This position seems unlikely to me at face value. It relies on a very long list of claims, and given the apparently massive improbability of the conjunction, there is no way this consideration is going to be the biggest impact of economic progress:
The most important determinant of future welfare is whether you get FAI or UFAI (this presupposes a relatively detailed model of how AI works, of what the danger looks like, etc.)
This will happen quite soon, and relevant AI work is already underway.
The main determinant of FAI vs. UFAI is whether an appropriate theoretical framework for goal-stability is in place.
As compared to UFAI work, the main difficulty for developing such a framework for goal-stability is the serial depth of the problem.
A 1% boost in economic activity this year has a non-negligible effect on the degree of parallelization of relevant AI work.
I don’t see how you can defend giving any of those points more than 1⁄2 probability, and I would give the conjunction less than 1% probability. Moreover, even in this scenario, the negative effect from economic progress is quite small. (Perhaps a 1% increase in sustained economic productivity makes the future 0.1% worse if this story is true, and each year of a 1% increase in economic productivity makes the future 0.002% worse?)
So on balance it seems to me like this would say that a 1% increase in economic productivity would make the future 0.00002% worse? That is a ridiculously tiny effect; even if you couldn’t see any particular reason that economic progress helped or hurt I think your prior should expect other effects to dominate, and you can use other considerations to get a handle on the sign. I guess you think that this is an underestimate, but I would be interested to know where you disagree.
I would guess that the positive effect from decreasing the cumulative risk of war alone are several orders of magnitude higher than that. I know you think that a world war that killed nearly everyone might be positive rather than negative, but in that case slowing down economic progress would still have positive effects that are orders of magnitude larger than the effect on FAI parallelization.
Even without doing any calculations, it is extraordinarily hard to imagine that the difference between “world at war” and “world at peace” is less than the difference between “world with slightly more parallelization in AI work” and “world with slightly less parallelization;” almost everyone would disagree with you, and as far as I can tell their reasons seem better than yours. Similarly, it is hard to imagine that the interaction between generational turnover and economic activity wouldn’t swamp this by several orders of magnitude.
General remark: At some point I need to write a post about how I’m worried that there’s an “unpacking fallacy” or “conjunction fallacy fallacy” practiced by people who have heard about the conjunction fallacy but don’t realize how easy it is to take any event, including events which have already happened, and make it look very improbable by turning one pathway to it into a large series of conjunctions. E.g. I could produce a long list of things which allegedly have to happen for a moon landing to occur, some of which turned out to not be necessary but would look plausible if added to the list ante facto, with no disjunctive paths to the same destination, and thereby make it look impossible. Generally this manifests when somebody writes a list of alleged conjunctive necessities, and I look over the list and some of the items seem unnecessary (my model doesn’t go through them at all), obvious disjunctive paths have been omitted, the person has assigned sub-50% probability to things that I see as mainline 90% probabilities, and conditional probabilities when you assume the theory was right about 1-N would be significantly higher for N+1. Most of all, if you imagine taking the negation of the assertion and unpacking it into a long list of conjunctive probabilities, it would look worse—there should be a name for the problem of showing that X has weak arguments but not considering that ~X has even weaker arguments. Or on a meta level, since it is very easy to make things look more conjunctive, we should perhaps not be prejudiced against things which somebody has helpfully unpacked for us into a big conjunction, when the core argument still seems pretty simple on some level.
When I look over this list, my reaction is that:
(1) is a mainline assumption with odds of 5:1 or 10:1 - of course future intergalactic civilization bottlenecks through the goals of a self-improving agency, how would you get to an intergalactic civilization without that happening? If this accounts for much of our disagreement then we’re thinking about entirely different scenarios, and I’m not sure how to update from your beliefs about mostly scenario B to my beliefs about mostly scenario A. It makes more sense to call (1) into question if we’re really asking about global vs. local, but then we get into the issue of whether global scenarios are mostly automatic losses anyway. If (1) is really about whether we should be taking into account a big chunk of survivable global scenarios then this goes back to a previous persistent disagreement.
(2) I don’t see the relevance—why does a long time horizon vs. a short time horizon matter? 80 years would not make me relax and say that we had enough serial depth, though it would certainly be good news ceteris paribus, there’s no obvious threshold to cross.
Listing (3) and (4) as separate items was what originally made my brain shout “unpacking fallacy!” There are several subproblems involved in FAI vs. UFAI, of which the two obvious top items are the entire system being conducive to goal stability through self-improvement which may require deducing global properties to which all subsystems must be conducive, and the goal loading problem. These both seem insight-heavy which will require serial time to solve. The key hypothesis is just that there are insight-heavy problems in FAI which don’t parallelize well relative to the wide space of cobbled-together designs which might succeed for UFAI. Odds here are less extreme than for (1) but still in the range of 2:1-4:1. The combined 3-4 issue is the main weak point, but the case for “FAI would parallelize better than UFAI” is even weaker.
