On your general point about paying attention to political biases, I agree that’s worthwhile. A quibble related to that which might matter to you: the Wikipedia article you’re quoting seems to attribute the incorrect predictions to TTAPS but I could only trace them to Sagan specifically. I could be missing something due to dead/inaccessible links.
Kit
Karma: 1,382
There are a whole bunch of things I love about this work. Among other things:
An end-to-end model of nuclear winter risk! I’m really excited about this.
The quantitative discussions of many details and how they interact are very insightful. e.g. ones which were novel for me included how exactly smoke causes agriculture loss, and roughly where the critical thresholds for agricultural collapse might be. The concrete estimates for the difference in smoke production between counterforce and countervalue, which I knew the sign of but not the magnitude, are fascinating and make this much clearer.
I really appreciate the efforts to make the (huge) uncertainty transparent, notably the list of simplifying assumptions, and running specific scenarios for heavy countervalue targeting. Most of all, though, the Guesstimate model is remarkably legible, which makes absorbing all this info so much easier.
I have one material issue with the model structure, which I think may reverse your bottom line. The scenario full-scale countervalue attack against Russia has a median smoke estimate of 60Tg and a scenario probability of 0.27 x 0.36 = ~0.1. This means the probability of total smoke exceeding 60Tg has to be >5%, but Total smoke generated by a US-Russia nuclear exchange calculates a probability of only 0.35% for >60Tg smoke.
What seems to be going on is that the model incorporates estimated smoke from each countervalue targeting scenario as {scenario probability x scenario amount of smoke} in all Monte Carlo samples, when I figure you actually want it to count {scenario amount of smoke} in the appropriate proportion of samples. This would give a much more skewed distribution.
Sampling properly (as I see it) seems to be a bit fiddly in Guesstimate, but I put something together for Smoke that would be generated as a result of countervalue targeting against the US in an ‘Alternative 3.4.4’ section here. (I figured making a copy would be the easiest way to communicate the idea.)
I also redirected the top-level smoke calculation to point to the above alternate calculation to see what difference it makes. (Things I’ve added are marked with [KH] in the copy to make the differences easy to spot.) Basically every distribution now has two humps: either there is a countervalue strike and everything has a high chance of collapsing, or there isn’t and things are awful but probably recoverable. Some notable conclusions that change:
~15% chance of getting into the 50Tg+ scenarios that you flag as particularly concerning, up from ~1%.
~13% chance that corn cultivation becomes impossible in Iowa, and 6% chance that Ukraine cannot grow any of the crops you focus on, both from <1%. I don’t know whether still being able to grow some amount of barley helps much.
Your bottom-line ~5% chance of 96% population collapse jumps to ~16%, with most of that on >99% collapse. On the bright side, expected deaths drop by ~1bn.
Obviously, all these numbers are hugely unstable. I list them only to illustrate the difference made by sampling in this way, not to suggest that the actual numbers should be taken super seriously.
As above, these changes are just from adjusting the sampling for Smoke that would be generated as a result of countervalue targeting against the US. Doing the same adjustment for Smoke that would be generated as a result of countervalue targeting against Russia would add additional risk of extreme nuclear winter. For example, I think your model would imply a few % chance of all the crops you focus on becoming impossible to grow in both Iowa and Ukraine.
Before exploring your work, I hadn’t understood just how heavily extinction risk may be driven by the probability of a full-scale countervalue strike occurring. This certainly makes me wonder whether there’s anything one can do to specifically reduce the risk of such strikes without too significantly increasing the overall risk of an exchange. In general, working through your model and associated text and sources has been super useful to my understanding.
Neat. Happy to be a little bit helpful!
Agreed. The discussion of the likelihood of countervalue targetting throughout this piece seems very important if countervalue strikes would typically produce considerably more soot than counterforce strikes. In particular, the idea that any countervalue component of a second strike would likely be small seems important and is new to me.
I really hope the post is right that any countervalue targetting is moderately unlikely even in a second strike for the countries with the largest arsenals. That one ‘point blank’ line in the 2010 NPR was certainly surprising to me. On the other hand, I’m not compelled by most of the arguments as applied to second strikes specifically.
This is fascinating, especially with details like different survivability of US and Russian SLBMs. My main takeaway is that counterforce is really not that effective, so it remains hard to see why it would be worth engaging in a first strike. I’d be interested to hear if you ever attempt to quantify the risk that cyber, hypersonic, drone and other technologies (appear to) change this, or if this has been attempted by someone already.
