Concerning x-risks, my personal point of disagreement with the community is that I feel more skeptical of the chances to optimize our influence on the long-term future “in the dark” than what seems to be the norm. By “in the dark”, I mean in the absence of concrete short-term feedback loops. For instance, when I see the sort of things that MIRI is doing, my instinctive reaction is to want to roll my eyes (I’m not an AI specialist, but I work as a researcher in an academic field that is not too distant). The funny thing is that I can totally see myself from 10 years ago siding with “the optimists”, but with time I came to appreciate more the difficulty of making anything really happen. Because of this I feel more sympathetic to causes in which you can measure incremental progress, such as (but not restricted to) climate change.
Often times climate change is dismissed on the basis that there is already a lot of money going into this. But it’s not clear to me that this proves the point. For instance, it may well be that these large resources that are being deployed are poorly directed. Some effort to reallocate these resources could have a tremendously large effect. (E.g. supporting the Clean Air Task Force, as suggested by the Founders Pledge, may be of very high impact, especially in these times of heavy state intervention, and of coming elections in the US.) We should be careful to apply the “Importance-Neglectedness-Tractability” framework with caution. In the last analysis, what matters is the impact of our best possible action, which may not be small just on the basis of “there is already a lot of money going into this”. (And, for the record, I would personally rate AI safety technical research as having very low tractability, but I think it’s good that some people are working on it.)
Jc_Mourrat
How to find EA documents on a particular topic
Assumptions about the far future and cause priority
I had arrived at similar conclusions. Lately we were busy preparing material to put up on our freshly minted website at EA France. When it comes to recommendations, we naturally turn to charities such as GiveWell or ACE and refer to their work. But before putting these recommendations onto our website, I wanted to double-check that the evaluations are sound, because I want that we be ready to back them up publicly with confidence. So I tried to locate contradicting voices. For GiveWell, I found out that they disagree with other organizations such as the Campbell institute or Cochrane on the effectiveness of deworming programs. So I spent quite a bit of time reading the arguments of each party, and after that I came out extremely impressed by the depth and seriousness of the analysis of GiveWell. Of course I did not check all of what they do, but this experience gave me very high confidence that they are doing an outstanding work.
Then I moved to ACE. I took Nathan’s article as a starting point for the disagreeing voice. Of course I was appalled by some of the points raised there, in particular in relation with leafletting. Also, a friend at AEF dig up a Facebook thread that happened around the time of publication of this article, and my recollection is that half of the people discussing this where just busy explaining that Nathan was really a very very mean person that we could not possibly imagine talking to.
I understand that this is old news, but I want to pause for a moment and reflect on the bigger picture. On human poverty, GiveWell is one among several very serious actors. It engages very thoroughly in discussions and explanations when diverging views emerge. We can argue about whether Nathan was diplomatic enough, etc, certainly he did not take all the precautions that Halstead has taken when writing this piece. But we have to realize that when it comes to animal suffering, as far as I know ACE is the only game in town. In my opinion, this is a precarious state of affairs, and we should do our best to protect criticism of ACE, even when it does not come with the highest level of politeness. Of course, I do not mean that people at ACE have bad intentions, but checks and balances are important, we are all human beings, and right now there seems to be precious little of these.
And as it turns out, as pointed out here by Halstead, at least some of the criticism of Nathan was actually correct, and is now acknowledged on the website of ACE (e.g. on leafletting).
And then I must stay that I fell off of my chair when I looked up the evaluation of corporate outreach and found out that the single argument in support of cage-free reform was this paper by De Mol et al (2016). To give an element of context, I have no previous exposure to animal wellfare, and yet it jumped at me that this was very bogus. How can this possibly happen?? I know that now ACE changed their position about this, but how they could come up with that in the first place, and how on Earth can this argument be still available online completely eludes me. All this while, as far as I can tell, corporate outreach is one of the flagship interventions advocated by ACE.
But again, I want to pause and think about the bigger picture for a while. The fact is that at the time of writing this argument, the organisation Direct Action Everywhere (DxE) had put up a rather comprehensive report explaining that they had come up with the opposite conclusion! (That cage-free reform is actually detrimental to animal wellfare.) I will refrain from discussing it at length here because this comment is already long, but this report of DxE was, in my opinion, dismissed with precious little good argument.
So again I see a small dissenting voice in the otherwise rather monopolistic position of ACE which is being dismissed without due consideration. And of course I get more worried. (To be perfectly clear, my point has nothing to do with whether DxE’s conclusions were right or not; only with the fact that they were dismissed without proper consideration.)