(5) makes no sense to ask as a conditionally independent question separate from (1); if (1) is true then the only astronomical effects of modern-day economic growth are whatever effects that growth has on AI work, and to determine if economic growth is qualitatively good or bad, we ask about the sign of the effect neglecting its magnitude. I suppose if the effect were trivial enough then we could just increase the planet’s growth rate by 5% for sheer fun and giggles and it would have no effect on AI work, but this seems very unlikely; a wealthier planet will ceteris paribus have more AI researchers. Odds of 10:1 or better.
On net, this says that in my visualization the big question is just “Does UFAI parallelize better than FAI, or does FAI parallelize better than UFAI?” and we find that the case for the second clause is weaker than the first; or equivalently “Does UFAI inherently require serial time more than FAI requires serial time?” is weaker than “Does FAI inherently require serial time more than UFAI requires serial time?” This seems like a reasonable epistemic state to me.
The resulting shove at the balance of the sign of the effect of economic growth would have to be counterbalanced by some sort of stronger shove in the direction of modern-day economic growth having astronomical benefits. And the case for e.g. “More econ growth means friendlier international relations and so they endorse ideal Y which leads them to agree with me on policy Z” seems even more implausible when unpacked into a series of conjunctions. Lots of wealthy people and relatively friendly nations right now are not endorsing policy Z.
To summarize and simplify the whole idea, the notion is:
Right now my estimate of the sign of the astronomical effect of modern-day economic growth is dominated by a 2-node conjunction of, “Modern-day econ growth has a positive effect on resources into both FAI and UFAI” and “The case for FAI parallelizing better than UFAI is weaker than the converse case”. For this to be not true requires mainly that somebody else demonstrate an effect or set of effects in the opposite direction which has better net properties after its own conjunctions are taken into account. The main weakness in the argument and lingering hope that econ growth is good, isn’t that the original argument is very conjunctive, but rather it’s that faster econ growth seems like it should have a bunch of nice effects on nice things and so the disjunction of other econ effects might conceivably swing the sign the other way. But it would be nice to have at least one plausible such good effect without dropping our standards so low that we could as easily list a dozen equally (im)plausible bad effects.
Even without doing any calculations, it is extraordinarily hard to imagine that the difference between “world at war” and “world at peace” is less than the difference between “world with slightly more parallelization in AI work” and “world with slightly less parallelization;”
With small enough values of ‘slightly’ obviously the former will have a greater effect, the question is the sign of that effect; also it’s not obvious to me that moderately lower amounts of econ growth lead to world wars, and war seems qualitatively different in many respects from poverty. I also have to ask if you are possibly maybe being distracted by the travails of one planet as a terminal value, rather than considering that planet’s instrumental role in future galaxies.
At some point I need to write a post about how I’m worried that there’s an “unpacking fallacy” or “conjunction fallacy fallacy” practiced by people who have heard about the conjunction fallacy but don’t realize how easy it is to take any event, including events which have already happened, and make it look very improbable by turning one pathway to it into a large series of conjunctions.
Luke asked me to look into this literature for a few hours. Here’s what I found.
The original paper (Tversky and Koehler 1994) is about disjunctions, and how unpacking them raises people’s estimate of the probaility. So for example, asking people to estimate the probability someone died of “heart disease, cancer, or other natural causes” yields a higher probability estimate than if you just ask about “natural causes.”
They consider the hypothesis this might be because they take the researcher’s apparent emphasis as evidence that’s it’s more likely, but they tested & disconfirmed this hypothesis by telling people to take the last digit of their phone number and estimate the percentage of couples that have that many children. Percentages sum to greater than 1.
Finally, they check whether experts are vulnerable to this bias by doing an experiment similar to the first experiment, but using physicians at Stanford University as the subjects and asking them about a hypothetical case of a woman admitted to an emergency room. They confirmed that yes, experts are vulnerable to this mistake too.
This phenomenon is known as “subadditivity.” A subsequent study (RottenStreich and Tversky 1997) found that subadditivity can even occur when dealing with explicit conjunctions. Macci et al. (1999) found evidence of superadditivity: ask some people how probable it is that the freezing point of alcohol is below that of gasoline, other people how probable it is that the freezing point of gasoline is below that of alcohol, average answers sum to less than 1.
Other studies try to refine the mathematical model of how people make judgements in these kinds of cases, but the experiments I’ve described are the most striking empirical results, I think. One experiment that talks about unpacking conjunctions (rather than disjunctions, like the experiments I’ve described so far) is Boven and Epley (2003, particularly their first experiment, where they ask people how much an oil refinery should be punished for pollution. This pollution is described either as leading to an increase in “asthma, lung cancer, throat cancer, or all varieties of respiratory diseases,” or just as leading to an increase in “all varieties of respiratory diseases.” In the first condition, people want to punish refinery more. But, in spite of being notably different from previous unpacking experiments, still not what Eliezer was talking about.