Relatedly:
If improvements in technology allowed either country to reliably locate and destroy those targets, they would be able to eliminate the others’ secure second strike, thereby limiting the degree to which a nuclear war could escalate.
Perhaps reading into this too much, but I wondered if you think the development of some kinds of effective counterforce are net positive in expectation from an extinction risk perspective. My amateur impression is that these developments are kind of all bad (most prominently because the ability to destroy weapons seems to force ‘launch on warning’ to be the default, making accidental escalation (from zero) more likely), but I’m potentially generalising too much.
Quibbles/queries:
The one significant thing I was confused about was why the upper bound survivability for stationary, land-based ICBMs is only 25%. It looks like these estimates are specifically for cases where a rapid second strike (which could theoretically achieve survivability of up to 100%) is not attempted. Do you intend to be taking a position on whether a rapid second strike is likely? It seems like you are using these numbers in some places, e.g. when talking about ‘Countervalue targeting by Russia in the US’ in your third post, when you might be using significantly larger numbers if you thought a rapid second strike was likely. The reason I’m interested in this question is that it seems likely to feed into your planned research into nuclear winter, which I particularly look forward to.
Also, maybe you intend for your adjustment for US missile defence systems to be negating 15% of the lost warheads rather than adding 15% to the total arsenal? The current calculation suggests that missile defences reduce counterforce effectiveness by ~61%, which seems like not your intention given what you’ve said about interceptor success rates and diminishing returns on a counterforce strike. (I think this change would decrease surviving, deployed US warheads by ~163, so possibly has moderately large implications for your later work.)
This series (#2, #3) has begun as the most interesting-to-me on the Forum in a long time. Thanks very much. If you have written or do write about how future changes in arsenals may change your conclusions about what scenarios to pay the most attention to, I’d be interested in hearing about it.
In case relevant to others, I found your spreadsheet with raw figures more insightful than the discrete system in the post. To what extent do you think the survey you use for the probabilities of particular nuclear scenarios is a reliable source? (I previously distrusted it for heuristic reasons like the authors seeming to hype some results that didn’t seem that meaningful.) I’m interested because, as well as the numbers you use it for, the survey implies ~15% chance of use of nuclear weapons conditional on a conventional conflict occurring between nuclear-armed states, which seemed surprisingly low to me and would change my thinking about conflicts between great powers in general if I believed it.
effect from boosting efficacy of current long-termist labor + effect from increasing the amount of long-termist labor
Let’s go. Upside 1:
effect from boosting efficacy of current long-termist labor
Adding optimistic numbers to what I already said:
Let’s say EAs contribute $50m† of resources per successful drug being rolled out across most of the US (mainly contributing to research and advocacy). We ignore costs paid by everyone else.
This somehow causes rollout about 3 years earlier than it would otherwise have happened, and doesn’t trade off against the rollout of any other important drug.
At any one time, about 100 EAs†† use the now-well-understood, legal drug, and their baseline productivity is average for long-term-focused EAs.
This improves their productivity by an expected 5%††† vs alternative mental health treatment.
Bottom line: your $50m buys you about 100 x 5% x 3 = 15 extra EA-years via this mechanism, at a price of $3.3m per person-year.
Suppose we would trade off $300k for the average person-year††††. This gives a return on investment of about $300k/$3.3m = 0.09x. Even with optimistic numbers, upside 1 justifies a small fraction of the cost, and with midline estimates and model errors I’d expect more like a ~0.001x multiplier. Thus, this part of the argument is insignificant.
-----
Also, I’ve decided to just reply to this thread, because it’s the only one that seems decision-relevant.
† Various estimates of the cost of introducing a drug here, with a 2014 estimate being $2.4bn. I guess EAs could only cover the early stages, with much of the rest being picked up by drug companies or something.
†† Very, very optimistically, 1,000 long-term-focused EAs in the US, 10% of the population suffer from relevant mental health issues, and all of them use the new drug.
††† This looks really high but what do I know.
†††† Pretty made up but don’t think it’s too low. Yes, sometimes years are worth more, but we’re looking at the whole population, not just senior staff.
Psychedelic interventions seem promising because they can plausibly increase the number of capable people focused on long-termist work, in addition to plausibly boosting the efficacy of those already involved.
Pointing out that there are two upsides is helpful, but I had just made this claim:
The math for [the bold part] seems really unlikely to work out.
It would be helpful if you could agree with or contest with that claim before we move on to the other upside.