I know that ACE no longer considers the article of De Mol et al (2016) as relevant, and things are clearly moving in a positive direction. Yet my confidence in ACE’s work is at an all-time low. For full disclosure, I was planning to propose to my colleagues at AEF that we spend some time doing a “case study” in relation with the recommendations of ACE (similarly to studying the controversy about deworming for GiveWell). The “case study” I have in mind is the comparative evaluations of the Good Food Institute vs. New Harvest (As a side remark, I am interested in any previous discussion about this point.)
To sum up, I want to stress that I write this with the best intentions, and I appreciate that ACE has been improving a lot on all the points I have raised so far. Also, I understand that we cannot “go meta” and ask for evaluators of evaluators, evaluators of evaluators of evaluators, etc. Yet it is my impression that when it comes to human poverty, the situation is much, much healthier. In this area, there are several influential actors that have their own decision processes, can then compare their conclusions, and engage in serious debate about them. For animal suffering, I think it would do us a lot of good to make sure that dissenting voices are heard and protected, and that ACE engages with their arguments with much greater consideration than has been the case so far.
As a side comment, I think it can make perfect sense to work on some area and donate to another one. The questions “what can I do with my money to have maximal impact” and “how do I use my skill set to have maximal impact” are very different, and I think it’s totally fine if the answers land on different “cause areas” (whatever that means).
Thanks for this interesting post! I particularly like your point that instant-runoff voting “has a track record in competitive elections and is much more in line with conventional notions of “majority””. Paraphrasing, the point is of course not to debate whether or not IRV produces a theoretically valid notion of majority; rather, it is about the psychological perception of the voting process and of the perceived legitimacy of the winner. I think these psychological aspects are very important, and are essentially impossible to capture by any theory.
Relatedly, I found this paragraph in Wikipedia’s article on approval voting, which I find worrisome: “Approval voting was used for Dartmouth Alumni Association elections for seats on the College Board of Trustees, but after some controversy, it was replaced with traditional runoff elections by an alumni vote of 82% to 18% in 2009.” My understanding is that voters in approval voting most often choose to support only one candidate, despite being given a much broader range of options; this, in elections with more than 2-3 possible candidates, often leads to a winner who collected only a small fraction of the votes, and who is then perceived as lacking legitimacy.
I have not studied the question in detail, but as of know my guess would be that instant-runoff voting should be preferred over approval voting.
Thanks for the link, that’s very interesting! I’ve seen that you direct donations to the Clean Energy Innovation program of the Information Technology and Innovation Foundation. How confident are you that the funds are actually fully used for the purpose? I understand that their accounting will show that all of your funds will go there, but how confident are you that they will not reduce their discretionary spending to this progam as a consequence? (I glanced at some of their recent work, and they have some pieces that are fairly confrontational towards China. While this may make sense from a short-term US perspective, it might even be net harmful if one takes a broader view and/or takes into account the possiblity of a military escalation between the US and China.) Did you consider the Clean Air Task Force when looking for giving opportunities?
For what it’s worth, if I could choose between this form existing or not existing, I would prefer that it exists. But we can also try to think about something in-between. Like:
(1) We agree in advance that there will be some clean-up of the form before release. We clarify what this means, I suppose that we will want to say that offensive or ad hominem content will be removed. Maybe we propose a list of made-up examples to explain what we want to be removed. This will be subject to some debate, but we can figure out something reasonable.
(2) We collect all the answers without disclosing them.
(3) We ask for a pool of people to volunteer for cleaning up the form.
(4) We select a small subset of these volunteers at random and they do the job. They check on each other for the cleaning, and then release the cleaned-up form.
I suppose that the simple fact of having this structure in place will already essentially clean up the form, whatever we mean by that.
I’ve just read the results of an interesting new study on the effect of red-flagging some information on social media, with flags such as “Multiple fact-checking journalists dispute the credibility of this news”, and variations with “Multiple fact-checking journalists” replaced by, alternatively, “Major news outlets”, “A majority of Americans”, or “Computer algorithms using AI”. The researchers tested the effect this had on the propensity of people to share the content. The effect of the “fact-checking” phrasing was the most pronounced, and very significant (a reduction of about 40% of the probability to share content; which jumps to 60% for people who identify as Democrats). Overall the effect of the “AI” phrasing was also very significant, but quite counterintuitively it has the effect of increasing the probability of sharing content for people who identify as Republicans! (By about 8%; it decreases that same probability by 40% for people who identify as Democrats.)
https://engineering.nyu.edu/news/researchers-find-red-flagging-misinformation-could-slow-spread-fake-news-social-media
Ok, I understand your point better now, and find that it makes sense. To summarize, I believe that the art of good planning to a distant goal is to find a series of intermediate targets that we can focus on, one after the other. I was worried that your argument could be used against any such strategy. But in fact your point is that as it stands, health interventions have not been selected by a “planner” who was actually thinking about the long-term goals, so it is unlikely that the selected interventions are the best we can find. That sounds reasonable to me. I would really like to see more research into what optimizing for long-term growth could look like (and what kind of “intermediate targets” this would select). (There is some of this in Christiano’s post, but there is clearly room for more in-depth analysis in my opinion.)
