# Nuclear war tail risk has been exaggerated?

The views expressed here are my own, not those of Alliance to Feed the Earth in Disasters (ALLFED), for which I work as a contractor.

# Summary

• I calculated a nearterm annual risk of human extinction from nuclear war of 5.93*10^-12 (more).

• I consider grantmakers and donors interested in decreasing extinction risk had better focus on artificial intelligence (AI) instead of nuclear war (more).

• I would say the case for sometimes prioritising nuclear extinction risk over AI extinction risk is much weaker than the case for sometimes prioritising natural extinction risk over nuclear extinction risk (more).

• I get a sense the extinction risk from nuclear war was massively overestimated in The Existential Risk Persuasion Tournament (XPT) (more).

• I have the impression Toby Ord greatly overestimated tail risk in The Precipice (more).

• I believe interventions to decrease deaths from nuclear war should be assessed based on standard cost-benefit analysis (more).

• I think increasing calorie production via new food sectors is less cost-effective to save lives than measures targeting distribution (more).

# Extinction risk from nuclear war

I calculated a nearterm annual risk of human extinction from nuclear war of 5.93*10^-12 (= (6.36*10^-14*5.53*10^-10)^0.5) from the geometric mean between[1]:

• My prior of 6.36*10^-14 for the annual probability of a war causing human extinction.

• My inside view estimate of 5.53*10^-10 for the nearterm annual probability of human extinction from nuclear war.

By nearterm annual risk, I mean that in a randomly selected year from 2025 to 2050. I computed my inside view estimate of 5.53*10^-10 (= 0.0131*0.0422*10^-6) multiplying:

• 1.31 % annual probability of a nuclear weapon being detonated as an act of war.

• 4.22 % probability of insufficient calorie production given at least one nuclear detonation.

• 10^-6 probability of human extinction given insufficient calorie production.

I explain the rationale for the above estimates in the next sections. Note nuclear war might have cascade effects which lead to civilisational collapse[2], which could increase longterm extinction risk while simultaneously having a negligible impact on the nearterm one I estimated. I do not explicitly assess this in the post, but I guess the nearterm annual risk of human extinction from nuclear war is a good proxy for the importance of decreasing nuclear risk from a longtermist perspective:

• My prior implicitly accounts for the cascade effects of wars. I derived it from historical data on the deaths of combatants due to not only fighting, but also disease and starvation, which are ever-present indirect effects of war.

• Nuclear war might have cascade effects, but so do other catastrophes.

• Global civilisational collapse due to nuclear war seems very unlikely to me. For instance, the maximum destroyable area by any country in a nuclear 1st strike was estimated to be 65.3 k km^2 in Suh 2023 (for a strike by Russia), which is just 70.8 % (= 65.3*10^3/​(92.2*10^3)) of the area of Portugal, or 3.42 % (= 65.3*10^3/​(1.91*10^6)) of the global urban area.

• Even if nuclear war causes a global civilisational collapse which eventually leads to extinction, I guess full recovery would be extremely likely. In contrast, an extinction caused by advanced AI would arguably not allow for a full recovery.

• I am open to the idea that nuclear war can have longterm implications even in the case of full recovery, but considerations along these lines would arguably be more pressing in the context of AI risk.

• For context, William MacAskill said the following on The 80,000 Hours Podcast. “It’s quite plausible, actually, when we look to the very long-term future, that that’s [whether artificial general intelligence is developed in “liberal democracies” or “in some dictatorship or authoritarian state”] the biggest deal when it comes to a nuclear war: the impact of nuclear war and the distribution of values for the civilisation that returns from that, rather than on the chance of extinction”.

• Nevertheless, value lock-in (for better or worse) is arguably more cost-effectively ensured via influencing the development of AI.

• Appealing to cascade effects or other known unknowns feels a little like a regression to the inscrutable, which is characterised by the following pattern:

• Arguments for high existential risk initially focus on aspects of the risk which are relatively better understood (e.g. famine deaths due to the climatic effects of nuclear war).

• Further analysis frequently shows the risk from such aspects has been overestimated, and is in fact quite low (e.g. nearterm risk of human extinction from nuclear war).

• Then discussions move to more poorly understood aspects of the risk (e.g. how the distribution of values after a nuclear war affects the longterm values of transformative AI).

In any case, I recognise it is a crucial consideration whether nearterm annual risk of human extinction from nuclear war is a good proxy for the importance of decreasing nuclear risk from a longtermist perspective. I would agree further research on this is really valuable.

Additionally, I appreciate one should be sceptical whenever a model outputs a risk as low as the ones I mentioned at the start of this section. For example, a model predicting a 1 in a trillion chance of the global real gross domestic product (real GDP) decreasing from 2008 to 2009 would certainly not be capturing most of the actual risk of recession then, which would come from that model being (massively) wrong. On the other hand, one should be careful not to overgeneralise this type of reasoning, and conclude that any model outputting a small probability must be wrong by many orders of magnitude (OOMs). The global real GDP decreased 0.743 % (= 1 − 92.21/​92.9) from 2008 to 2009, largely owing to the 2007–2008 financial crisis, but such a tiny drop is a much less extreme event than human extinction. Basic analysis of past economic trends would have revealed global recessions are unlikely, but perfectly plausible. In contrast, I see historical data suggesting a war causing human extinction is astronomically unlikely.

Finally, one could claim I am underestimating the risk due to not adequately accounting for unknown unknowns. I agree, but:

• I might as well be overestimating it for the same reasons. To illustrate, one knows nothing about absolutely unknown unknowns, and therefore should not expect them to move the best guess for the risk up or down[3].

• In the real world of probabilities, if not in that of logic, absence of evidence is evidence of absence.

• I have the impression best guesses for tail risk and cost-effectiveness usually go down[4].

• It is harder to decrease the risks from unknown unknowns because there is less information about them.

• Unknown unknowns also affect other risks, and it is unclear whether the unknown unknowns surrounding nuclear and AI risk are such that I am underestimating the importance of the former relative to the latter.

## Annual probability of a nuclear weapon being detonated as an act of war

I estimated an annual probability of a nuclear weapon being detonated as an act of war of 1.31 % (= 1 - (1 − 0.29)^(1/​(2050 − 2024))), which I got from Metaculus’ community prediction on 23 January 2024 of 29 % before 2050. My annual probability is 1.03 (= 0.0131/​0.0127) times the base rate of 1.27 % (= 179), respecting nuclear detonations in one year over the last 79 (= 2023 − 1945 + 1), which seems reasonable.

## Probability of insufficient calorie production given at least one nuclear detonation

I determined a probability of (globally) insufficient calorie production given at least one nuclear detonation of 4.22 %. I computed this running a Monte Carlo simulation with 1 M samples and independent distributions[5], and supposing:

• The number of nuclear detonations given at least one being detonated as an act of war, as a fraction of the total of 12.5 k in 2023, is described by a beta distribution with 61st percentile (= 1 − 0.39) of 0.800 % (= 100/​(12.5*10^3)), and 89th percentile (= 1 − 0.11) of 8.00 % (= 1*10^3/​(12.5*10^3)), which has alpha and beta parameters of 0.190 and 6.68, and mean of 2.77 %. The 61st and 89th percentiles correspond to Metaculus’ community predictions on 2 February 2024 of 39 % and 11 % probability of over 100 and 1 k offensive nuclear detonations before 2050 given at least one nuclear detonation causing a fatality before 2050.

• The fraction of nuclear detonations which are countervalue[6] is represented by a beta distribution with 25th and 75th percentiles equal to 3.7 % and 63.0 %, in agreement with Metaculus’ community predictions on 2 February 2024. This beta distribution has alpha and beta parameters of 0.364 and 0.682, and mean of 34.8 %.

• The mean equivalent yield of the countervalue nuclear detonations is 121 kt[7] (= 2,559*10^6/​21,234), which I got from the ratio between:

• The soot injected into the stratosphere per equivalent yield is the maximum likelihood lognormal distribution given 2 independent estimates of 3.15*10^-5 and 0.00215 Tg/​kt.

• I arrived at these by adjusting results from Reisner 2018 and Reisner 2019, and Toon 2008 and Toon 2019.

• The mean and standard deviation of the logarithm of the distribution I just mentioned are equal to the mean and unadjusted standard deviation of the logarithms of the 2 estimates, which are −8.25 and 2.11[9].

• For references, my mean soot injected into the stratosphere per equivalent yield is 0.00242 Tg/​kt, which is 1.13 (= 0.00242/​0.00215) times my higher estimate. The reasons for this are the distribution having to be quite wide for one to be maximally likely to observe 2 very different estimates, and the mean of a lognormal distribution increasing with its uncertainty[10].

• Minimum soot injected into the stratosphere for insufficient calorie production of 84.2 Tg (= 47 + (150 − 47)/​(2.38 − 1.08)*(2.38 − 1.91)). This is the minimum for insufficient calorie consumption in year 2[11], less than 1.91 k kcal/​person/​d, given equitable food distribution, consumption of all edible livestock feed, and no household food waste, linearly interpolating the data of Fig. 5a of Xia 2022:

• 47 Tg for 2.38 k kcal/​person/​d[12].

• 150 Tg for 1.08 k kcal/​person/​d[12].

• The net effect on calorie production of all the adaptation measures is similar to assuming equitable food distribution, consumption of all edible livestock feed, and no household food waste. To the extent these 3 are needed to mitigate famine nationally, I guess they would be roughly fully implemented nationally, but not globally. Nevertheless, there are other factors contributing towards Xia 2022 overestimating famine (relatedly, see resilient food solutions):

• The baseline conditions in Xia 2022 refer to 2010, but the world is becoming increasingly more resilient against starvation. The death rate from protein-energy malnutrition decreased 77.7 % (= 1 - (0.00274 %)/​(0.0123 %)) from 1990 to 2019[13].

