collapse” is underspecified, and some types of recovery from some scenarios may be unlikely even if others are likely.
- Begin to construct in Guesstimate one or more models of the odds of various types of recovery, given various collapse scenarios.
- Begin to estimate the parameters of this model(s).
- Iteratively refine the model(s), estimates, and conceptual underpinnings.
- Present the results in EA Forum posts, and possibly as a paper.
- Perhaps iterate further based on new feedback from the EA and academic communities.
Steps 1-4 could involve armchair reasoning; reading relevant academic and/or EA writings; and getting ideas and input from EAs and academics. I’ve previously collected sources relevant to this matter, and, if I were doing this project, I’d likely start by (re-)reading some of those sources and contacting some of their authors. I’d also read works from and talk to experts outside the global catastrophic risk field (e.g., experts on minimum viable populations or the history of the industrial revolution), especially when estimating parameters. I expect the output would resemble that of Rethink Priorities’ series on nuclear risks, and it might be useful to in some ways emulate how that research was conducted.
This project would be unlikely to provide definitive answers, but could plausibly inform our point estimates, narrow our uncertainty, indicate what further research would be most valuable, and suggest points for intervention.
It may also end up seeming worthwhile to (also) model the odds of various collapse scenarios occurring in the first place. However, that topic seems to me somewhat less neglected than the odds of recovery.
(See here for further thoughts.)
Grateful acknowledgements and delicious disclaimers
My thanks to David Denkenberger, Aron Mill, Luke Kemp, Siebe Rozendal, Ozzie Gooen, Seth Baum, and Luisa Rodriguez for their useful input on this research project idea. Also, this idea might have originally been inspired by Rethink Priorities indicating interest in “Analyzing the likelihood of civilization recovering from a population collapse” (I can’t remember for sure). This does not necessarily imply any of these people’s endorsement of this idea or this post.
This post doesn’t necessarily reflect the views of my past or present employers.
I have a other project ideas I could pursue instead of this. And someone else might have a more fitting skillset for this than I do. So I’m potentially happy for someone else to take this project on instead of me. ↩︎
Maybe someone else is already doing, or would by default do, something similar to this. And/or maybe someone is doing something vaguely related, and they or I could benefit from us talking. ↩︎
I'm a bit skeptical about the value of formal modelling here. The parameter estimates would be almost entirely determined by your assumptions, and I'd expect the confidence intervals to be massive.
I think a toy model would be helpful for framing the issue, but going beyond that (to structural estimation) seems not worth it.
Also, if you're aware of Rethink Priorities/Luisa Rodriguez's work on modelling the odds and impacts of nuclear war (e.g., here), I'd be interested to hear whether you think making parameter estimates was worthwhile in that case. (And perhaps, if so, whether you think you'd have predicted that beforehand, vs being surprised that there ended up being a useful product.)
This is because that seems like the most similar existing piece of work I'm aware of (in methodology rather than topic). And to me it seems like that project was probably worthwhile, including the parameter estimates, and that it provided outputs that are perhaps more useful and less massively uncertain than I would've predicted. And that seems like weak evidence that parameter estimates could be worthwhile in this case as well.
Thanks for the comment. That seems reasonable. I myself had been wondering if estimating the parameters of the model(s) (the third step) might be:
If that's the case, this would also reduce the extent to which this model could "plausibly inform our point estimates" and "narrow our uncertainty". Though the model might still capture the other two benefits (indicating what further research would be most valuable and suggesting points for intervention).
That said, if one goes to the effort of building a model of this, it seems to me like it's likely at least worth doing something like:
And then perhaps this project could stop there, or perhaps it could then involve somewhat deeper/more rigorous investigation of the parameters where that seems most valuable.
Any thoughts on whether that seems worthwhile?
[1] Perhaps this step could benefit from use of Elicit; I should think about that if I pursue this idea further.
I hadn't seen this post before, but this is basically a summary of the project I've just started doing with a LTFF grant.
Can we see it yet?
It ended up being more of a coding project that would enable us to ask the question than a research project looking into the answer. The post I wrote on the approach I'm using is here. Still working on giving the code a UI, but you can see it at https://github.com/Arepo/lrisk_calculator - and in theory at least, the two calculators will work if you insert a breakpoint, and query the Markov Chain object they generate.
I think a moderate amount of this work has actually already been done - I'd be happy to help connect you.
Thanks, I've sent you a PM :)
ETA: Turns out I was aware of the work Peter had in mind; I think it's relevant, but not so similar as to strongly reduce the marginal value this project could provide.
ETA (even later): Peter had in mind Luisa Rodriguez's work, some of which was published as What is the likelihood that civilizational collapse would directly lead to human extinction (within decades)? and some of which made its way into Will MacAskill's What We Owe the Future.
Did you ever start/do this project, as per your linked G-doc?
No, I didn't - I ended up getting hired by Rethink Priorities and doing work on nuclear risk instead, among other things.