Thanks a lot, this is super helpful! I particularly appreciated that you took the time to explain the internal workings of a typical think tank, this was not at all clear to me.
Thanks for your detailed and kind comments! It’s true that naming this a “plateau” is not very accurate. It was my attempt to make the reader’s life a bit easier by using a notion that is relatively easier to grasp in the main text (with some math details in a footnote for those who want more precision). About the growth rate, mathematically a function is fully described by its growth rate (and initial condition), and here the crux is whether or not the growth rate will go to zero relatively quickly, so it seems like a useful concept to me.
(When you refer to footnote 15, that can make sense, but I wonder if you were meaning footnote 5 instead.)
I agree with all the other things you say. I may be overly worried about our community becoming more and more focused on one particular cause area, possibly because of a handful of disappointing personal experiences. One of the main goals of this post was to make people more aware of the fact that current recommendations are based in an important way on a certain belief on the trajectory of the far future, and maybe I should have focused on that goal only instead of trying to do several things at once and not doing them all very well :-)
As a first pass the rate of improvement should asymptote towards zero so long as there’s a theoretical optimum and declining returns to further research before the heat death of the universe, which seems like pretty mild assumptions.
I think it would be really useful if this idea was explained in more details somewhere, preferably on the 80k website. Do you think there is a chance that this happens at some point? (hopefully not too far in the future ;-) )
I have two comments concerning your arguments against accelerating growth in poor countries. One is more “inside view”, the other is more “outside view”.
The “inside view” point is that Christiano’s estimate only takes into account the “price of a life saved”. But in truth GiveWell’s recommendations for bednets or deworming are to a large measure driven by their belief, backed by some empirical evidence, that children who grow up free of worms or malaria become adults who can lead more productive lives. This may lead to better returns than what his calculations suggest. (Micronutrient supplementation may also be quite efficient in this respect.)
The “outside view” point is that I find our epistemology really shaky and worrisome. Let me transpose the question into AI safety to illustrate that the point is not related to growth interventions. If I want to make progress on AI safety, maybe I can try directly to “solve AI alignment”. Let’s say that I hesitate between this and trying to improve the reliability of current-day AI algorithms. I feel that, at least in casual conversations (perhaps especially from people who are not actually working in the area), people would be all too willing to jump to “of course the first option is much better because this is the real problem, if it succeeds we win”. But in truth there is a tradeoff with being able to make any progress at all, it is not just better to turn your attention to the most maximally long-term thing you can think of. And, I think it is extremely useful to have some feedback loop that allows you to track what you are doing, and by necessity this feedback loop will be somewhat short term. To summarize, I believe that there is a “sweet spot” where you choose to focus on things that seem to point in the right direction and also allow you at least some modicum of feedback over shorter time scales.
Now, consider the argument “this intervention cannot be optimal in the long run because it has been optimized for the short term”. This argument essentially allows you to reject any intervention that has shown great promise based on the observations we can gather. So, effective altruism started as being “evidence based” etc., and now we reached a situation where we have built a theoretical construct that, not only allows to place certain interventions above all others without us having to give any empirical evidence backing this, but moreover, if another intervention is proposed that comes with good empirical backing, we can use this fact as an argument against the intervention!
I may be pushing the argument a bit too far. This still makes me feel very uncomfortable.
You’re right that these are indeed important considerations that I swept under the rug… Thanks again for all the references.
Thanks a lot, this looks all very useful. I found these texts by Tomasik and Baumann particularly interesting, and was not aware of them.
When I search for capacity building as suggested in the post, it seems to me that the results are ok: I get about 600 results, and the first result is actually the page you mentioned. It’s surprising that one cannot get the custom search engine to reproduce this. (And it’s good to know that we can get around this by just generating Google queries as in the link above!) Thanks for pointing out searchstack.co, it looks very interesting!
Let me call X the statement: “our rate of improvement remains bounded away from zero far into the future”. If I understand correctly, you are saying that we have great difficulties imagining a scenario where X happens, therefore X is very unlikely.
Human imagination is very limited. For instance, most of human history shows very little change from one generation to the next; in other words, people were not able to imagine ways for future generations to do certain things in better ways than how they already knew. Here you ask our imagination to perform a spectacularly difficult task, namely to imagine what extremely advanced civilizations are likely to be doing in billions of years. I am not surprised if we do not manage to produce a credible scenario where X occurs. I do not take this as strong evidence against X.