• Foreign aid to the more affected countries, including international food assistance.

• Increase in meat production per capita from 2010, which is the reference year in Xia 2022.

• Increase in real GDP per capita from 2010, which is relevant because poverty is a major risk factor for famines.

• Replacing forest and grazing land by cropland:

• In 2016, grazing land was 2.06 (= 3.28/​1.59) times as large as cropland, so this would become 3.06 (= 1 + 2.06) times as large given full replacement.

• In 2019, forest land was 85.5 % (= 0.3758/​0.4394) as large as cropland, so this would become 1.86 (= 1 + 0.855) times as large given full replacement.

• I am not claiming full replacement would be possible or needed, but the above illustrates there is great margin to increase cropland.

• “Scenarios assume that all stored food is consumed in Year 1”, so there is room for better rationing.

• “We do not consider farm-management adaptations such as changes in cultivar selection, switching to more cold-tolerating crops or greenhouses31 and alternative food sources such as mushrooms, seaweed, methane single cell protein, insects32, hydrogen single cell protein33 and cellulosic sugar34”.

• “Large-scale use of alternative foods, requiring little-to-no light to grow in a cold environment38, has not been considered but could be a lifesaving source of emergency food if such production systems were operational”.

• “Byproducts of biofuel have been added to livestock feed and waste27. Therefore, we add only the calories from the final product of biofuel in our calculations”. However, it would have been better to redirect to humans the crops used to produce biofuels.

• It is possible to have a relatively low famine death rate with a calorie consumption lower than 1.91 k kcal/​person/​d:

• The calorie supply (to households) in the Central African Republic (CAR) in 2015 was 1.73 k kcal/​person/​d. I assume household waste is quite negligible there, such that the calorie consumption is similar to the calorie supply.

• The deaths from protein-energy malnutrition there in that year were 1.38 k, equal to 0.0286 % (= 1.38*10^3/​(4.82*10^6)) of CAR’s population in 2015. For context, global deaths from protein-energy malnutrition in 2019 were 238 k, equal to 0.00307 % (= 238*10^3/​(7.76*10^6)) of the global population.

• One of the anonymous reviewers commented low reported calorie supply values like CAR’s in 2015 are underestimates due to smuggling, which would imply a greater death rate from malnutrition than the above if the real supply matched the reported one. Yet, this effect is offset by Xia 2022 not considering the underreported calories. In other words, it is still possible to have a relatively low famine death rate with a reported, if not actual, calorie consumption lower than 1.91 k kcal/​person/​d.

• The same reviewer commented that an actual calorie consumption of 1.7 k kcal/​person/​d “is not sustainable, and literally killed people in WW2”, as described in Taste of War: World War II and the Battle for Food. I agree 1.7 k kcal/​person/​d is far from optimal for adults[14], but I doubt it would reduce life expectancy to less than 2 years, such that it could be sustained during the worst years of the nuclear winter in Xia 2022, 2 and 3. Calorie consumption in the coastal village of Kaul (Papua New Guinea) was 1.68 k kcal/​person/​d (= (1.94 + 1.42)/​2) based on the mean values provided by Norgan 1974 for 51 adult men and 69 adult women[15].

## Probability of human extinction given insufficient calorie production

I obtained a probability of human extinction given insufficient calorie production of 10^-6 (= 1/​10^6), considering 1 M years is the typical lifespan of a mammal species[16]. For context:

• See Luisa Rodriguez’ and Carl Shulman’s general arguments and considerations about the possibility of civilisation collapse leading to extinction. Here are Luisa’s:

• “Historical survival and resilience”.

• “The grace period”.

• “With population loss comes “decorrelation” of survivors”.

• “Non-uniformity of the initial catastrophe’s impacts”.

• “The population loss would have to be incredibly extreme to lead to extinction”.

• Inequitable food distribution tendentially decrease extinction risk:

• For example, with 1 k kcal/​person/​d and equitable distribution, everyone would starve because that is less than the resting energy expenditure, which is 1.14 k kcal/​person/​d according to Fig. 5 of Xia 2022.

• Nonetheless, with inequitable distribution, there is room for part of the population to have enough calories. From Table S2 of the supplementary information of Xia 2022, Australia’s major food crops and marine fish production in year 2 of a nuclear winter involving 47 and 150 Tg would be 36.0 % and 24.2 % higher than under normal conditions.

• My probability seems compatible with Luke Oman, one of the 3 authors of Robock 2007, having guessed a risk of human extinction of 0.001 % to 0.01 % for an injection of soot into the stratosphere of 150 Tg.

• According to Fig. 5a of Xia 2022, 150 Tg would result in a calorie consumption 56.5 % (= 1.08*10^3/​(1.91*10^3)) as large as that for 84.2 Tg given equitable food distribution, consumption of all edible livestock feed, and no household food waste.

• So Luke’s guess for the extinction risk would presumably be significantly lower for 84.2 Tg.

# Grantmakers and donors interested in decreasing extinction risk had better focus on artificial intelligence instead of nuclear war

Supposedly cause neutral grantmakers aligned with effective altruism have influenced 15.3 M\$[17] (= 0.03 + 5*10^-4 + 2.70 + 3.56 + 0.0488 + 0.087 + 5.98 + 2.88) towards efforts aiming to decrease nuclear risk[18]:

I encourage grantmakers to be more transparent by sharing further information about their grants. The extension of the public write-ups respecting the grants above ranged from 1 sentence to 6 paragraphs, with the median being 1 paragraph[21].

I consider the grant to Will was worth it, as I can see it having contributed to him now being a “researcher in longtermist AI strategy” at Metaculus. All of the others seem way less cost-effective than the current marginal grants of LTFF, which are overwhelmingly aimed at decreasing AI risk:

• I guess the nearterm annual extinction risk from AI is 1.69 M (= 10^-5/​(5.93*10^-12)) times that from nuclear war. This assumes an nearterm annual extinction risk from AI of 0.001 %, which I motivate later in the section.

• I consider the annual spending on decreasing extinction risk from AI is 35.4 (= 4.04*10^9/​(114*10^6)) times that on decreasing extinction risk from nuclear war. I determined this from the ratio between:

• 4.04 G\$ (4.04 billion USD) on nuclear risk in 2020, which I got from the mean of a lognormal distribution with 5th and 95th percentile equal to 1 and 10 G\$, corresponding to the lower and upper bound guessed in 80,000 Hours’ profile on nuclear war. “This issue is not as neglected as most other issues we prioritise. Current spending is between \$1 billion and \$10 billion per year (quality-adjusted)” (see details).

• 114 M\$ (= (79.8 + 32 + 2*1)*10^6) on “AI safety research that is focused on reducing risks from advanced AI” in 2023:

• 79.8 M\$ from the National Science Foundation (NSF), LTFF, Open Philanthropy, SFF and “other”.

• “~\$32m per year” from “for-profit companies” (Anthropic, Conjecture, Google Deepmind and OpenAI).

• 2 times “probably at least \$1m per year” from “individual donors”.

• So the nearterm annual extinction risk per annual spending for AI risk is 59.8 M (= 1.69*10^6*35.4) times that for nuclear risk.

• It would be super hard for the best interventions to decrease nuclear risk to be so many OOMs more tractable that they overturn the massive difference in importance and neglectedness illustrated above (relatedly).

• Consequently, I consider grantmakers and donors interested in decreasing extinction risk had better focus on AI instead of nuclear war.

Some caveats:

• I expect AI risk will become much less neglected in the next few decades, and the cost-effectiveness of interventions to decrease AI risk to significantly drop as a result.

• Interventions to decrease nuclear risk have indirect effects which will tend to make their cost-effectiveness more similar to that of the best interventions to decrease AI risk. I guess the best marginal grants to decrease AI risk are much less than 59.8 M times as cost-effective as those to decrease nuclear risk. At the same time:

• It can still make sense for cause neutral grantmakers to recommend donors who are not so to support interventions to decrease nuclear risk[22]. The alternative may well be less cost-effective, and supporting interventions to decrease nuclear risk could be a pathway towards influencing more pressing areas.

I arrived at a nearterm annual extinction risk from AI of 0.001 % as follows. I think looking into how species have gone extinct in the past is the best reference class to estimate AI risk. Jacob Steinhardt did an analysis which has some relevant insights:

Thus, in general most species extinctions are caused by:

• A second species which the original species has not had a chance to adapt to. This second species must also not be reliant on the original species to propagate itself.

• A catastrophic natural disaster or climate event.

• Habitat destruction or ecosystem disruption caused by one of the two sources above.

I believe we have pretty good reasons to think the 2nd point applies much more weakly to humans than animals, but the 1st holds if one sees advanced AI as analogous to a new species[23]. I would still claim deaths in past terrorist attacks and wars provide a strong basis for arguing that humans will not go extinct via an AI war or terrorist attack. However, the 1st point alludes to what seems to me to be the greatest risk from AI, natural selection favouring AIs over humans. Since 1 M years is the typical lifespan of a mammal species, my prior extinction risk from AI in a random year this century is 10^-6 (= 1/​10^6). Further accounting for inside view considerations, I guess the extinction risk from AI in a random year from 2025 to 2050 is 0.001 %. Relatedly, I encourage readers to check Zach Freitas-Groff’s post on AGI Catastrophe and Takeover: Some Reference Class-Based Priors.

