Speedrun: Demonstrate the ability to rapidly scale food production in the case of nuclear winter

By Buhl @ 2023-02-13T19:00 (+39)

Introduction

This post is a shallow investigation of the intervention of running pilots of various resilient food strategies, in order to demonstrate the viability of using these strategies to rapidly scale food production during a nuclear winter. The post is part of a sequence of speedrun research projects by Rethink Priorities’ General Longtermism team. I recommend you begin by reading the introductory post in the sequence if you haven’t already for information about the context of this post.

A quick context tl;dr:

1. The aim of this investigation was to help our team decide if we should take steps towards incubating an organization focusing on this project. Keep in mind that some of the conclusions take into account considerations (such as our team’s comparative advantage) that may not be relevant to the reader.
2. The investigation was intended to be very early-stage and prioritize speed over rigor. We would have conducted a more in-depth investigation before launching any projects in this space (and recommend that others do the same).
3. This post was written in early fall 2022; the funding situation has changed significantly since then, but the investigation has not been updated to reflect this. As such, the funding bar alluded to in the post is probably outdated. I feel uncertain about how that should affect my bottom line, but overall I’m probably a little less enthusiastic about this intervention as a result.

Epistemic status

I spent ~10 hours researching and writing up this speedrun, then got input from ALLFED staff and spent a further 5-10 hours updating it. This was my first attempt at civilisational collapse-related research and I don’t have a natural science background. So: I’m a junior generalist who has thought about this for a couple of work days, and as a result, this post should be considered very preliminary and likely to contain mistakes and bad takes. My goal in publishing this regardless is that it may be useful as (a) a primer gathering useful information in one place, and (b) an example of the kind of research done by junior generalists. 

Summary

• This speedrun evaluates the project of running pilots to test rapidly scaling up food production using resilient food solutions such as the ones identified by the Alliance to Feed the Earth in Disaster (ALLFED).
• I think this project is probably worth someone in the longtermist space doing, especially someone with expertise in resilient foods, food production, entrepreneurship, and/or NGOs working with governments.
  • This is somewhat based on my (extremely rough) estimate that this project reduces x-risk with a cost-effectiveness of ~$260 million per 0.01% absolute reduction^[1] (~70% confidence interval: 2.2 million to 2.7 billion). If this estimate were accurate, then this project would clear our median roughly estimated cost-effectiveness bar of $500M per basis-point of x-risk averted (though it wouldn’t clear the more stringent bar of $100M per basis-point averted). However, I only place limited weight on this model as I think it’s likely very inaccurate. (more)
    • Retroactive comment: As noted in the introduction, this cost-effectiveness bar is probably now out of date given recent changes to the EA funding landscape. I haven’t thought carefully about how this updates my bottom line.^[2] 
  • This is to a greater extent based on me thinking that the project has several plausible paths to impact (more) and low downside risk (more).
• I think this project is relatively unlikely (~10%?) to be in the top 10 of the project ideas on our current list excluding considerations about Rethink’s fit for supporting a project like this.
  • This is somewhat based on this project looking ~10x less cost-effective than the previous project I evaluated (using a similarly rough cost-effectiveness model), namely developing an affordable super PPE, despite the estimate in this speedrun being potentially more optimistic (e.g., the PPE estimate involved a very low lower bound on biohazard-related x-risk).
    • However, it’s worth keeping in mind that the error bars on both estimates are very wide (the 70% confidence interval spans ~3 orders of magnitude in both cases).
  • It is also somewhat based on me feeling less confident in the degree to which food insecurity contributes to existential risk relative to factors addressed by other interventions (such as pandemics) – though this is one of my main uncertainties. (more)
• Given the amount of existing work by better-suited actors, if I were to make a decision now about whether Rethink Priorities should try to incubate a new project in this space, I would say that it should not.
  • This is partly based on the previous point, and partly based on organizations with more relevant expertise already doing a moderate amount of work in this area. (more)

What is this project?