Separately from this, I personally do not find it very likely that we will ultimately settle most of the accessible universe, as you suppose, because I would be surprised if human beings hold such a special position. (In my opinion, either advanced civilizations are not so interested in expanding in space; or else, we will at some point meet a much more advanced civilization, and our trajectory after this point will probably depend little on what we can do before it.)
Concerning the point you put in parentheses about safety being “infinitely” preferred, I meant to use phrases such as “virtually infinitely preferred” to convey that the preference is so strong that any actual empirical estimate is considered unnecessary. In footnote 5 above, I mentioned this 80k article intended to summarize the views of the EA community, where it is said that speedup interventions are “essentially morally neutral” (which, given the context, I take as being equivalent to saying that risk mitigation is essentially infinitely preferred).
Thanks a lot for your very interesting work. While I am very sympathetic to the views you expressed here, I want to play the devil’s advocate for a moment and try to explore some counter-arguments.
Like Sanjay I think it would be desirable that you explain more why fears of morphine dependence and misuse are unwarranted. The article you linked to as a response to Sanjay argues that restricting the prescription of opioids in the US is counterproductive. This article is not very strongly convincing to me, in part because of lack of sources, and in part because the intervention context is really very different. I want to note that problems with opioids are observed in some African countries, see https://www.economist.com/the-economist-explains/2018/08/23/west-africas-opioid-crisis . Also, even if the probability that this intervention turns out to be detrimental in some important way is small, the negative consequences it would have if this were the case could be rather devastating, especially if the cause is tightly associated with the EA movement. Finally, this is much more speculative, but I was also wondering how “flow-through”, more long-term effects of the intervention would compare with other interventions. (E.g. the longterm positive effects could be smaller if most people need pain relief in the last few years of their lives.)
After all that criticizing of mine, I want to end by expressing again my sincere appreciation for your work. I view my critique not as a way to discourage people to work in this direction, but as a modest attempt at trying to help adjust as best as we can what we can do there.
The impression I get from a (I admit relatively casual) look is that you are saying something along the following lines:
1) there is a big mystery concerning the fact that the rate of growth has been accelerating,
2) you will introduce a novel tool to explain that fact, which is stochastic calculus,
3) using this tool, you arrive at the conclusion that infinite explosion will occur before 2047 with 50% probability.
For starters, as you point out if we read you sufficiently carefully, there is no big mystery in the fact that the rate of growth of humanity has been super-exponential. This can be simply explained by assuming that innovation is an important component of the growth rate, and the amount of innovation effort itself is not constant, but grows with the size of the population, maybe in proportion to this size. So if you decide that this is your model of the world, and that the growth rate is proportional to innovation effort, then you write down some simple math and you conclude that infinite explosion will occur at some point in the near future. This has been pointed numerous times. For instance, as you point out (if we read you carefully), Michael Kremer (1993) checked that, going back as far as a million years ago, the idea that population growth rate is roughly proportional to (some positive power of the) population size gives you a good fit with the data up to maybe a couple of centuries ago. And then we know that the model stops to work, because for some reason at some level of income people stop to transform economic advancement into having more children. I don’t think we should ponder for long about the fact that a model that matched well past data stopped to work at some point. This seems to me to be the natural fate of models of early growth of anything. So instead of speculating about this, Kremer adjusts his model to make it more realistic.
It is of course legitimate to argue that human progress over recent times is not best captured by population size, and that maybe gross world product is a better measure. For this measure, we have less direct evidence that a slowdown of the “naive model” is coming (By “naive model” I mean the model in which you just fit growth with a power law, without any further adjustment). Altough I do find works such as this or this quite convincing that future trends will be slower than what the “naive” model would say.
After reading a (very small) bit of your technical paper, my sense is that your main contribution is that you fixed a small inconsistency in how we go about estimating the parameters of the “naive model”. I don’t deny that this is a useful technical contribution, but I believe that this is what it is: a technical contribution. I don’t think that it brings any new insight into questions such as, for instance, whether or not there will indeed be a near-infinite explosion of human development in the near future.
I am not comfortable with the fact that, in order to convey the idea of introducing randomness into the “naive model”, you invoke “E = mc2″, the introduction of calculus by Newton and Leibnitz, the work of Nobel prize winners, or the fact that “you experienced something like what [this Nobel prize winner] experienced, except for the bits about winning a Nobel”. Introducing some randomness into a model is, in my opinion, a relatively common thing to do. That is, once we have a deterministic model that we find relatively plausible and that we want to refine somewhat.