I should note I do not consider humans being outcompeted by AI as necessarily bad (relatedly). I strongly endorse expected total hedonistic utilitarianism (ETHU), and I would be surprised if humans were the most efficient way of increasing welfare longterm. At the same time, minimising nearterm extinction risk from AI seems like a good heuristic to align it with ETHU.

# The case for sometimes prioritising nuclear extinction risk over AI extinction risk is much weaker than the case for sometimes prioritising natural extinction risk over nuclear extinction risk

## Cost-effectiveness of decreasing extinction risk from nuclear war

I guess lobbying for nuclear arsenal limitation is one of the most cost-effective interventions to decrease nearterm extinction risk from nuclear war. The Centre for Exploratory Altruism Research (CEARCH) estimated it averts disability-adjusted life years (DALYs) 5.25 k times as cost-effectively as GiveWell’s top charities, although:

The headline cost-effectiveness will almost certainly fall if this cause area is subjected to deeper research: (a) this is empirically the case, from past experience; and (b) theoretically, we suffer from optimizer’s curse (where causes appear better than the mean partly because they are genuinely more cost-effective but also partly because of random error favouring them, and when deeper research fixes the latter, the estimated cost-effectiveness falls).

Despite this, lobbying for nuclear arsenal limitation still looks promising among interventions to decrease nuclear risk. For context, CEARCH estimated, subject to the caveat above too, that conducting a pilot study of a resilient food source would be 14 times as cost-effective as GiveWell’s top charities, i.e. just 0.267 % (= 14/​(5.25*10^3)) as cost-effective as lobbying for nuclear arsenal limitation.

CEARCH determined lobbying for nuclear arsenal limitation decreases 9*10^-10 of the nuclear risk per dollar, but I guess the actual cost-effectiveness is only 1 % as high, such that it is only 52.5 (= 0.01*5.25*10^3) times as cost-effective as GiveWell’s top charities at averting DALYs. Consequently, I guess lobbying for nuclear arsenal limitation decreases 9*10^-12 (= 0.01*9*10^-10) of the nuclear risk per dollar, which respects a cost-effectiveness of decreasing nearterm extinction risk from nuclear war of 5.34*10^-7 bp/​T\$[24] (= 9*10^-12*5.93*10^-12).

## Cost-effectiveness of decreasing extinction risk from asteroids and comets

Salotti 2022 estimated the extinction risk from 2023 to 2122 from asteroids and comets is 2.2*10^-12 (see Table 1). This comes from the probability of long period comets with a diameter larger than 100 km colliding with Earth[25], for which the warning time is shorter than 5 years (see Table 1). The nearterm annual extinction risk from asteroids and comets respecting Salotti 2022 is 2.20*10^-14 (= 1 - (1 − 2.2*10^-12)^(1/​100)).

Jean-Marc Salotti, the author of Salotti 2022, guesses it would cost hundreds of billions of dollars to design and test shelters which would decrease the extinction risk from asteroids and comets by 50 %[26]. I supposed a cost of 182 G\$ (= 2/​(1/​10^3 + 1100)*10^9), which is the reciprocal of the mean of the reciprocal of a uniform distribution ranging from 100 to 1 k G\$[27]. So I guess the cost-effectiveness of decreasing nearterm extinction risk from asteroids and comets is 6.04*10^-10 bp/​T\$ (= 0.50*2.20*10^-14/​(182*10^9)).

## Comparisons

According to my estimates, the cost-effectiveness of decreasing nearterm extinction risk from nuclear war via lobbying for nuclear arsenal limitation is 884 (= 5.34*10^-7/​(6.04*10^-10)) times that from decreasing nearterm extinction risk from asteroids and comets via shelters. However, I do not think one can conclude from this high ratio that lobbying for nuclear arsenal limitation is better than working on shelters, as these would decrease extinction risk from not only asteroids and comets, but also other risks, including nuclear war.

On the other hand, I would say the case for sometimes prioritising nuclear extinction risk over AI extinction risk is much weaker than the case for sometimes prioritising natural extinction risk over nuclear extinction risk:

• The ratio of 884 between the cost-effectiveness of decreasing nuclear and asteroids and comets risk is many orders of magnitude lower than the ratio of 59.8 M I calculated between the nearterm annual extinction risk per annual spending of AI and nuclear risk.

• The conclusion just above is reinforced if one believes there are more pressing natural risks besides those from asteroids and comets. According to Toby Ord’s guesses given in The Precipice, the existential risk from 2021 to 2120 from supervolcanic eruptions, his largest natural risk, is 100 (= 10^-4/​10^-6) times that from asteroids and comets.

• However, I am not that moved by Toby’s estimate for the existential risk from supervolcanic eruptions.

• I believe extinction risk from these is many OOMs lower, as arguably proved to be the case for asteroids and comets.

Further research to increase the resilience of my cost-effectiveness estimates would be useful.

# Extinction risk from nuclear war was massively overestimated in The Existential Risk Persuasion Tournament

I collected in the table below the predictions of the superforecasters, domain experts, general existential risk experts, and non-domain experts of XPT for the risk of human extinction from nuclear war. The estimates respect the medians across 88 superforecasters, 13 domain experts, 14 general existential risk experts, and 58 non-domain experts.

The superforecasters’, domain experts’, general existential risk experts’, and non-domain experts’ annual risk of human extinction from nuclear war from 2023 to 2050 is 602 k (= 3.57*10^-6/​(5.93*10^-12)), 7.23 M (= 4.29*10^-5/​(5.93*10^-12)), 10.3 M (= 6.08*10^-5/​(5.93*10^-12)) and 4.22 M (= 2.50*10^-5/​(5.93*10^-12)) times my nearterm annual risk. So I get a sense the extinction risk from nuclear war was massively overestimated in XPT. Do you agree? If yes, should one put little trust in other estimates of extinction risk from XPT? I think so. Still, I believe the XPT was quite valuable given the wealth of information shared in the report explaining the rationale for the forecasts (see Appendix 7).

One could argue the large gap between XPT’s estimates and mine points to me not having sufficiently updated my independent impression. I agree epistemic deference is valuable in general, but it is unclear to me whether I should be deferring more:

• I am familiar with what informed XPT’s nuclear extinction risk predictions, having read the respective sections “Sources of agreement, disagreement and uncertainty”, “Arguments given for low-end forecasts”, and “Arguments given for higher-end forecasts” (pp. 298 to 303).

• Some participants in the XPT seemed to believe in a much lower nuclear extinction risk than the medians I presented (emphasis mine):

• “Most forecasters whose probabilities were near the median factored in a range of possible risks, including world wars, nuclear winters, and even artificial-intelligence-driven NERs [nuclear extinction risks], but concluded that even under worst case scenarios, the extinction of humanity (give or take 5000 people) would be near impossible...even if an NER [nuclear existential risk] had set humanity on a path that made eventual extinction a foregone conclusion, existing resources on earth would allow at least 5000 survivors to hang on for seventy-eight years”.

• “For many, the thought of getting to less than 5000 humans alive was simply too far fetched an outcome and they couldn’t be persuaded otherwise in what they saw as credible scenarios”.

• “[T]he set of circumstances required for this to happen are quite low, though obviously not impossible. These circumstances are that there will be a nuclear conflict between 2 nations both capable and willing to fire at everyone everywhere between the two of them: ‘very bad case scenarios’ where India and Pakistan, or the US and Russia, or China and anyone else, fired everything they had at just each other, or even at each other and each other’s close allies, would likely not cause extinction…it requires some of the big nuclear powers to decide to try to take literally everyone down with them, and that they actually succeed”.

• “So we think that the probabilities in this question are dominated by scenarios of total nuclear war before 2050 which cause civilizational and climate collapse to the point where long-term survival becomes impossible to save for very well-prepared shelters. But even pessimistic scenarios seem unlikely to lead to a collapse that is fast enough to reduce the global population to below 5000 by 2100”.

• There aren’t compelling arguments on the higher end for this question again due to the fact that this is a very high bar to achieve”.

• “The team predicts that there will be pockets of people who survive in various regions of the world. Their survival may be at Neolithic standards, but there will be tribes of people who band together and restart mankind. After all, many mammals survived the asteroid and ice age that killed the dinosaurs”.

• “[A] certain number of team members feel that even if there was a full strategic exchange and usage of all of the world’s nuclear arsenal still humanity would be able to keep its numbers over 5000. The argument for this is the number [a]nd population of uncontacted tribes, or isolated human populations like the Easter island population pre-contact, that have managed to hold numbers of over 5000 in extremely harsh conditions”.

• [A]lmost certainly some people would survive on islands or in caves given even the worst of worst cases”.

• “Southern Hemisphere likely to be less impacted – New Zealand, Madagascar, Pacific Islands, Highlands of Papua New Guinea, unlikely to be targeted and include areas with little global and technology dependence…Just the population of Antarctica in its summer is ~5000 people. Even small islands surviving could easily mean more than 5k people”.

• [There are s]everal regions in the world that would not be affected by nuclear conflict directly and have decent climatic conditions to support 100 of millions even in a NW [nuclear winter]”.

• I believe my estimate involved much more explicit modelling than XPT’s.

• There is very little formal evidence on the accuracy of forecasting very rare events like human extinction[30].

• In general, I suspect there is a tendency to give probabilities between 1 % and 99 % for events whose mechanics we do not understand well, like the factors involved in a product to estimate the chance of extinction.

• Such a range encompasses the vast majority (98 % = 0.99 − 0.01) of the available linear space (from 0 to 1), and forecasting questions are often formulated with the aim of reasonable predictions falling in that range.

• However, the available logarithmic space is infinitely vast, and it is hard to rule out an astronomically low extinction risk. In contrast, extinction risk could be overly high if it implies a too low probability of our current existence.