-
Rationality projects: I don’t care to arbitrate what counts as EA. I’m going to steer clear of present-day statements about specific orgs, but you can see my donation record from when I was a trader on my LinkedIn profile.
I’m not arguing against trying to compare things. I was saying that the comparison wasn’t informative. Comparing dissimilar effects is valuable when done well, but comparing d-values of different effects from different interventions tells you very little.
To explicitly separate out two issues that seem to be getting conflated:
Long-term-focused EAs should make use of the best mental health care available, which would make them more effective.
Some long-term-focused EAs should invest in making mental health care better, so that other long-term-focused EAs can have better mental health care and be more effective.
The former seems very likely true.
The latter seems very likely false. You would need the additional cost of researching, advocating for and implementing a specific new treatment (here, psilocybin) across some entire geography to be justified by the expected improvement in mental health care (above what already exists) for specifically long-term-focused EAs in that geography (<0.001% of the population). The math for that seems really unlikely to work out.
I continue to focus on the claims about this being a good long-term-focused intervention because that’s what is most relevant to me.
-----
Non-central notes:
We’ve jumped from emotional blocks & unhelpful personal narratives to life satisfaction & treatment-resistant depression, which are very different.
As you note, the two effects you’re now comparing (life satisfaction & treatment-resistant depression) aren’t really the same at all.
I don’t think that straightforwardly comparing two Cohen’s d measurements is particularly meaningful when comparing across effect types.
I believe you when you say that psychedelic experiences have an effect of some (unknown) size on emotional blocks & unhelpful personal narratives, and that this would change workers’ effectiveness by some (unknown) amount. However, even assuming that the unknown quantities are probably positive, this doesn’t tell me whether to prioritise it any more than my priors suggest, or whether it beats rationality training.
Nonetheless, I think your arguments should be either compelling or something of a wake-up call for some readers. For example, if a reader does not require careful, quantified arguments to justify their favoured cause area†, they should also not require careful, quantified arguments about other things (including psychedelics).
† For example, but by no means exclusively, rationality training.
[Edited for kindness while keeping the meaning the same.]
Boring answer warning!
The best argument against most things being ‘an EA cause area’† is simply that there is insufficient evidence in favour of the thing being a top priority.
I think future generations probably matter morally, so the information in sections 3(a), 3(b) and 4 matter most to me. I don’t see the information in 3(a) or 3(b) telling me much about how leveraged any particular intervention is. There is info about what a causal mechanism might be, but analysis of the strength is also needed. (For example, you say that psychedelic interventions are plausibly in the same ballpark of effectiveness of other interventions that increase the set of well-intentioned + capable people. I only agree with this because you use the word ‘plausibly’, and plausibly...in the same ballpark isn’t enough to make something an EA cause area.) I think similarly about previous discussion I’ve seen about the sign and magnitude of psychedelic interventions on the long-term future. (I’m also pretty sceptical of some of the narrower claims about psychedelics causing self-improvement.††)
I did appreciate your coverage in section 4 of the currently small amount of funding and what is getting done as a result, which seems like it could form part of a more thorough analysis.†††
My amateur impression is that Michael Plant has made a decent start on quantifying near-term effects, though I don’t think anyone should take my opinion on that very seriously. Regardless of that start looking good, I would be unsurprised if most people who put less weight on future generations than me still wanted a more thorough analysis before directing their careers towards the cause.
As I said, it’s a boring answer, but it’s still my true objection to prioritising this area. I also think negative PR is a material consideration, but I figured someone else will cover that.
-----
† Here I’m assuming that ‘psychedelics being an EA cause area’ would eventually involve effort on a similar scale to the areas you’re directly comparing it to, such as global health (say ~100 EAs contributing to it, ~$10m in annual donations by EA-aligned people). If you weaken ‘EA cause area’ to mean ‘someone should explore this’, then my argument doesn’t work, but the question would then be much less interesting.
†† I think mostly this comes from me being pretty sceptical of claims of self-improvement which don’t have fairly solid scientific backing. (e.g. I do deep breathing because I believe that the evidence base is good, but I think most self-improvement stuff is random noise.) I think that the most important drivers of my intuitions for how to handle weakly-evidenced claims have been my general mathematical background, a few week-equivalents trying to understand GiveWell’s work, this article on the optimiser’s curse, and an attempt to simulate the curse to get a sense of its power. Weirdness aversion and social stuff may be incorrectly biasing me, but e.g. I bought into a lot of the weirdest arguments around transformative AI before my friends at the time did, so I’m not too worried about that.