What this shows is that people are inconsistent in a certain way. If you ask them the same question in two different ways (packed vs. unpacked) you get different answers. Is there any indication of which is the better way to ask the question, or whether asking it some other way is better still? Without an answer to this question, it’s unclear to me whether we should talk about an “unpacking fallacy” or a “failure to unpack fallacy”.
Here’s a handy example discussion of related conjunction issues from the Project Cyclops report:
We have outlined the development of technologically competent life on Earth as a succession of steps to each of [which] we must assign an a priori probability less than unity. The probability of the entire sequence occurring is the product of the individual (conditional) probabilities. As we study the chain of events in greater detail we may become aware of more and more apparently independent or only slightly correlated steps. As this happens, the a priori probability of the entire sequence approaches zero, and we are apt to conclude that, although life indeed exists here, the probability of its occurrence elsewhere is vanishingly small.
The trouble with this reasoning is that it neglects alternate routes that converge to the same (or almost the same) end result. We are reminded of the old proof that everyone has only an infinitesimal chance of existing. One must assign a fairly small probability to one’s parents and all one’s grandparents and (great)^n-grandparents having met and mated. Also one must assign a probability on the order of 2^-46 to the exact pairing of chromosomes arising from any particular mating. When the probabilities of all these independent events that led to a particular person are multiplied, the result quickly approaches zero. This is all true. Yet here we all are. The [explanation] is that, if an entirely different set of matings and fertilizations had occurred, none of “us” would exist, but a statistically indistinguishable generation would have been born, and life would have gone on much the same.
Regarding the “unpacking fallacy”: I don’t think you’ve pointed to a fallacy here. You have pointed to a particular causal pathway which seems to be quite specific, and I’ve claimed that this particular causal pathway has a tiny expected effect by virtue of its unlikeliness. The negation of this sequence of events simply can’t be unpacked as a conjunction in any natural way, it really is fundamentally a disjunction. You might point out that the competing arguments are weak, but they can be much stronger in the cases were they aren’t predicated on detailed stories about the future.
As you say, even events that actually happened can also be made to look quite unlikely. But those events were, for the most part, unlikely ex ante. This is like saying “This argument can suggest that any lottery number probably wouldn’t win the lottery, even the lottery numbers that actually won!”
If you had a track record of successful predictions, or if anyone who embraced this view had a track record of successful predictions, maybe you could say “all of these successful predictions could be unpacked, so you shouldn’t be so skeptical of unpackable arguments.” But I don’t know of anyone with a reasonably good predictive record who takes this view, and most smart people seem to find it ridiculous.
I don’t understand your argument here. Yes, future civilization builds AI. It doesn’t follow that the value of the future is first determined by what type of AI they build (they also build nanotech, but the value of the future isn’t determined by the type of nanotech they build, and you haven’t offered a substantial argument that discriminates between the cases). There could be any number of important events beforehand or afterwards; there could be any number of other important characteristics surrounding how they build AI which influence whether the outcome is positive or negative.
Do you think the main effects of economic progress in 1600 were on the degree of parallelization in AI work? 1800? The magnitude of the direct effects of economic progress on AI work depends on how close the economic progress is to the AI work; as the time involved gets larger, indirect effects come to dominate.
You have a specific view, that there is a set of problems which need to be solved in order to make AI friendly, and that these problems have some kind of principled relationship to the problems that seem important to you now. This is as opposed to e.g. “there are two random approaches to AI, one of which leads to good outcomes and one of which leads to bad outcomes,” or “there are many approaches to AI, and you have to think about it in advance to figure out which lead to good outcomes” or “there is a specific problem that you can’t have solved by the time you get to AI if you want to to have a positive outcome” or an incredible variety of alternative models. The “parallelization is bad” argument doesn’t apply to most of these models, and in some you have “parallelization is good.”
Even granting that your picture of AI vs. FAI is correct, and there are these particular theoretical problems that need to be solved, it is completely unclear that more people working in the field makes things worse. I don’t know why you think this follows from 3 or can be sensibly lumped with 3, and you don’t provide an argument. Suppose I said “The most important thing about dam safety is whether you have a good theoretical understanding the dam before building it” and you said “Yes, and if you increase the number of people working on the dam you are less likely to understand it by the time it gets built, because someone will stumble across an ad hoc way to build a dam.” This seems ridiculous both a priori and based on the empirical evidence. There are many possible models for the way that important problems in AI get solved, and you seem to be assuming a particular one.