• So there is margin for moderate guesses (e.g. between 1 % and 99 %) to be major overestimates.

As a side note, the superforecasters predicted the annual risk from 2023 to 2100 is 7.59 (= 9.49*10^-6/​(1.25*10^-6)) times that from 2023 to 2030, the domain experts 2.83 (= 7.07*10^-5/​(2.50*10^-5)) times, the general existential risk experts 2.40 (= 9.01*10^-5/​(3.75*10^-5)) times, and the non-domain experts 1.95 (= 2.44*10^-5/​(1.25*10^-5)) times, i.e. all expected the risk to increase throughout this century. Interestingly, none foresaw major changes to the median number of nuclear warheads by 2040, which is some evidence against large increases in nuclear arsenals. Relative to the 12,705 in 2022 (see pp. 532 and 533):

• 31 superforecasters predicted 13,500, i.e. an increase of 6.26 % (= 13,50012,705 − 1).

• 1 domain expert predicted 11,990, i.e. a decrease of 5.63 % (= 1 − 11,99012,705).

• 5 general existential risk experts predicted 10,200, i.e. a decrease of 19.7 % (= 1 − 10,20012,705).

• 10 non-domain experts predicted 12,952.5, i.e. a decrease of 1.95 % (= 1 − 12,952.5/​12,705).

Consequently, I think the superforecasters, domain experts, general existential risk experts, and non-domain experts implicitly predicted at least one of the following. Nuclear war becoming more frequent, having a greater potential to escalate[31], or humanity becoming less resilient to it. I only seem to agree with the 2nd of these.

# Toby Ord greatly overestimated tail risk in The Precipice

I collected in the table below Toby’s annual existential risk from 2021 to 2120 from AI, nuclear war, and asteroids and comets based on his guesses given in The Precipice. I also added my estimates for the nearterm annual extinction risk from the same 3 risks, and the ratio between Toby’s values and mine. The values are not directly comparable, because Toby’s refer to existential risk and mine to extinction risk. Nonetheless, I still have the impression Toby greatly overestimated tail risk. This is in agreement with David Thorstad’s series exaggerating the risks, which includes subseries on climate, AI and bio risk, and discusses Toby’s book The Precipice.

The estimates of the tail risk from asteroids and comets are arguably the most robust, so it is interesting there is a large difference between Toby’s and mine even there. There are many concepts of existential catastrophe[34], but I do not think one can say existential risk from asteroids and comets is anything close to 455 k times as high as extinction risk from these:

• In The Precipice, Toby says the probability of an asteroid larger than 10 km colliding with Earth in the next 100 years is lower than 1 in 150 M (Table 3.1), and guesses that the risk from comets larger than 10 km is similarly large (p. 72), which implies a total risk from asteroids and comets larger than 10 km of around 1.33*10^-8 (= 2/​(150*10^6)). This is only 1.33 % (= 1.33*10^-8/​10^-6) of Toby’s guess for the existential risk from asteroids and comets, which implies Toby expects the vast majority of existential risk to come from asteroids and comets smaller than 10 km.

• The last mass extinction “was caused by the impact of a massive asteroid 10 to 15 km (6 to 9 mi) wide”, and happened 66 M years ago. It involved an impact winter, which played a role in the extinction of the dinosaurs, and may well have contributed to the emergence of mammals and ultimately humans.

• So Toby would expect an asteroid impact similar to that of the last mass extinction to be an existential catastrophe. Yet, at least ignoring anthropics, I believe the probability of not fully recovering would only be 0.0513 % (= e^(-10^9/​(132*10^6))), assuming:

• An exponential distribution with a mean of 132 M years (= 66*10^6*2) represents the time to go from i) human extinction due to such an asteroid to ii) evolving a species as capable as humans at steering the future. I supposed this on the basis that:

• An exponential distribution with a mean of 66 M years describes the time between extinction threats as well as that to go from i) to ii) conditional on no extinction threats.

• Given the above, extinction and full recovery are equally likely. So there is a 50 % chance of full recovery, and one should expect the time until full recovery to be 2 times (= 10.50) as long as that conditional on no extinction threats.

• The above evolution could take place in the next 1 billion years during which the Earth will remain habitable.

• In addition, one should arguably suppose a species as capable as humans at steering the future would have similarly good values, even if different.

• Setting the existential risk from asteroids and comets to the extinction risk estimated in Salotti 2022 seems much more legitimate, as it relies on a threshold of 100 km for the impactor. This is 1 OOM larger, and 3 OOMs more energetic[35] than the asteroid involved in the last mass extinction, thus having the potential to cause the extinction of not only humans, but also of many other species in our evolutionary path.

# Interventions to decrease deaths from nuclear war should be assessed based on standard cost-benefit analysis

I believe interventions to decrease deaths from nuclear war should be assessed based on standard cost-benefit analysis (CBA):

• Having in mind my astronomically low nearterm annual extinction risk from nuclear war, it is unclear to me whether interventions to decrease deaths from nuclear war decrease extinction risk more cost-effectively than broader ones, like the best interventions to boost economic growth or decrease disease burden (e.g. GiveWell’s top charities).

• I expect extinction risk can be decreased much more cost-effectively by focussing on AI risk rather than nuclear risk. So I would argue interventions to decrease deaths from nuclear war can only be competitive under an alternative worldview, like ones where the goal is boosting economic growth or decreasing disease burden.

Moreover, I would propose using standard CBAs not only in the political sphere, as argued by Elliott Thornley and Carl Shulman, but also outside of it. In terms of what grantmakers aligned with effective altruism have been doing[36]:

I wonder whether the best interventions to decrease deaths from nuclear war would, based on in-depth CBAs, be better than donating to GiveWell’s All Grants Fund. From the ones above, I guess only nuclear arsenal limitation would be so.

# Increasing calorie production via new food sectors is less cost-effective to save lives than measures targeting distribution

Nuclear winter is a major source of risk of global catastrophic food failures. Nonetheless, my estimates imply the annual probability of a nuclear war causing (globally) insufficient calorie production is 0.0553 % (= 0.0131*0.0422). This suggests food distribution rather than production will be the bottleneck to decrease famine deaths in the vast majority of circumstances, as is the case today[37]. So I think increasing calorie production via new (or massively scaled up) food sectors, like greenhouse crop production, lignocellulosic sugar, methane single cell protein or seaweed, is less cost-effective to save lives than measures targeting distribution, like ones aiming to ensure the continuation of international food trade.

One of the anonymous reviewers commented the aforementioned new food sectors “are definitely helpful for loss of international trade scenarios”. I suspect the reviewer has something like the following in mind:

• From Fig. 5b of Xia 2022, the minimum soot injected into the stratosphere for insufficient calorie consumption is 10 Tg[38] given no international food trade, consumption of all edible livestock feed, and no household food waste.

• In contrast, I estimated a minimum soot injected into the stratosphere for insufficient calorie production of 84.2 Tg, supposing the net effect on calorie production of all the adaptation measures is similar to assuming equitable food distribution, consumption of all edible livestock feed, and no household food waste.

I think the reviewer may be concluding from the above that, given no international food trade, calorie consumption would be much lower, and therefore increasing food production via new food sectors would become much more important relative to distribution. I agree with the former, but not the latter. Loss of international food trade is more of a problem of food distribution than production. If this increased thanks to new food sectors, but could not be distributed to low-income food-deficit countries (LIFDCs) due to loss of trade, there would still be many famine deaths there. Many LIFDCs are in tropical regions too, where there is a smaller decrease in crop yields during a nuclear winter (see Fig. 4 of Xia 2022).

Furthermore, greater loss of trade and supply chain disruptions will be associated with greater loss of population and infrastructure, which in turn will arguably make solutions relying on new food sectors less likely to be successful relative to ones leveraging existing sectors. Examples of the latter include decreasing animal and biofuel production which relies on edible crops, expanding crop area, and using more cold-tolerant crops.

My point about distribution rather than production being a bottleneck loses strength as the severity of the nuclear winter increases. For an injection of soot into the stratosphere of 150 Tg, the calorie consumption given equitable food distribution, consumption of all edible livestock feed, and no household food waste would be 1.08 k kcal/​person/​d (see Fig. 5a of Xia 2022), which is just 56.5 % (= 1.08*10^3/​(1.91*10^3)) of the minimum caloric requirement. Producing more calories would be crucial in this case. Moreover, Xia 2022’s 150 Tg scenario involves 4.4 k nuclear detonations (see Table 1). The disruptions to international food trade caused by these would be so extensive that it would be especially useful for countries to have local resilience, such as by producing their own food.

Finally, there is a risk that focussing on new food sectors counterfactually increases the suffering of farmed animals without decreasing starvation (not to mention the meat-eater problem). Some countries may not need to consume all edible livestock feed to mitigate starvation, in which case increasing production from new food sectors could allow for greater consumption of farmed animals with bad lives. Somewhat relatedly, I have very mixed feelings about promoting resilient food solutions which rely on increasing factory-farming, such as ALLFED mentioning insects.

# Acknowledgements

Thanks to Anonymous Person 1, Anonymous Person 2, Anonymous Person 3, Anonymous Person 4, Carl Robichaud, Ezra Karger, Farrah Dingal, Matthew Gentzel, Nuño Sempere and Ross Tieman for feedback on the draft[39]. Thanks to Jean-Marc Salotti for guessing the cost of shelters which would decrease the extinction risk from asteroids and comets.

1. ^

The geometric mean between 2 small probabilities is similar to the probability linked to the geometric mean of the odds of the 2 probabilities.