††† I also appreciated the prize incentive, without which I might not have written this comment.
As an aside, I wouldn’t say that any Good Ventures things are ‘housed under Open Phil’. I’d rather say that Open Phil makes recommendations to Good Ventures. i.e. Open Phil is a partner to Good Ventures, not a subsidiary.
Technically, I’ve therefore answered a different question to the one you asked: I’ve answered the question ‘why aren’t these grants on the Open Phil website’.
From Good Ventures’ grantmaking approach page:
In 2018, Good Ventures funded $164 million in grants recommended by the Open Philanthropy Project, including $74 million to GiveWell’s top charities, standout charities, and incubation grants. (These grants generally appear in both the Good Ventures and Open Philanthropy Project grants databases.)
Good Ventures makes a small number of grants in additional areas of interest to the foundation. Such grants totaled around $19 million in 2018. Check out Our Portfolio and Grants Database to learn more about the grants we’ve made so far.
I figured the OP was suggesting that people go to the retreat? (or maybe be generically supportive of the broader project of running retreats)
Not sure where this is going; doesn’t immediately seem like it counters what I said about your comparison to specific fundraising + analysis posts, or about why readers might be confused as to why this is here.
Right. The stuff about psychedelics as Cause X was maybe a bit of a red herring. You probably know how to sell your business much better than I do, but something which I think is undervalued in general is simply opening your pitch with why exactly you think someone should care about your thing. I actually hadn’t considered creative problem-solving or career choice as reasons to go on this retreat.
My earlier comment was a reply to the challenge of ‘how this post is substantively different from previous content like...’ and this now seems fairly obvious, so I probably have little more useful to say :)
I can see where you’re coming from, but I think there’s a lot of missing info here, and this will make the post confusing to most readers. Some* of the other posts you link to also ask things of their readers, but they also present a case for why that ask is a particularly exceptional use of resources.
I happen to know of some topics which psychedelics might be relevant to, some of which are mentioned in the post and in your later comment, e.g.
Potentially strong treatment for depression
Drug liberalisation could reduce unnecessary incarceration
Very speculative things like maybe psychedelics make you a better or more effective person (or increases your risk of psychosis), or maybe psychedelics could help us study sentience
but it’s pretty unclear how EAs going on a psychedelic retreat is an effective way to make progress in these fields. i.e. even with what I guess is an above-median amount of context on the subject, I don’t know what the case is. Given that, I think Khorton’s reaction is very reasonable.
Maybe I’m missing the point and the post is just saying that there’s a cool thing you can do with other EAs, not trying to claim that it’s an effectively altruistic use of resources. In that case, the difference between the posts appears to be simple.
A disclosure of my own: I previously worked for CEA. Of course, these are my opinions only.
*Giving What We Can is still growing at a surprisingly good pace doesn’t justify an ask, but it doesn’t have an ask either.
On the specific questions you’re asking about whether empirical data from the Kuwaiti oil field destruction is taken into account: it seems that the answer to each is simply ‘yes’. The post says that the data used is adapted from Toon et al. (2007), which projects how much smoke would reach the stratosphere specifically. The paper explicitly considers that event and what the model would predict about them:
The details of the paper could be wrong – I’m a complete amateur and would be interested to hear the views of people who’ve looked into it, especially given substantial reliance on this particular paper in the post – but it seems to have already considered the things you raise.
However, this still got me thinking. Why look at smoke from burning oil fields, with their much lower yields, when one could look at smoke from Hiroshima or Nagasaki? It’s a grim topic, but more relevant for projecting the effects of other nuclear detonations. After a surprisingly long search, I found this paper, which attempts to measure the height of the ‘mushroom cloud’ over Hiroshima, which isn’t what we’re looking for. Fortunately for me, they seem to think that Photo ‘(a) Around Kurahashi Island’ is another photo of the ‘mushroom cloud’, but in fact it appears* to be the cloud produced by the resulting fires. This explains their surprising result:
16km (range 14.54-16.88km) is well into the stratosphere across Russia and most of the US, so it seems that history is compatible with theories which say that weapons on the scale of ‘Little Boy’ (13–18kt) are likely to cause substantial smoke in the stratosphere.
[*update 17-Sep-2021: various people familiar with nuclear weapons agree that photo (a) is of the smoke from the firestorm, not the ‘mushroom cloud’.]