A large-scale nuclear war could cause nuclear winter, a phenomenon where sunlight is blocked by smoke, leading to significant temperature reductions. This would make current agricultural methods much less effective and so probably significantly reduce food supply, causing famine.

Some existing research organizations (e.g., ALLFED) have done work to identify resilient food solutions, i.e., food production methods that would help maintain global food supplies even in global catastrophic risk (GCR) scenarios such as a nuclear winter. Resilient food solutions can involve alternative food sources (such as seaweed and leaf-based protein) or changes to how “mainstream” food sources are manufactured (such as crop relocation and expanding the proportion of land used for food production^[3]. While existing organizations have identified several potential resilient food solutions, there have been few real-world demonstrations of their viability and efficacy.

In this speedrun, I evaluate the project of running large-scale pilots to test rapidly scaling up one or more resilient food solutions. This could involve running experiments to see how rapidly and how much production could be scaled up. It could also involve drawing up plans for how to ramp up production at scale in a global disaster, building infrastructure needed to quickly ramp up production, and spreading information about all of the above.

Lessons from these pilots could be used to:

1. Improve our understanding of resilient foods and develop more robust disaster preparedness plans,
2. If the resilient food solutions identified so far are indeed effective, provide robust evidence of this effectiveness in order to aid with prioritization (e.g., funding decisions by philanthropists) and be used in advocacy (e.g., to governments).

Which resilient food to pilot?

Below is a quick survey of the main resilient food solutions listed on ALLFED’s website^[4]. I wrote this to get a sense of (a) what resilient food solutions look like, and (b) which might be good fits for pilots/demonstrations. Note that this is a highly technical topic and, given my lack of technical understanding, I would place especially little weight on my conclusions in this section.

Paths to impact

Note that in this section, I focus exclusively on how this project could help prevent existential risk. I don’t consider other types of impact, even important ones such as lives saved in the short term in a nuclear winter. This is because my team generally operates under a longtermist worldview, and under this worldview, effects on existential risk seem likely to be the most important effects in determining the expected impact of a project.

Here’s a very rough sketch of what I think is the most important path to x-risk reduction for resilient food pilots^[5]:

1. Prevent civilisational collapse in a nuclear winter:
   1. Resilient food pilots have at least one of these two main effects:
      1. They demonstrate convincingly the viability of using resilient food solutions to rapidly scale up food production in a nuclear winter; this makes governments more likely to adopt resilient food plans.
      2. They help identify key bottlenecks and problems with resilient food solutions that then go on to be addressed by further work; this significantly improves the resilient food plans of governments (should they make any such plans) and possibly other actors like philanthropists.
   2. Nuclear winter causes large agricultural shortfalls (~10-90%) and mass famine.
   3. In part due to the resilient food pilots, governments either have high-quality resilient food plans in place or are willing to quickly make such plans briefly after a disaster in collaboration with researchers.
   4. These plans are (at least partially) implemented, and this makes a significant difference to the speed at and degree to which global famine is alleviated.
   5. The reduced length and severity of a famine causes reduced levels of conflict (since resource scarcity is a source of conflict). This increases cooperation in recovering from the nuclear winter and reduces the risk of compounding global catastrophes (e.g., more war), thereby reducing the chance of collapse.

Here are some other plausible paths to x-risk reduction (again very roughly sketched), roughly in descending order of how important I think they are:

1. Researchers aiming to reduce x-risk gain expertise, network, and credibility, especially among adjacent academic communities and governments, which enables further x-risk reducing work in the future. This expertise/network/credibility could develop partly via working with experts in adjacent fields and governments, and partly via the project constituting a legible win for x-risk motivated researchers.
2. Prevent totalitarianism following a nuclear winter: As pathway (1), but instead of preventing collapse, the reduced length and severity of famine decreases the risk of totalitarianism (food scarcity is very plausibly a risk factor for totalitarianism).
3. Prevent civilisational collapse in the case of a widespread crop disease. As pathway (1), but instead of a nuclear winter, the large agricultural shortfalls are caused by a widespread crop disease (either natural, accidentally released, or deliberately engineered).
   1. I’m very uncertain how much weight to place on this path as I have very little understanding of how likely a widespread crop disease is.
4. Prevent civilisational collapse in other GCR scenarios with severe climate effects. (e.g., heating and extreme weather events due to climate change, supervolcanic eruption, asteroid): As pathway (1), but instead of a nuclear winter, the large agricultural shortfalls are caused by other GCR events with severe climate effects, including:
   1. GCR events that lead to cooling, e.g., a supervolcanic eruption or asteroid impact. These scenarios are highly analogous to the nuclear winter scenario, but also 1 to 3 orders of magnitude less likely to cause an existential catastrophe, according to Toby Ord’s estimates in The Precipice (p. 287).
      1. I place limited weight on this path because I think the likelihood of these scenarios leading to an existential catastrophe is very low (primarily based on deferral to Toby Ord).
   2. GCR events that lead to heating and extreme weather events, e.g., severe climate change^[6]. These scenarios are less analogous to the nuclear winter scenario; but some resilient food solutions are likely about equally helpful in heating scenarios (e.g., scaling up seaweed production  (though I haven’t checked the growing requirements of seaweed in detail) and crop relocation (though the concrete relocation plans would be different in heating and cooling scenarios). Climate change is also about equally likely to cause an existential catastrophe as nuclear weapons, according to Toby Ord’s estimates in The Precipice (p. 287).
      1. I place limited weight on this path because I think extreme climate change would happen slowly enough that governments would be likely to adopt resilient food plans regardless of any pilots run now before climate change would cause severe global agricultural shortfall (e.g., >10%).^[7]
5. Aid civilisational recovery (in scenarios with severe climate effects):
   1. Some catastrophe, most likely nuclear winter, results in civilisational collapse and severe climate effects, most likely cooling.
   2. In part due to resilient food pilots, researchers and philanthropists have documented food production methods that (a) work relatively well even in extreme climate change scenarios, and (b) are relatively low-tech and low-coordination. Those food production methods could involve “alternative foods” like leaf protein concentrate or just things like directories of which crops can grow in which climate conditions.
   3. Some post-collapse survivors know about or find this documentation and use it to grow food production and thereby help rebuild civilization.
      1. I place limited weight on this path because (a) it is unclear to me that information from a pilot would make much marginal difference compared to already existing information about how to grow food in a scenario like this, and (b) it seems to me like many versions of a resilient food pilot would not be geared towards this scenario.
6. Prevent civilisational collapse in other GCR scenarios that don’t involve severe climate effects (e.g., pandemics, cyberattacks)^[8]:
   1. As pathway (1), but instead of a nuclear winter, the large agricultural shortfalls are caused by other GCR events, such as:
      1. A highly deadly pandemic which significantly disrupts food production, e.g., by killing a high proportion of agricultural workers or making agricultural workers too scared to go to work.
      2. Cyberattacks on critical food infrastructure.
      3. High-altitude electromagnetic pulses resulting from nuclear war (which could impact food production even absent a nuclear winter).
   2. Some resilient food solutions seem like they could plausibly be cost-effective solutions to food shortages in such scenarios. For example, resilient food interventions that focus on optimizing or scaling up existing (especially low-tech) production (e.g., greenhouse deployment, crop relocation).
   3. I place limited weight on this path because (a) at least some of the example scenarios (other than pandemics) seem to me highly unlikely to result in an existential catastrophe, and (b) it seems to me like many versions of a resilient food pilot would not be geared towards this scenario.