2. ^

For instance, Bailey 2017 analyses the effects of interruptions at chokepoints in global food trade. “Critical junctures on transport routes through which exceptional volumes of trade pass”. A reviewer highlighted other cascade effects which might lead to civilisational collapse: loss of major world governments, major changes in the distribution military of power; loss of power grids, fuel supply chains, and many machines and devices through direct destruction and nuclear electromagnetic pulses (nuclear EMPs); loss of major nodes in the financial and transportation system; uncontrolled wildfires; and further crop and animal losses from radiation.

3. ^

In reality, people use the term unknown unknowns to refer to considerations about which we have some understanding.

4. ^

Note that best guesses going down is often weak evidence that they were overestimates. A best guess should in expectation stay the same, but this is compatible with it being more likely to go down than up. The expected value of a heavy-tailed distribution can be much larger than its median, so it can be quite likely that one’s best, respecting the expected value, goes down as one updates towards a distribution with less uncertainty.

5. ^

The running time is 0.5 s.

6. ^

Jeffrey Lewis clarified on The 80,000 Hours Podcast there is not a sharp distinction between counterforce and countervalue:

And so just to explain that a little bit, or unpack that: if you look at what the United States says about its nuclear weapons today, we are explicit that we target things that the enemy values, and we are also explicit that we follow certain interpretations of the law of armed conflict. And it is absolutely clear in those legal writings that the United States does not target civilians intentionally, but that in conducting what you might call “counterforce,” there is a list of permissible targets. And they include not just nuclear forces. I think often in the EA community, people assume counterforce means nuclear forces, because it’s got the word “force,” right? But it’s not true. So traditionally, the US targets nuclear forces and all of the supporting infrastructure — including command and control, it targets leadership, it targets other military forces, and it targets what used to be called “war-supporting industries,” but now are called “war-sustaining industries.”

In the context of the Metaculus’ prediction:

A strike is considered countervalue for these purposes if credible media reporting does not widely consider a military or industrial target as the primary target of the attack (except in the case of strikes on capital cities, which will automatically be considered countervalue for this question even if credible media report that the rationale for the strike was disabling command and control structures).

7. ^

I did not model this as a distribution because its uncertainty is much smaller than that in other factors for the cases I am interested in (relatedly). I am analysing extinction risk, so I want the distribution to be accurate for cases with many detonations. Since the mean equivalent yield tends to a constant as the detonations tend to the available nuclear warheads, I think using a constant is appropriate. In addition, the importance of modelling a factor in a product as a distribution decreases with the number of factors which are already being modelled as a distribution. If N factors follow independent lognormal distributions whose ratio between the 95th and 5th percentile is r, the ratio between the 95th and 5th percentile of the distribution of the product is r^(N^0.5). The exponent grows sublinearly with the number of factors, so the relative increase in the uncertainty of the product is smaller if one is already modelling many of its factors as distributions.

8. ^

The equivalent yield is defined in Suh 2023 such that it is proportional to the destroyable area. From equations 1 and 2, the equivalent yield is proportional to the yield to the power of 23 if the yield is smaller than 1 Mt, and to the yield to the power of 12 if the yield is larger than 1 Mt. I actually think the (maximum) burnable area is proportional to the yield, thus being larger than the destroyable area estimated in Suh 2023. On the other hand, the actual burned area will be smaller than the burnable area, which counteracts the effect of using a higher exponent of 1. In any case, using an exponent of 1 instead of 23 to estimate the equivalent yield only makes the burnable area 1.14 times as large for the nuclear arsenal of the United States in 2023. So I guess the question of which exponent to use is not that important, especially in the context of estimating extinction risk.

9. ^

By unadjusted standard deviation, I mean the square root of the unadjusted variance.

10. ^

The mean of a lognormal distribution can be expressed as m*e^(sigma^2/​2), where m is the median of the lognormal, and sigma is the standard deviation of the logarithm of the lognormal.

11. ^

In Xia 2022, “the soot is arbitrarily injected during the week starting on May 15 of Year 1”.

12. ^

I obtained high precision based on the pixel coordinates of the relevant points, which I retrieved with Paint.

13. ^

Interestingly, the annual FAO Food Price Index (FFPI), which “is a measure of the monthly [and annual] change in international prices of a basket of food commodities”, increased 51.0 % (= 95.1/​63.0 − 1) during the same period (calculated based on values in column B of tab “Annual” of the excel file “Excel: Nominal and real indices from 1990 onwards (monthly and annual)”). So the FFPI is not a good proxy for the death rate from protein-energy malnutrition. I believe this is explained by most people on the edge of starvation being subsistence farmers who are not much affected by market prices. Apparently, “roughly 65 percent of Africa’s population relies on subsistence farming. Subsistence farming, or smallholder agriculture, is when one family grows only enough to feed themselves. Without much left for trade, the surplus is usually stored to last the family until the following harvest”.

14. ^

From Akisaka 1996, “the energy intake of the Okinawan centenarians living at home was about 1,100 kcal/​day for both sexes, which was similar to that of centenarians throughout Japan”. I do not particularly trust this because food consumption was assessed based on self-reports. “The dietary survey was done by one 24h recall method, as was done for centenarians living throughout Japan (3)”.

15. ^

In these studies, I am always worried about food consumption being estimated based on self-reports, but this should not be an issue in Norgan 1974. “All of the food eaten by each individual subject was weighed after cooking (where applicable) and immediately before consumption. Food consumed in the house was weighed on a robust Avery balance, weighing to 1 kg in 10 g divisions, using a large bowl scale-pan. The balances were frequently calibrated. Masses were recorded to the nearest 5 g. Left-over portions or inedible portions were also weighed and subtracted from the initial mass. Subjects were followed when they left the immediate vicinity of the house and food eaten away from the house was weighed on a portable Salter dietary balance weighing up to 500 g in 5 g divisions. A light plastic jug and plate were used for liquids such as coconut water”.

There would also have been margin to further decrease calorie consumption via reducing physical activity. “The way of life for all the people was moderately active—more so in the highlands [not in Kaul] - since they were subsistence farmers cultivating their own gardens for food”.

16. ^

Humans are a mammal species.

17. ^

Excluding the grants from Longview Philanthropy’s Nuclear Weapons Policy Fund (NWPF), whose size is not publicly available, but I do not think including them would significantly change the total. The grants to decrease nuclear risk from Longview’s Emerging Challenges Fund (ECF) only represent 0.569 % (= 0.087/​15.3) of my total, and I guess NWPF has not granted more than 1 OOM more money than that linked to ECF’s grants to decrease nuclear risk.

I included a grant of 500 \$ made by the Effective Altruism Infrastructure Fund (EAIF), but decided not to describe it (besides mentioning the size here). This is quite small, so I was worried identifying the grantee could be a little mean. In addition, the grantee asked me not to mention the grant.

18. ^

I listed the grantmakers alphabetically.

19. ^

I emailed Longview’s Head of Grants Management & Compliance, Andrew Player, about this on 29 January 2024. He said it was a busy time, and that he would respond in due course.

20. ^

These grants were made in the context of Open Philanthropy’s global catastrophic risks portfolio. In contrast, this and this grants to increase food resilience against abrupt sunlight reduction scenarios (nuclear, volcanic or impact winters) were made under the global health and wellbeing portfolio.

21. ^

16th smallest/​largest write-up of a total of 31. 12 were no longer than 1 sentence, and 26 no longer than 1 paragraph.

22. ^

For example, Founders Pledge is a cause neutral organisation that advises some donors who are not so, such as ones partial to climate. Longview Philanthropy is a philanthropic advisory service, so I guess it operates under similar constraints, supporting some donors who are not cause neutral.

23. ^

I would update towards a higher extinction risk from wars relative to advanced AI systems if interspecific competition was more common relative to intraspecific one.

24. ^

1 bp/​T\$ corresponds to 0.01 percentage points per 1 trillion dollars.

25. ^

One of the anonymous reviewers guessed comets larger than 10 km would still have a 20 % chance of causing extinction while being 500 times as likely as ones larger than 100 km. This would imply the extinction risk from comets larger than 10 km being 100 (= 0.2*500) times as large as that from ones larger than 100 km. As a result, the point I am making in this section would become stronger by 2 OOMs.

Salotti 2022 justifies the threshold of 100 km as follows:

A 10 km sized asteroid could threaten large populations on Earth but there would still exist safe places on Earth to survive (Sloan et al., 2017, Toon et al., 1994, Chapman and Morrison, 1994, Mathias et al., 2017, RUMPF et al., 2017, Collins et al., 2005).

I opted to rely on Salotti 2022’s mainline estimate in my post, but I have not looked into the studies above. Less importantly, I also think a lower extinction risk per time makes more sense for shorter periods, given a less strict requirement for extended survival, and my nearterm annual extinction risk from nuclear war respects a period of 26 years (= 2050 − 2025 + 1), whereas Salotti 2022’s estimate concerns one of 100 years.

26. ^

Information provided via email.

27. ^

I used this because E(“cost-effectiveness”) = E(“benefits”/​“cost”) = E(“benefits”)*E(1/​“cost”) = E(“benefits”)/​(1/​E(1/​“cost”)), assuming benefits and cost are independent.

28. ^

See pp. 293 and 294.

29. ^

“Annual risk” = 1 - (1 - “total risk”)^(1/​“duration of the period in years”). The periods have durations of 8 (= 2030 − 2023 + 1), 28 (= 2050 − 2023 + 1) and 78 (= 2100 − 2023 + 1) years.

30. ^

Additionally, there is very little formal evidence on the accuracy of long-range forecasting (I am only aware of Tetlock 2023), but this is arguably not as important because I am only relying on XPT’s extinction risk until 2030.

31. ^

Including via a more heavy-tailed distribution of the number of nuclear warheads.