Suppose that I airdrop in a million knowledge workers this year and they leave next year, corresponding to an exogenous boost in productivity this year. You are claiming that this obviously increases the degree of parallelization on relevant AI work. This isn’t obvious, unless a big part of the relevant work is being done today (which seems unlikely, casually?)
I agree that I’ve only argued that your argument has a tiny impact; it could still dominate if there was literally nothing else going on. But even granting 1-5 there seem to be other big effects from economic growth.
The case in favor of growth seems to be pretty straightforward; I linked to a blog post in the last comment. Let me try to make the point more clearly:
Increasing economic activity speeds up a lot of things. Speeding up everything is neutral, so the important point is the difference between what it speeds up and what it doesn’t speed up. Most things people are actually trying to do get sped up, while a bunch of random things (aging and disease, natural disasters, mood changes) don’t get sped up. Lots of other things get sped up but significantly less than 1-for-1, because they have some inputs that get sped up and some that don’t (accidents of all kinds, conflicts of all kinds, resource depletion). Given that things people are trying to do get sped up, and the things that happen which they aren’t trying to do get sped up less, we should expect the effect to be to positive, as long as people are trying to do good things.
What’s a specific relevant example of something people are trying to speed up / not speed up besides AGI (= UFAI) and FAI? You pick out aging, disease, and natural disasters as not-sped-up but these seem very loosely coupled to astronomical benefits.
Increasing capital stocks, improving manufacturing, improving education, improving methodologies for discourse, figuring out important considerations. Making charity more efficient, ending poverty. Improving collective decision-making and governance. All of the social sciences. All of the hard sciences. Math and philosophy and computer science. Everything that everyone is working on, everywhere in the world.
I picked out conflict, accidents, and resource depletion as not being sped up 1-for-1, i.e. such that a 1% boost in economic activity corresponds to a <1% boost in those processes. Most people would say that war and accidents account for many bad things that happen. War is basically defined by people making decisions that are unusually misaligned with aggregate welfare. Accidents are basically defined by people not getting what they want. I could have lumped in terrorism, and then accounted for basically all of the ways that we can see things going really badly in the present day.
You have a particular story about how a bad thing might happen in the future. Maybe that’s enough to conclude the future will be entirely unlike the present. But it seems like (1) that’s a really brittle way to reason, however much you want to accuse its detractors of the “unpacking fallacy,” and most smart people take this view, and (2) even granting almost all of your assumptions, it’s pretty easy to think of scenarios where war, terrorism, or accidents are inputs into AI going badly, or where better education, more social stability, or better decision-making are inputs into AI going well. People promoting these positive changes are also working against forces that wouldn’t be accelerated, like people growing old and dying and thereby throwing away their accumulated human capital, or infrastructure being stressed to keep people alive, etc. etc.
Increasing capital stocks, improving manufacturing, improving education, improving methodologies for discourse, figuring out important considerations. Making charity more efficient, ending poverty. Improving collective decision-making and governance. All of the social sciences. All of the hard sciences. Math and philosophy and computer science. Everything that everyone is working on, everywhere in the world.
How is an increased capital stock supposed to improve our x-risk / astronomical benefit profile except by being an input into something else? Yes, computer science benefits, that’s putatively the problem. We need certain types of math for FAI but does math benefit more from increased capital stocks compared to, say, computing power? Which of these other things are supposed to save the world faster than computer science destroys it, and how? How the heck would terrorism be a plausible input into AI going badly? Terrorists are not going to be the most-funded organizations with the smartest researchers working on AGI (= UFAI) as opposed to MIT, Google or Goldman Sachs.
Does your argument primarily reduce to “If there’s no local FOOM then economic growth is a good thing, and I believe much less than you do in local FOOM”? Or do you also think that in local FOOM scenarios higher economic growth now expectedly results in a better local FOOM? And if so is there at least one plausible specific scenario that we can sketch out now for how that works, as opposed to general hopes that a higher economic growth exponent has vague nice effects which will outweigh the shortening of time until the local FOOM with a correspondingly reduced opportunity to get FAI research done in time? When you sketch out a specific scenario, this makes it possible to point out fragile links which conjunctively decrease the probability of that scenario, and often these fragile links generalize, which is why it’s a bad idea to keep things vague and not sketch out any concrete scenarios for fear of the conjunction fallacy.
It seems to me that a lot of your reply, going by the mention of things like terrorism and poverty, must be either prioritizing near-term benefits over the astronomical future, or else being predicated on a very different model from local FOOM. We already have a known persistent disagreement on local FOOM. This is an important modular part of the disagreement on which other MIRIfolk do not all line up on one side or another. Thus I would like to know how much we disagree about expected goodness of higher econ growth exponents given local FOOM, and whether there’s a big left over factor where “Paul Christiano thinks you’re just being silly even assuming that a FOOM is local”, especially if this factor is not further traceable to a persistent disagreement about competence of elites. It would then be helpful to sketch out a concrete scenario corresponding to this disagreement to see if it looks even more fragile and conjunctive.