Cost-effectiveness BOTEC: ~$260million / 0.01% x-risk reduction

I made a quick, extremely rough cost-effectiveness Guesstimate model. But first, four things to note on the limitations of this model:

1. It’s based only on the path-to-impact flowing via nuclear winter, since other paths to impact listed in the previous section seemed either (a) significantly less important to me or (b) difficult to quantify (specifically for x-risk-motivated researchers developing expertise and credibility). Since the most significant-seeming other effects I can think of are positive, my estimate is likely a (small?) underestimate.
2. The uncertainty in this estimate is extremely high. So the main implication of this estimate is that I’m extremely uncertain about the impact of this project. The model outputs a ~70% confidence interval of 2.2million to 2.7 billion, i.e., 3 orders of magnitude wide.^[9] In addition, each time you refresh Guesstimate, you get different results (and sometimes negative numbers), though the median is usually within the same order of magnitude.
3. I’m relatively inexperienced with using Guesstimate and making cost-effectiveness estimates, and I have little domain expertise to inform my intuitions on this topic. So, I’m guessing that this model contains errors and inaccuracies^[10].
4. My estimate differs significantly from an estimate made by ALLFED about their own cost-effectiveness – I discuss the differences below.

With that said, here’s a description of the model inputs and how I estimated them.

Cost estimate: 

• Because of my team’s focus on scalable projects, I estimate the cost (and impact) of not just a single pilot, but a series of increasingly large-scale pilots – roughly “as many as we would need” (though I realize this is vague).
• I made a quick guess of ~10-100 million.
• This is roughly in line with ALLFED’s cost estimates (source):
  • $1-30 million for planning
  • $10-100 million for research
  • $10-100 million for development
  • $10-100 billion for training

Impact estimate: 

• X-risk from nuclear war this century: ~0.18%. (based on the numbers below and sanity checked against Toby Ord’s estimate in The Precipice (p. 287) of ~0.1%)
  • Risk of a major nuclear war this century: ~5-26%.
    • Toby Ord estimates a 5% chance of a “full-scale nuclear war” this century. (source)
    • Luisa Rodriguez estimates a 0.38% annual probability of nuclear war between Russia and the US. Using this as a rough proxy for a “full-scale nuclear war” and multiplying through for 78 more years gives ~26% chance of a US-Russia nuclear war. (Though this calculation assumes no other x-risk this century, so it is likely a significant overestimate.) (source)
  • Risk of nuclear winter conditional on nuclear war: ~11%
    • I just used Luisa Rodriguez’ estimate. (source)
  • X-risk constituted by a nuclear winter: 1-20%
    • This is just my intuition.
• Proportion of x-risk (conditional on nuclear winter) solved if a series of pilots succeed in ensuring that governments or philanthropists have strong resilient food plans (taking into account that this might have happened anyway): 5-30%
  • This is just my intuition.
• Probability of a series of pilots succeeding in making governments or philanthropists have strong resilient food plans in place: 15-60%.
  • This is just my intuition.

Overall estimate: ~$260million per 0.01% x-risk reduction (in absolute terms) (70% confidence interval: 2.2 million to 2.7B^[11]). Because the Guesstimate model varied so much, I refreshed it 11 times (not counting refreshes that resulted in a negative final number) and used the median of the values displayed across the refreshes.