32. ^

Ordered from the largest to the smallest.

33. ^

“Annual risk” = 1 - (1 - “total risk”)^(1/​“duration of the period in years”). The period has a duration of 100 years (= 2120 − 2021 + 1).

34. ^

I prefer focussing on clearer metrics.

35. ^

Kinetic energy is proportional to mass, and the mass of a sphere is proportional to its diameter to the power of 3. Kinetic energy is also proportional to speed to the power of 2, but I am guessing the impact speed is independent of the size.

36. ^

I listed the grantmakers alphabetically.

37. ^

Note I am not arguing prices will stay constant. I am claiming prices will go up mostly due to limitations in food distribution rather than production.

38. ^

Eyeballed.

39. ^

Names ordered alphabetically.

40. ^

I obtained high precision based on the pixel coordinates of the relevant points, which I retrieved with Paint.

• I agree that people should not focus on nuclear risk as a direct extinction risk (and have long argued this), see Toby’s nuke extinction estimates as too high, and would assess measures to reduce damage from nuclear winter to developing neutral countries mainly in GiveWell-style or ordinary CBA terms, while considerations about future generations would favor focus on AI, and to a lesser extent bio.

However, I do think this wrongly downplays the effects on our civilization beyond casualties and local damage of a nuclear war that wrecks the current nuclear powers, e.g. on disrupting international cooperation, rerolling contingent nice aspects of modern liberal democracy, or leading to release of additional WMD arsenals (such as bioweapons, while disrupting defense against those weapons). So the ‘can nuclear war with current arsenals cause extinction’ question misses most of the existential risk from nuclear weapons, which is indirect in contributing to other risks that could cause extinction or lock-in of permanent awful regimes. I think marginal philanthropic dollars can save more current lives and help the overall trajectory of civilization more on other risks, but I think your direct extinction numbers above do greatly underestimate how much worse the future should be expected to be given a nuclear war that laid waste to, e.g. NATO+allies and the Russian Federation.

You dismiss that here:

> Then discussions move to more poorly understood aspects of the risk (e.g. how the distribution of values after a nuclear war affects the longterm values of transformative AI).

But I don’t think it’s a huge stretch to say that a war with Russia largely destroying the NATO economies (and their semiconductor supply chains), leaving the PRC to dominate the world system and the onrushing creation of powerful AGI, makes a big difference to the chance of locked-in permanent totalitarianism and the values of one dictator running roughshod over the low-hanging fruit of many others’ values. That’s very large compared to these extinction effects. It also doesn’t require bets on extreme and plausibly exaggerated nuclear winter magnitude.

Similarly, the chance of a huge hidden state bioweapons program having its full arsenal released simultaneously (including doomsday pandemic weapons) skyrockets in an all-out WMD war in obvious ways.

So if one were to find super-leveraged ways reduce the chance of nuclear war (this applied less to measures to reduce damage to nonbelligerent states) then in addition to beating GiveWell at saving current lives, they could have big impacts on future generations. Such opportunities are extremely scarce, but the bar for looking good in future generation impacts is less than I think this post suggests.

• Thanks for sharing your thoughts, Carl!

I agree that people should not focus on nuclear risk as a direct extinction risk (and have long argued this), see Toby’s nuke extinction estimates as too high, and would assess measures to reduce damage from nuclear winter to developing neutral countries mainly in GiveWell-style or ordinary CBA terms, while considerations about future generations would favor focus on AI, and to a lesser extent bio.

Thanks for mentioning these points. Would you also rely on ordinary CBAs to assess interventions to decrease the direct damage of nuclear war? I think this would still make sense.

So the ‘can nuclear war with current arsenals cause extinction’ question misses most of the existential risk from nuclear weapons, which is indirect in contributing to other risks that could cause extinction or lock-in of permanent awful regimes.

At the same time, the nearterm extinction risk from AI also misses most of the existential risk from AI? I guess you are implying that the ratio between nearterm extinction risk and total existential risk is lower for nuclear war than for AI.

Related to your point above, I say that:

Regarding:

You dismiss that [“effects on our civilization beyond casualties and local damage of a nuclear war”] here:

> Then discussions move to more poorly understood aspects of the risk (e.g. how the distribution of values after a nuclear war affects the longterm values of transformative AI).

Note I mention right after this that:

In any case, I recognise it is a crucial consideration whether nearterm annual risk of human extinction from nuclear war is a good proxy for the importance of decreasing nuclear risk from a longtermist perspective. I would agree further research on this is really valuable.

You say that:

I don’t think it’s a huge stretch to say that a war with Russia largely destroying the NATO economies (and their semiconductor supply chains), leaving the PRC to dominate the world system and the onrushing creation of powerful AGI, makes a big difference to the chance of locked-in permanent totalitarianism and the values of one dictator running roughshod over the low-hanging fruit of many others’ values, one very large compared to these extinction effects. It also doesn’t require bets on extreme and plausibly exaggerated nuclear winter magnitude.

I agree these are relevant considerations. On the other hand:

• The US may want to attack China in order not to relenquish its position as global hegemon.

• I feel like there has been little research on questions like:

• How much it would matter if powerful AI was developped in the West instead of China (or, more broadly, in a democracy instead of autocracy).

• The likelihood of lock-in.

On the last point, your piece is a great contribution, but you say:

Note that we’re mostly making claims about feasibility as opposed to likelihood.

However, the likelihood of lock-in is crucial to assess the strength of your points. I would not be surprised if the chance of an AI lock-in due to a nuclear war was less than 10^-8 this century.

In terms of nuclear war indirectly causing extinction:

at least ignoring anthropics, I believe the probability of not fully recovering would only be 0.0513 % (= e^(-10^9/​(132*10^6))), assuming:

• An exponential distribution with a mean of 132 M years (= 66*10^6*2) represents the time to go from i) human extinction due to such an asteroid to ii) evolving a species as capable as humans at steering the future. I supposed this on the basis that:

• An exponential distribution with a mean of 66 M years describes the time between extinction threats as well as that to go from i) to ii) conditional on no extinction threats.

• Given the above, extinction and full recovery are equally likely. So there is a 50 % chance of full recovery, and one should expect the time until full recovery to be 2 times (= 10.50) as long as that conditional on no extinction threats.

• The above evolution could take place in the next 1 billion years during which the Earth will remain habitable.

In contrast, if powerful AI caused extinction, control over the future would arguably permanently be lost.

plausibly exaggerated nuclear winter magnitude.

Similarly, the chance of a huge hidden state bioweapons program having its full arsenal released simultaneously (including doomsday pandemic weapons) skyrockets in an all-out WMD war in obvious ways.

Is there any evidence for this?

this applied less to measures to reduce damage to nonbelligerent states

Makes sense. If GiveWell’s top charities are not a cost-effective way of improving the longterm future, then decreasing starvation in low income countries in a nuclear winter may be cost-effective in terms of saving lives, but has semingly negligible impact on the longterm future too. Such countries just have too little influence on transformative technologies.

• Rapid fire:

• Nearterm extinction risk from AI is wildly closer to total AI x-risk than the nuclear analog

• My guess is that nuclear war interventions powerful enough to be world-beating for future generations would look tremendous in averting current human deaths, and most of the WTP should come from that if one has a lot of WTP related to each of those worldviews

• Re suspicious convergence, what do you want to argue with here? I’ve favored allocation on VOI and low-hanging fruit on nuclear risk not leveraging AI related things in the past less than 1% of my marginal AI allocation (because of larger more likely near term risks from AI with more tractability and neglectedness); recent AI developments tend to push that down, but might surface something in the future that is really leveraged on avoiding nuclear war

• I agree not much has been published in journals on the impact of AI being developed in dictatorships

• Re lock-in I do not think it’s remote (my views are different from what that paper limited itself to) for a CCP-led AGI future,

• Thanks for following up!

Re suspicious convergence, what do you want to argue with here?

Sorry for the lack of clarity. Some thoughts:

• The 15.3 M\$ grantmakers aligned with effective altruism have influenced aiming to decrease nuclear risk seem mostly optimised to decrease the nearterm damage caused by nuclear war (especially the spending on nuclear winter), not the more longterm existential risk linked to permanent global totalitarianism.

• As far as I know, there has been little research on how a minor AI catastrophe would influence AI existential risk (although wars over Taiwan have been wargamed). Looking into this seems more relevant than investigating how a non-AI catastrophe would influence AI risk.

• The risk from permanent global totalitarianism is still poorly understood, so research on this and how to mitigate it seems more valuable than efforts focussing explicitly on nuclear war. There might well be interventions to increase democracy levels in China which are more effective to decrease that risk than interventions aimed at ensuring that China does not become the sole global hegemon after a nuclear war.

• I guess most of the risk from permanent global totalitarianism does not involve any major catastrophes. As a data point, the Metaculus’ community predicts an AI dystopia is 5 (= 0.19/​0.037) times as likely as a paperclipalypse by 2050.

I agree not much has been published in journals on the impact of AI being developed in dictatorships

More broadly, which pieces would you recommend reading on this topic? I am not aware of substantial blogposts, although I have seen the concern raised many times.

• Any probability as low as 5.93*10^-12 about something as difficult to model as the effects of nuclear war on human society seems extremely overconfident to me. Can you really make 15.93*10^-12 (170 billion) predictions about independent topics and expect to be wrong only once? Are you 99.99% [edit: fixed this number] sure that there is no unmodeled set of conditions under which civilizational collapse occurs quickly, which a nuclear war is at least 0.001% likely to cause? I think the minimum probabilities that one should have given these considerations is not much lower than the superforecasters’ numbers.

• Thanks for the comment, Thomas!