(Note that e.g. Wei Dai also thought it was obviously true that faster econ growth exponents had a negative-sign effect on FAI, though, like me, this debate made him question (but not yet reject) the ‘obvious’ conclusion.)
(Note that e.g. Wei Dai also thought it was obviously true that faster econ growth exponents had a negative-sign effect on FAI, though, like me, this debate made him question (but not yet reject) the ‘obvious’ conclusion.)
I’m confused by the logic of this sentence (in particular how the ‘though’ and ‘like me’ fit together). Are you saying that you and Wei both at first accepted that faster econ growth meant less chance of FAI, but then were both caused to doubt this conclusion by the fact that others debated the claim?
Even given a very fast local foom (to which I do assign a pretty small probability, especially as we make the situation more detailed and conclude that fewer things are relevant), I would still expect higher education and better discourse to improve the probability that people handle the situation well. It’s weird to cash this out as a concrete scenario, because that just doesn’t seem like how reasonable reasoning works.
But trying anyway: someone is deciding whether to run an AI or delay, and they correctly choose to delay. Someone is arguing that research direction X is safer than research direction Y, and others are more likely to respond selectively to correct arguments. Someone is more likely to notice there is a problem with a particular approach and they should do something differently, etc. etc.
Similarly, I expect war or external stressors to make things worse, but it seems silly to try and break this down as very specific situations. In general, people are making decisions about what to do, and if they have big alternative motivations (like winning a war, or avoiding social collapse, or what have you), I expect them to make decisions that are less aligned with aggregate welfare. They choose to run a less safe AI, they pursue a research direction that is less safe, etc. Similarly, I expect competent behavior by policy-makers to improve the situation across a broad distribution of scenarios, and I think that is less likely given other pressing issues. We nationalize AI projects, we effectively encourage coordination of AI researchers, we fund more safety-conscious research, etc. Similarly, I expect that an improved understanding of forecasting and decision-making would improve outcomes, and improved understanding of social sciences would play a small role in this. And so on.
But at any rate, my main question is how you can be so confident of local foom that you think this tiny effect given local foom scenarios dominates the effect given business as usual? I don’t understand where you are coming from there. The secondary objection is to your epistemic framework. I have no idea how you would have thought about the future if you lived in 1800 or even 1900; it seems almost certain that this framework reasoning would have led you to crazy conclusions, and I’m afraid that the same thing is true in 2000. You just shouldn’t expect to be able to think of detailed situations that determine the whole value of the universe, unless you are in an anomalous situation, but that doesn’t mean that your actions have no effect and that you should condition on being in an anomalous situation.
Even given a very fast local foom (to which I do assign a pretty small probability, especially as we make the situation more detailed and conclude that fewer things are relevant), I would still expect higher education and better discourse to improve the probability that people handle the situation well. It’s weird to cash this out as a concrete scenario, because that just doesn’t seem like how reasonable reasoning works.
But trying anyway: someone is deciding whether to run an AI or delay, and they correctly choose to delay. Someone is arguing that research direction X is safer than research direction Y, and others are more likely to respond selectively to correct arguments. Someone is more likely to notice there is a problem with a particular approach and they should do something differently, etc. etc.
How did this happen as a result of economic growth having a marginally greater exponent? Doesn’t that just take us to this point faster and give less time for serial thought, less time for deep theories, less time for the EA movement to spread faster than the exponent on economic growth, etcetera? This decision would ceteris paribus need to be made at some particular cumulative level of scientific development, which will involve relatively more parallel work and relatively less serial work if the exponent of econ growth is higher. How does that help it be made correctly?
Exposing (and potentially answering) questions like this is very much the point of making the scenario concrete, and I have always held rather firmly on meta-level epistemic grounds that visualizing things out concretely is almost always a good idea in math, science, futurology and anywhere. You don’t have to make all your predictions based on that example but you have to generate at least one concrete example and question it. I have espoused this principle widely and held to it myself in many cases apart from this particular dispute.
But at any rate, my main question is how you can be so confident of local foom that you think this tiny effect given local foom scenarios dominates the effect given business as usual?
Procedurally, we’re not likely to resolve that particular persistent disagreement in this comment thread which is why I want to factor it out.
My secondary objection is to your epistemic framework. I have no idea how you would have thought about the future if you lived in 1800 or even 1900; it seems almost certain that this framework reasoning would have led you to crazy conclusions, and I’m afraid that the same thing is true in 2000.