How this relates to ALLFED’s own cost-effectiveness estimates

• ALLFED, an organization researching resilient foods, has published several estimates of its own cost-effectiveness (most recently in this paper based on survey responses from GCR researchers; previously in this model).
• ALLFED’s cost-effectiveness estimates are estimates of the impact of ALLFED’s project as a whole, not just of implementing pilots of one of ALLFED’s proposed resilient food solutions. However, I think the estimates are still roughly comparable as:
  • We should expect the cost-effectiveness of the whole project and a small subject of the project to be relatively similar, unless we think pilots are an unusually important or unimportant part of the project.
  • It seems to me like the difference between our model outputs is less due to the difference in what we’re measuring and more due to different inputs going into the model.
• Overall, ALLFED’s own cost-effectiveness estimates are approximately 100x as optimistic as mine (i.e., they estimate that this kind of intervention is ~100x more effective than I do), though both of our median estimates imply reasonable cost-effectiveness.
  • Retroactive comment: As noted in the introduction, the cost-effectiveness bar has probably changed as a result of recent changes to the EA funding landscape, so it’s now less clear whether this project has “reasonable cost-effectiveness” on my estimate (though it’s plausible to me that the estimated cost-effectiveness would still be above the new funding bar).
• Ways in which ALLFED’s model seems to differ from mine^[12]:
  • Key crux 1: X-risk from nuclear war.
    • ALLFED’s most recent paper estimates a 1-90% x-risk conditional on full-scale nuclear war; I estimate a 0.12-2% x-risk conditional on full-scale nuclear war.
    • I’m not sure what underlying factors we disagree about. I’m guessing it’s a mix of (a) me thinking relatively narrowly about the impacts of food shortages while ALLFED’s survey respondents thought more broadly about all impacts of a nuclear war, and (b) differences in beliefs / intuitions about the importance of food shortages as an x-risk factor.
  • Key crux 2: X-risk from 10% agricultural shortfall scenarios (and similar).
    • In addition to x-risk via a nuclear winter leading to major (>30%) agricultural shortfalls, ALLFED’s estimates include x-risk via 10% agricultural shortfalls (which is estimated in the most recent paper to be 0.001% to 40%). I naively guessed that the x-risk from 10% agricultural shortfalls would be trivial relative to the x-risk from nuclear winter scenarios, so did not include it.
  • Key crux 3: Optimism about a project like this reducing x-risk.
    • It’s harder to pin down our exact disagreements here since I estimate the effect of a particular, narrow intervention while ALLFED estimates the effect of a much broader set of projects than just this one. But it seems roughly like ALLFED is something like 2-5x more optimistic than me about the x-risk reduction caused by projects like this.

Downside risks

Overall, the downsides of this project seem relatively limited to me.

Here are some potential downside risks I thought of in a 10 minute brainstorm:

1. Moral hazard (discussed by ALLFED and included in their model): If governments know that they have solid plans in place to feed their populations through a nuclear winter, the incentive to avoid nuclear war could be reduced. (And equivalently for other GCRs.)
2. Misallocation of resources: I (and other researchers and grantmakers in relatively privileged positions) might be biased towards interventions that might help (relatively privileged) people survive in the event of a global disaster, and overestimate the extent to which it’s justified/cost-effective to spend resources on these projects relative to projects that help the most disadvantaged today.
   1. This risk could be mitigated by (a) designing the project to focus on ensuring that it is not only the most privileged who would survive a potential global disaster, and (b) where possible and cost-effective, sharing findings from the project that could be applied to near-term poverty alleviation.
3. Reputation risk to project founders and affiliates: Even if the project does not fall prey to the above risk, it could be portrayed as doing so, which could cause reputational risk to project founders and affiliates. Similar mitigations apply.
4. Dual use technological development: Information about how to rapidly scale production of goods could be misused to scale the production of harmful things.

Other actors in this space

There are already EA-aligned actors working in this space, one of which is planning to run pilots of this kind of project (at a relatively small scale) if funding is secured.

• The Alliance to Feed the Earth in Disasters (ALLFED) researches resilient food solutions that could be scaled up in case of a nuclear winter and other disasters (and their work is the main inspiration for this project).
  • So far, ALLFED’s primary focus has been research, but ALLFED has run some pilot projects to test the feasibility of rapidly scaling up food production.
  • ALLFED is planning to run further pilots if they can secure funding for it.
• A team at Penn State University is working to “develop, test, and optimize strategies for emergency food resilience”, with ~$3 million in funding from Open Philanthropy^[13]; I haven’t looked into their work and don’t know the extent to which they’re planning to run demonstrations / pilots.
• (There might be other actors in this space that I've missed – I would love to hear about them if so.)

This makes me think that marginal work in this area is less valuable than I would otherwise have thought, and I’m not confident that it would be better for a potential entrepreneur to start a new project rather than join an existing one.

However, there are still reasons why new projects in this space could be valuable, including:

1. Having more independent projects might mean having more shots at succeeding in developing solid food production plans for a disaster.
2. Having more actors in this space with different approaches could lead to fruitful exchanges, collaboration and/or competition to improve the existing projects.
3. Actors with more entrepreneurial experience might add significant value relative to existing projects.
4. I’m not confident in the degree to which the strategy of existing projects is geared specifically towards x-risk reduction.