Any probability as low as 5.93*10^-12 about something as difficult to model as the effects of nuclear war on human society seems extremely overconfident to me.

I feel like this argument is too general. The human body is quite complex too, but the probability of a biological human naturally growing to have 10 m is still astronomically low. Likewise for the risk of asteroids and comets, and supervolcanoes.

Nuclear war being difficult to model means more uncertainty, but not necessarily higher risk. There are infinitely many orders of magnitude between 0 and 5.93*10^-12, so I think I can at least in theory be quite uncertain while having a low best guess. I understand greater uncertainty (e.g. higher ratio between the 95th and 5th percentile) holding the median constant tends to increase the mean of heavy-tailed distributions (like lognormals), but it is unclear to which extent this applies. I have also accounted for this by using heavy-tailed distributions whenever I thought appropriate (e.g. I modelled the soot injected into the stratosphere per equivalent yield as a lognormal).

Can you really make 15.93*10^-12 (170 billion) predictions about independent topics and expect to be wrong only once? Are you 99.9999% sure that there is no unmodeled set of conditions under which civilizational collapse occurs quickly, which a nuclear war is at least 0.001% likely to cause?

Nitpick. I think you have to remove 2 9s in the 2nd sentence, because the annual chance of nuclear war is around 1 %.

I do not think I have calibrated intuitions about the probability of rare events. To illustrate, I suppose it is easy for someone (not aware of the relevant dynamics) to guess the probability of a ticket winning the lottery is 0.1 %, whereas it could in fact be 10^-8. Relatedly:

Additionally, I appreciate one should be sceptical whenever a model outputs a risk as low as the ones I mentioned at the start of this section. For example, a model predicting a 1 in a trillion chance of the global real gross domestic product (real GDP) decreasing from 2008 to 2009 would certainly not be capturing most of the actual risk of recession then, which would come from that model being (massively) wrong. On the other hand, one should be careful not to overgeneralise this type of reasoning, and conclude that any model outputting a small probability must be wrong by many orders of magnitude (OOMs). The global real GDP decreased 0.743 % (= 1 − 92.21/​92.9) from 2008 to 2009, largely owing to the 2007–2008 financial crisis, but such a tiny drop is a much less extreme event than human extinction. Basic analysis of past economic trends would have revealed global recessions are unlikely, but perfectly plausible. In contrast, I see historical data suggesting a war causing human extinction is astronomically unlikely.

I think the minimum probabilities that one should have given these considerations is not much lower than the superforecasters’ numbers.

I would be curious to see a model based as much as possible on empirical evidence suggesting a much higher risk.

• Don’t have time to reply in depth, but here are some thoughts:

• If a risk estimate is used for EA cause prio, it should be our betting odds /​ subjectie probabilities, that is, average over our epistemic uncertainty. If from our point of view a risk is 10% likely to be >0.001%, and 90% likely to be ~0%, this lower bounds our betting odds at 0.0001%. It doesn’t matter that it’s more likely to be 0%.

• Statistics of human height are much better understood than nuclear war because we have billions of humans but no nuclear wars. The situation is more analogous to finding the probability of a 10 meter tall adult human having only ever observed a few thousand monkeys (conventional wars), plus one human infant (WWII) and also knowing that every few individuals humans mutate into an entirely new species (technological progress).

• It would be difficult to create a model suggesting a much higher risk because most of the risk comes from black swan events. Maybe one could upper bound the probability by considering huge numbers of possible mechanisms for extinction and ruling them out, but I don’t see how you could get anywhere near 10^-12.

• If a risk estimate is used for EA cause prio, it should be our betting odds /​ subjectie probabilities, that is, average over our epistemic uncertainty. If from our point of view a risk is 10% likely to be >0.001%, and 90% likely to be ~0%, this lower bounds our betting odds at 0.0001%. It doesn’t matter that it’s more likely to be 0%.

Agreed. I expect my estimate for the nearterm extinction risk from nuclear war to remain astronomically low.

Statistics of human height are much better understood than nuclear war because we have billions of humans but no nuclear wars. The situation is more analogous to finding the probability of a 10 meter tall adult human having only ever observed a few thousand monkeys (conventional wars), plus one human infant (WWII) and also knowing that every few individuals humans mutate into an entirely new species (technological progress).

My study of the monkeys and infants, i.e. my analysis of past wars, suggested an annual extinction risk from wars of 6.36*10^-14, which is still 1.07 % (= 5.93*10^-12/​(5.53*10^-10)) of my best guess.

It would be difficult to create a model suggesting a much higher risk because most of the risk comes from black swan events. Maybe one could upper bound the probability by considering huge numbers of possible mechanisms for extinction and ruling them out, but I don’t see how you could get anywhere near 10^-12.

For the superforecasters’ annual extinction risk from nuclear war until 2050 of 3.57*10^-6 to be correct, my model would need to miss 99.9998 % (= 1 − 5.93*10^-12/​(3.57*10^-6)) of the total risk. You say most (i.e. more than 50 %) of the risk comes from black swan events, but I think it would be really surprising if 99.9998 % did? The black swan events would also have to absent in some sense from XPT’s report, because my estimate accounts for the information I found there.

I should also clarify my 10^-6 probability of human extinction given insufficient calorie production is supposed to account for unknown unknowns. Otherwise, my extinction risk from nuclear war would be orders of magnitude lower.

• My study of the monkeys and infants, i.e. my analysis of past wars, suggested an annual extinction risk from wars of 6.36*10^-14, which is still 1.07 % (= 5.93*10^-12/​(5.53*10^-10)) of my best guess.

The fact that one model of one process gives a low number doesn’t mean the true number is within a couple orders of magnitude of that. Modeling mortgage-backed security risk in 2007 using a Gaussian copula gives an astronomically low estimate of something like 10^-200, even though they did in fact default and cause the financial crisis. If the bankers adjusted their estimate upward to 10^-198 it would still be wrong.

IMO it is not really surprising for very near 100% of the risk of something to come from unmodeled risks, if the modeled risk is extremely low. Like say I write some code to generate random digits, and the first 200 outputs are zeros. One might estimate this at 10^-200 probability or adjust upwards to 10^-198, but the probability of this happening is way more than 10^-200 due to bugs.

• The fact that one model of one process gives a low number doesn’t mean the true number is within a couple orders of magnitude of that.

Agreed. One should not put all weight in a single model. Likewise, one’s best guess for the annual extinction risk from wars should not update to (Stephen Clare’s) 0.01 % just because one model (Pareto distribution) outputs that. So the question of how one aggregates the outputs of various models is quite important. In my analysis of past wars, I considered 111 models, and got an annual extinction risk of 6.36*10^-14 for what I think is a reasonable aggregation method. You may think my aggregation method is super wrong, but this is different from suggesting I am putting all weight into a single method. Past analyses of war extinction risk did this, but not mine.

IMO it is not really surprising for very near 100% of the risk of something to come from unmodeled risks, if the modeled risk is extremely low. Like say I write some code to generate random digits, and the first 200 outputs are zeros. One might estimate this at 10^-200 probability or adjust upwards to 10^-198, but the probability of this happening is way more than 10^-200 due to bugs.

If it was not for considerations like the above, my best guess for the nearterm extinction risk from nuclear war would be many orders of magnitude below my estimate of 10^-11. I would very much agree that a risk of e.g. 10^-20 would be super overconfident, and not pay sufficient attention to unknown unknowns.

• There are more discussions re: appropriate ways to think about model uncertainty in comments between Vasco, Stephen Clare, and myself here.

• guesses it would cost hundreds of billions of dollars to design and test shelters

I looked at the referenced article and could not find a mention of this sum of money—could you please point me to where exactly Salotti makes this guess?

• I think the reviewer may be concluding from the above that, given no international food trade, calorie consumption would be much lower, and therefore increasing food production via new food sectors would become much more important relative to distribution. I agree with the former, but not the latter. Loss of international food trade is more of a problem of food distribution than production. If this increased thanks to new food sectors, but could not be distributed to low-income food-deficit countries (LIFDCs) due to loss of trade, there would still be many famine deaths there. Many LIFDCs are in tropical regions too, where there is a smaller decrease in crop yields during a nuclear winter (see Fig. 4 of Xia 2022).

Another factor is that if countries are aware of the potential of scaling up resilient foods, they would be less likely to restrict trade. Therefore, I’m thinking the outcomes might be fairly bimodal, with one scenario of resilient food production and continued trade, and another scenario of not having resilient food production and loss of trade, potentially more than just food trade, perhaps with loss of industrial civilization or worse.

Yet, at least ignoring anthropics, I believe there would be a probability of full recovery of 100 % (= 1 - e^(-10^9/​(66*10^6))) even then, assuming:

• An exponential distribution for the time to go from i) human extinction due to such an asteroid to ii) evolving a species as capable as humans at steering the future, with mean equal to the aforementioned 66 M years.

• The above evolution could take place in the next 1 billion years during which the Earth will remain habitable.

I think this assumes a scenario where, after the asteroid that causes human extinction, the next billion years are large asteroid/​comet free, which is not a good assumption.

• Thanks for the comments, David.

Another factor is that if countries are aware of the potential of scaling up resilient foods, they would be less likely to restrict trade. Therefore, I’m thinking the outcomes might be fairly bimodal, with one scenario of resilient food production and continued trade, and another scenario of not having resilient food production and loss of trade, potentially more than just food trade, perhaps with loss of industrial civilization or worse.

I agree that is a factor, but I guess the distribution of the severity of catastrophes caused by nuclear war is not bimodal, because the following are not binary:

• Awareness of mitigation measures:

• More or less countries can be aware.