I could make analogies about smart-people-will-then-decide and don’t-worry-the-elite-wouldn’t-be-that-stupid reasoning to various historical projections that failed, but I don’t think we can get very much mileage out of nonspecifically arguing which of us would have been more wrong about 2000 if we had tried to project it out while living in 1800. I mean, obviously a major reason I don’t trust your style of reasoning is that I think it wouldn’t have worked historically, not that I think your reasoning mode would have worked well historically but I’ve decided to reject it because I’m stubborn. (If I were to be more specific, when I listen to your projections of future events they don’t sound very much like recollections of past events as I have read about them in history books, where jaw-dropping stupidity usually plays a much stronger role.)
I think an important thing to keep in mind throughout is that we’re not asking whether this present world would be stronger and wiser if it were economically poorer. I think it’s much better to frame the question as whether we would be in a marginally better or worse position with respect to FAI today if we had the present level of economic development but the past century from 1913-2013 had taken ten fewer years to get there so that the current date were 2003. This seems a lot more subtle.
past events as I have read about them in history books, where jaw-dropping stupidity
usually plays a much stronger role.
How sure are you that this isn’t hindsight bias, that if various involved historical
figures had been smarter they would have understood the situation and not done things
that look unbelievably stupid looking back?
We are discussing the relative value of two different things: the stuff people do intentionally (and the byproducts thereof), and everything else.
In the case of the negative scenarios I outlined this is hopefully clear: wars aren’t sped up 1-for-1, so there will be fewer wars between here and any relevant technological milestones. And similarly for other stressors, etc.
Regarding education: Suppose you made everything 1% more efficient. The amount of education a person gets over their life is 1% higher (because you didn’t increase the pace of aging / turnover between people, which is the thing people were struggling against, and so people do better at getting what they want).
Other cases seem to be similar: some things are a wash, but more things get better than worse, because systematically people are pushing on the positive direction.
Procedurally, we’re not likely to resolve that particular persistent disagreement in this comment thread which is why I want to factor it out.
This discussion was useful for getting a more precise sense of what exactly it is you assign high probability to.
I wish you two had the time for a full-blown adversarial collaboration on this topic, or perhaps on some sub-problem within the topic, with Carl Shulman as moderator.
At some point I need to write a post about how I’m worried that there’s an “unpacking fallacy” or “conjunction fallacy fallacy” practiced by people who have heard about the conjunction fallacy...
Please do this. I really, really want to read that post. Also I think writing it would save you time, since you could then link to it instead of re-explaining it in comments. (I think this is the third time I’ve seen you say something about that post, and I don’t read everything you write.)
If there’s anything I can do to help make this happen (such as digging through your old comments for previous explanations of this point, copyediting, or collecting a petition of people who want to see the post to provide motivation), please please please let me know.
If there’s anything I can do to help make this happen (such as digging through your
old comments for previous explanations of this point, copyediting, or collecting a
petition of people who want to see the post to provide motivation), please please
please let me know.
My experience has been that asking people “let me know if I can help” doesn’t result
in requests for help. I’d suggest just going ahead and compiling a list of relevant
comments (like this one) and sending them along.
(If Eliezer doesn’t end up writing the post, well, you now have a bunch of comments
you could use to get started on a post yourself.)
See here and here.
This position seems unlikely to me at face value. It relies on a very long list of claims, and given the apparently massive improbability of the conjunction, there is no way this consideration is going to be the biggest impact of economic progress:
The most important determinant of future welfare is whether you get FAI or UFAI (this presupposes a relatively detailed model of how AI works, of what the danger looks like, etc.)
This will happen quite soon, and relevant AI work is already underway.
The main determinant of FAI vs. UFAI is whether an appropriate theoretical framework for goal-stability is in place.
As compared to UFAI work, the main difficulty for developing such a framework for goal-stability is the serial depth of the problem.
A 1% boost in economic activity this year has a non-negligible effect on the degree of parallelization of relevant AI work.
I don’t see how you can defend giving any of those points more than 1⁄2 probability, and I would give the conjunction less than 1% probability. Moreover, even in this scenario, the negative effect from economic progress is quite small. (Perhaps a 1% increase in sustained economic productivity makes the future 0.1% worse if this story is true, and each year of a 1% increase in economic productivity makes the future 0.002% worse?)
So on balance it seems to me like this would say that a 1% increase in economic productivity would make the future 0.00002% worse? That is a ridiculously tiny effect; even if you couldn’t see any particular reason that economic progress helped or hurt I think your prior should expect other effects to dominate, and you can use other considerations to get a handle on the sign. I guess you think that this is an underestimate, but I would be interested to know where you disagree.
I would guess that the positive effect from decreasing the cumulative risk of war alone are several orders of magnitude higher than that. I know you think that a world war that killed nearly everyone might be positive rather than negative, but in that case slowing down economic progress would still have positive effects that are orders of magnitude larger than the effect on FAI parallelization.