Key uncertainties

1. How strong should we expect the relationship between food scarcity and existential collapse to be? Is there (strong) evidence that food scarcity/insecurity are linked to conflict?
2. How robustly useful to different GCR scenarios would (the most robustly useful) resilient food solutions be? If so, can a (strong) case be made for the cost-effectiveness of particular resilient food solutions based on the diversity of scenarios in which they might be useful, despite those scenarios individually being relatively unlikely?
3. I generally feel like civilisational collapse and recovery are difficult to reason about and like I have only a vague understanding of what these actually look like in practice; and so I feel like I’ve probably missed some key considerations as a result.

Acknowledgements

David Denkenberger, Juan García Martínez, and Vasco Amaral Grilo of ALLFED provided me with generous input and discussion on this speedrun. I also got helpful comments from my colleagues Ben Snodin, Jam Kraprayoon, Linch Zhang, Michael Aird, Rachel Norman, and Renan Araujo.

This research is a project of Rethink Priorities. It was written by Marie Davidsen Buhl. If you like our work, please consider subscribing to our newsletter. You can explore our completed public work here.

1. ^{^}

    I’m not necessarily claiming that the project can in fact scale to this size or that the cost-effectiveness is constant with scale.

2. ^{^}

    My quick guess is that this project still clears our (new, implicit) cost-effectiveness bar; but the changes to the funding landscape reduce my confidence in this. It’s important to know that I was significantly uncertain both before and after the changes to the funding landscape.

3. ^{^}

    Thanks to Juan García Martínez for pointing out that resilient food interventions need not involve unconventional foods or production methods.

4. ^{^}

    Excluding those in the tab “Other solutions” as these are less cost-effective according to ALLFED.

5. ^{^}

    My purpose in sketching out these pathways is to get a rough sense of what the steps to an existential catastrophe might look like, in order to identify points of intervention or key weaknesses in the theory of change of this project. Note that I don’t think the particular scenarios I describe here are the only relevant ones – of course, any reasonably detailed sketch of a pathway to existential catastrophe is likely to be wrong.

6. ^{^}

    Vasco Grilo pointed out to me that in combined heating and cooling scenarios (e.g., extreme climate change + nuclear winter), the effects on agriculture would probably cancel each other out to some degree, such that resilient foods would be significantly less useful (see e.g., note 6.90 in the notes to What We Owe the Future).

7. ^{^}

    Potential counterpoint: According to Dave Denkenberg, the UK government has estimated that a 10% reduction in global food production due to simultaneous extreme weather events has an annual likelihood of 1% before 2050 and 3% for the rest of the century (which, if naively multiplied, adds up to about an 80% chance of such an event this century). I think Dave was referring to this report, but I haven’t cross-checked. If true, that would imply a high likelihood of climate change causing severe agricultural shortfalls this century, though I feel uncertain about how serious an x-risk factor a 10% agricultural shortage would constitute (the latter point is discussed more in this section).

8. ^{^}

    I would not have thought of this pathways if not for the fact that events without climate effects are mentioned as examples of catastrophic risks to food production on ALLFED’s website.

9. ^{^}

    It is somewhat misleading to interpret this as a 70% confidence interval as I did not consistently input 70% confidence intervals for each input. While it’s not accurate to call this a 70% confidence interval in a technical sense, I think it roughly gets across, in a more qualitative/less technical sense, the right level of uncertainty.

10. ^{^}

     E.g., this comment describes some errors in a previous cost-effectiveness model I made that also apply to this model.

11. ^{^}

     Again, it’s somewhat misleading to call this a 70% confidence interval (see footnote 9).

12. ^{^}

     I really appreciate David Denkenberger and Juan García Martínez’s engagement with a previous version of this section, which helped me understand the cruxes of disagreement much better.

13. ^{^}

     Thanks to Juan García Martínez for making me aware of this project.