• Any given country can be more or less aware.

• Ability to put in practice the mitigation measures.

• Export bans:

• More or less countries can enforce them.

• Any given country can enforce them more or less.

In addition, I have the sense historical more local catastrophes are not bimodal, following distributions which more closely resemble a power law, where more extreme outcomes are increasingly less likely.

I think this assumes a scenario where, after the asteroid that causes human extinction, the next billion years are large asteroid/​comet free, which is not a good assumption.

Good point! I have updated the relevant bullet in the post:

• So Toby would expect an asteroid impact similar to that of the last mass extinction to be an existential catastrophe. Yet, at least ignoring anthropics, I believe the probability of not fully recovering would only be 0.0513 % (= e^(-10^9/​(132*10^6))), assuming:

• An exponential distribution with a mean of 132 M years (= 66*10^6*2) represents the time to go from i) human extinction due to such an asteroid to ii) evolving a species as capable as humans at steering the future. I supposed this on the basis that:

• An exponential distribution with a mean of 66 M years describes the time between extinction threats as well as that to go from i) to ii) conditional on no extinction threats.

• Given the above, extinction and full recovery are equally likely. So there is a 50 % chance of full recovery, and one should expect the time until full recovery to be 2 times (= 10.50) as long as that conditional on no extinction threats.

• The above evolution could take place in the next 1 billion years during which the Earth will remain habitable.

Now the probability of not fully recovering is 0.0513 %, i.e. 1.95 k (= 5.13*10^-4/​(2.63*10^-7)) times as high as before. Yet, the updated unconditional existential risk (extinction caused by the asteroid and no full recovery afterwards) is still astronomically low, 3.04*10^-15 (= 5.93*10^-12*5.13*10^-4). So my point remains qualitatively the same.

I have also added the 2nd sentence in the following bullet:

• Even if nuclear war causes a global civilisational collapse which eventually leads to extinction, I guess full recovery would be extremely likely. In contrast, an extinction caused by advanced AI would arguably not allow for a full recovery.

• Nuclear risk philanthropy is about 30M/​y, it seems you are comparing overall nuclear risk effort to philanthropic effort for AI?

In terms of philanthropic effort AI risk strongly dominates nuclear risk reduction.

• Hi Johannes,

My intention is comparing quality-adjusted spending on decreasing nuclear and AI extinction risk, accounting for all sources (not just philanthropic ones).

• I consider the annual spending on decreasing extinction risk from AI is 50.6 (= 4.04*10^9/​(79.8*10^6)) times that on decreasing extinction risk from nuclear war. I determined this from the ratio between:

• 4.04 G\$ (4.04 billion USD) on nuclear risk in 2020, which I got from the mean of a lognormal distribution with 5th and 95th percentile equal to 1 and 10 G\$, corresponding to the lower and upper bound guessed in 80,000 Hours’ profile on nuclear war. “This issue is not as neglected as most other issues we prioritise. Current spending is between \$1 billion and \$10 billion per year (quality-adjusted)” (see details).

• 79.8 M\$ on “AI safety research that is focused on reducing risks from advanced AI” in 2023.

• I can’t open the GDoc on AI safety research.

But, in any case, I do not think this works, because philanthropic, private, and government dollars are not fungible, as all groups have different advantages and things they can and cannot do.

If looking at all resources, then 80M for AI safety research also seems an underestimate as this presumably does not include the safety and alignment work at companies?

• But, in any case, I do not think this works, because philanthropic, private, and government dollars are not fungible, as all groups have different advantages and things they can and cannot do.

I think I should be considering all sources of funding. Everything else equal, I expect a problem A which receives little philanthropic funding, but lots of funding from other sources, to be less pressing than a problem B which receives little funding from both philanthropic and non-philanthropic sources. The difference between A and B will not be as large as naively expected because philanthropic and non-philanthropic spending are not fungible. However, if one wants to define neglectedness as referring to just the spending from one source, then the scale should also depend on the source, and sources with less spending will be associated with a smaller fraction of the problem.

In general, I feel like the case for using the importance, tractability and neglectedness framework is stronger at the level of problems. Once one starts thinking about considerations within the cause area and increasingly narrow sets of interventions, I would say it is better to move towards cost-effectiveness analyses.

• So the nearterm annual extinction risk per annual spending for AI risk is 59.8 M (= 1.69*10^6*35.4) times that for nuclear risk.

Yet, given the above, I would say one should a priori expect efforts to decrease AI extinction risk to be more cost-effective at the current margin than ones to decrease nuclear extinction risk. Note: the sentence just above already includes the correction I will mention below.

I can’t open the GDoc on AI safety research.

Sorry! I have fixed the link now.

If looking at all resources, then 80M for AI safety research also seems an underestimate as this presumably does not include the safety and alignment work at companies?

It actually did not include spending from for-profit companies. I thought it included because I had seen they estimated just a few tens of millions of dollars coming from them:

I have now modified the relevant bullet in my analysis to the following:

• 114 M\$ (= (79.8 + 32 + 2*1)*10^6) on “AI safety research that is focused on reducing risks from advanced AI” in 2023:

• 79.8 M\$ from the National Science Foundation (NSF), LTFF, Open Philanthropy, SFF and “other”.

• “~\$32m per year” from “for-profit companies” (Anthropic, Conjecture, Google Deepmind and OpenAI).

• 2 times “probably at least \$1m per year” from “individual donors”.

My point remains qualitatively the same, as the spending on decreasing AI extinction risk only increased by 42.9 % (= 11479.8 − 1).

• (Last comment from me on this for time reasons)

• I think if you look at philanthropic neglectedness, the total sums across types of capital are not a good proxy. E.g., as far as I understand the nuclear risk landscape, it is both true that government spending is quite large but also that there is almost no civil society spending. This means that additional philanthropic funding should be expected to be quite effective on neglectedness grounds. Many obvious things are not done.

• The numbers on nuclear risk spending by 80k are entirely made up and not described otherwise (e.g. they do not cite a source and make no effort justifying the estimate, this is clearly a wild guess).

• If one constructed a similar number for AI risk, it could also be in the billions given it would presumably include stuff like the costs of government bureaucracies involved in tech regulation, emerging legislation etc.

I am fairly convinced your basic point will stand, but it seems important to not overplay the degree to which nuclear risk is not neglected, and to not underplay the degree to which government actors and others are now paying attention to AI risk (obviously, this also needs to be quality discounted, but this discounting does not reduce the value much for nuclear in your estimate).

• Thanks for elaborating.

I think if you look at philanthropic neglectedness, the total sums across types of capital are not a good proxy. E.g., as far as I understand the nuclear risk landscape, it is both true that government spending is quite large but also that there is almost no civil society spending. This means that additional philanthropic funding should be expected to be quite effective on neglectedness grounds.

I got this was your point, but I am not convinced it holds. I would be curious to understand which empirical evidence informs your views. Feel free to link to relevant pieces, but no worries if you do not want to engage further.

Many obvious things are not done.

I do not think this necessarily qualifies as satisfy empirical evidence that philanthropic neglectedness means high marginal returns. There may be non-obvious reasons for the ovious interventions not having been picked. In general, I am thinking that for any problem it is always possible to pick a neglected set of interventions, but that a priori we should assume diminishing returns in the overall spending, otherwise the government would fund the philanthropic interventions.

The numbers on nuclear risk spending by 80k are entirely made up and not described otherwise (e.g. they do not cite a source and make no effort justifying the estimate, this is clearly a wild guess).

For reference, here is some more context on 80,000 Hours’ profile:

### Who is working on this problem?

The area is a significant focus for governments, security agencies, and intergovernmental organisations.

Within the nuclear powers, some fraction of all work dedicated to foreign policy, diplomacy, military, and intelligence is directed at ensuring nuclear war does not occur. While it is hard to know exactly how much, it is likely to be in the billions of dollars or more in each country.

The US budget for nuclear weapons is comfortably in the tens of billions.8 Some significant fraction of this is presumably dedicated to control, safety, and accurate detection of attacks on the US.

In addition to this, some intergovernmental organisations devote substantial funding to nuclear security issues. For example, in 2016, the International Atomic Energy Agency had a budget of €361 million.9 Total philanthropic nuclear risk spending in 2021 was approximately \$57–190 million.

The spending of 4.04 G\$ I mentioned is just 4.87 % (= 4.04/​82.9) on the cost of maintaining and modernising nuclear weapons in 2022 of 82.9 G\$.

If one constructed a similar number for AI risk, it could also be in the billions given it would presumably include stuff like the costs of government bureaucracies involved in tech regulation, emerging legislation etc.

Good point. I guess the quality-adjusted contribution from those sources is currently small, but that it will become very significant in the next few years or decades.

I am fairly convinced your basic point will stand

Agreed. I estimated a difference of 8 OOMs (factor of 59.8 M) in the nearterm annual extinction risk per funding.

it seems important to not overplay the degree to which nuclear risk is not neglected, and to not underplay the degree to which government actors and others are now paying attention to AI risk (obviously, this also needs to be quality discounted, but this discounting does not reduce the value much for nuclear in your estimate).

Agreed. On the other hand, I would rather see discussions move from neglectedness towards cost-effectiveness analyses.

• but that a priori we should assume diminishing returns in the overall spending, otherwise the government would fund the philanthropic interventions.

I think this is fundamentally the crux—many of the most valuable philanthropic actions in domains with large government spending will likely be about challenging /​ advising /​ informationally lobbying the government in a way that governments cannot self-fund.

Indeed, when additional government funding does not reduce risk (does not reduce the importance of the problem) but is affectable, there can probably be cases where you should get more excited about philanthropic funding to leverage as public funding increases.