Even without doing any calculations, it is extraordinarily hard to imagine that the difference between “world at war” and “world at peace” is less than the difference between “world with slightly more parallelization in AI work” and “world with slightly less parallelization;” almost everyone would disagree with you, and as far as I can tell their reasons seem better than yours. Similarly, it is hard to imagine that the interaction between generational turnover and economic activity wouldn’t swamp this by several orders of magnitude.
General remark: At some point I need to write a post about how I’m worried that there’s an “unpacking fallacy” or “conjunction fallacy fallacy” practiced by people who have heard about the conjunction fallacy but don’t realize how easy it is to take any event, including events which have already happened, and make it look very improbable by turning one pathway to it into a large series of conjunctions. E.g. I could produce a long list of things which allegedly have to happen for a moon landing to occur, some of which turned out to not be necessary but would look plausible if added to the list ante facto, with no disjunctive paths to the same destination, and thereby make it look impossible. Generally this manifests when somebody writes a list of alleged conjunctive necessities, and I look over the list and some of the items seem unnecessary (my model doesn’t go through them at all), obvious disjunctive paths have been omitted, the person has assigned sub-50% probability to things that I see as mainline 90% probabilities, and conditional probabilities when you assume the theory was right about 1-N would be significantly higher for N+1. Most of all, if you imagine taking the negation of the assertion and unpacking it into a long list of conjunctive probabilities, it would look worse—there should be a name for the problem of showing that X has weak arguments but not considering that ~X has even weaker arguments. Or on a meta level, since it is very easy to make things look more conjunctive, we should perhaps not be prejudiced against things which somebody has helpfully unpacked for us into a big conjunction, when the core argument still seems pretty simple on some level.
When I look over this list, my reaction is that:
(1) is a mainline assumption with odds of 5:1 or 10:1 - of course future intergalactic civilization bottlenecks through the goals of a self-improving agency, how would you get to an intergalactic civilization without that happening? If this accounts for much of our disagreement then we’re thinking about entirely different scenarios, and I’m not sure how to update from your beliefs about mostly scenario B to my beliefs about mostly scenario A. It makes more sense to call (1) into question if we’re really asking about global vs. local, but then we get into the issue of whether global scenarios are mostly automatic losses anyway. If (1) is really about whether we should be taking into account a big chunk of survivable global scenarios then this goes back to a previous persistent disagreement.
(2) I don’t see the relevance—why does a long time horizon vs. a short time horizon matter? 80 years would not make me relax and say that we had enough serial depth, though it would certainly be good news ceteris paribus, there’s no obvious threshold to cross.
Listing (3) and (4) as separate items was what originally made my brain shout “unpacking fallacy!” There are several subproblems involved in FAI vs. UFAI, of which the two obvious top items are the entire system being conducive to goal stability through self-improvement which may require deducing global properties to which all subsystems must be conducive, and the goal loading problem. These both seem insight-heavy which will require serial time to solve. The key hypothesis is just that there are insight-heavy problems in FAI which don’t parallelize well relative to the wide space of cobbled-together designs which might succeed for UFAI. Odds here are less extreme than for (1) but still in the range of 2:1-4:1. The combined 3-4 issue is the main weak point, but the case for “FAI would parallelize better than UFAI” is even weaker.
(5) makes no sense to ask as a conditionally independent question separate from (1); if (1) is true then the only astronomical effects of modern-day economic growth are whatever effects that growth has on AI work, and to determine if economic growth is qualitatively good or bad, we ask about the sign of the effect neglecting its magnitude. I suppose if the effect were trivial enough then we could just increase the planet’s growth rate by 5% for sheer fun and giggles and it would have no effect on AI work, but this seems very unlikely; a wealthier planet will ceteris paribus have more AI researchers. Odds of 10:1 or better.
On net, this says that in my visualization the big question is just “Does UFAI parallelize better than FAI, or does FAI parallelize better than UFAI?” and we find that the case for the second clause is weaker than the first; or equivalently “Does UFAI inherently require serial time more than FAI requires serial time?” is weaker than “Does FAI inherently require serial time more than UFAI requires serial time?” This seems like a reasonable epistemic state to me.
The resulting shove at the balance of the sign of the effect of economic growth would have to be counterbalanced by some sort of stronger shove in the direction of modern-day economic growth having astronomical benefits. And the case for e.g. “More econ growth means friendlier international relations and so they endorse ideal Y which leads them to agree with me on policy Z” seems even more implausible when unpacked into a series of conjunctions. Lots of wealthy people and relatively friendly nations right now are not endorsing policy Z.
To summarize and simplify the whole idea, the notion is:
Right now my estimate of the sign of the astronomical effect of modern-day economic growth is dominated by a 2-node conjunction of, “Modern-day econ growth has a positive effect on resources into both FAI and UFAI” and “The case for FAI parallelizing better than UFAI is weaker than the converse case”. For this to be not true requires mainly that somebody else demonstrate an effect or set of effects in the opposite direction which has better net properties after its own conjunctions are taken into account. The main weakness in the argument and lingering hope that econ growth is good, isn’t that the original argument is very conjunctive, but rather it’s that faster econ growth seems like it should have a bunch of nice effects on nice things and so the disjunction of other econ effects might conceivably swing the sign the other way. But it would be nice to have at least one plausible such good effect without dropping our standards so low that we could as easily list a dozen equally (im)plausible bad effects.
With small enough values of ‘slightly’ obviously the former will have a greater effect, the question is the sign of that effect; also it’s not obvious to me that moderately lower amounts of econ growth lead to world wars, and war seems qualitatively different in many respects from poverty. I also have to ask if you are possibly maybe being distracted by the travails of one planet as a terminal value, rather than considering that planet’s instrumental role in future galaxies.
Related: There’s a small literature on what Tversky called “support theory,” which discusses packing and unpacking effects: Tversky & Koehler (1994); Ayton (1997); Rottenstreich & Tversky (1997); Macchi et al. (1997); Fox & Tversky (1998); Brenner & Koehler (1999); Chen et al. (2001); Boven & Epley (2003); Brenner et al. (2005); Bligin & Brenner (2008).
Luke asked me to look into this literature for a few hours. Here’s what I found.
The original paper (Tversky and Koehler 1994) is about disjunctions, and how unpacking them raises people’s estimate of the probaility. So for example, asking people to estimate the probability someone died of “heart disease, cancer, or other natural causes” yields a higher probability estimate than if you just ask about “natural causes.”
They consider the hypothesis this might be because they take the researcher’s apparent emphasis as evidence that’s it’s more likely, but they tested & disconfirmed this hypothesis by telling people to take the last digit of their phone number and estimate the percentage of couples that have that many children. Percentages sum to greater than 1.
Finally, they check whether experts are vulnerable to this bias by doing an experiment similar to the first experiment, but using physicians at Stanford University as the subjects and asking them about a hypothetical case of a woman admitted to an emergency room. They confirmed that yes, experts are vulnerable to this mistake too.
This phenomenon is known as “subadditivity.” A subsequent study (RottenStreich and Tversky 1997) found that subadditivity can even occur when dealing with explicit conjunctions. Macci et al. (1999) found evidence of superadditivity: ask some people how probable it is that the freezing point of alcohol is below that of gasoline, other people how probable it is that the freezing point of gasoline is below that of alcohol, average answers sum to less than 1.
Other studies try to refine the mathematical model of how people make judgements in these kinds of cases, but the experiments I’ve described are the most striking empirical results, I think. One experiment that talks about unpacking conjunctions (rather than disjunctions, like the experiments I’ve described so far) is Boven and Epley (2003, particularly their first experiment, where they ask people how much an oil refinery should be punished for pollution. This pollution is described either as leading to an increase in “asthma, lung cancer, throat cancer, or all varieties of respiratory diseases,” or just as leading to an increase in “all varieties of respiratory diseases.” In the first condition, people want to punish refinery more. But, in spite of being notably different from previous unpacking experiments, still not what Eliezer was talking about.
Below are some other messy notes I took:
http://commonsenseatheism.com/wp-content/uploads/2013/10/Fox-Tversky-A-belief-based-account-of-decision-under-uncertainty.pdf Uses support theory to develop account of decision under uncertainty.
http://commonsenseatheism.com/wp-content/uploads/2013/10/Brenner-Koehler-Subjective-probability-of-disjunctive-hypotheses-local-weight-models-for-decomposition-and-evidential-support.pdf Something about local weights; didn’t look at this one much.
http://commonsenseatheism.com/wp-content/uploads/2013/10/Chen-et-al-The-relation-between-probability-and-evidence-judgment-an-extension-of-support-theory.pdf Tweaking math behind support theory to allow for superadditivity.
http://commonsenseatheism.com/wp-content/uploads/2013/10/Brenner-et-al-Modeling-patterns-of-probability-calibration-with-random-support-theory.pdf Introduces notion of random support theory.
http://bear.warrington.ufl.edu/brenner/papers/bilgin-brenner-jesp08.pdf Unpacking effects weaker when dealing with near future as opposed to far future.
Other articles debating how to explain basic support theory results: http://bcs.siu.edu/facultypages/young/JDMStuff/Sloman%20(2004)%20unpacking.pdf http://aris.ss.uci.edu/~lnarens/Submitted/problattice11.pdf http://eclectic.ss.uci.edu/~drwhite/pw/NarensNewfound.pdf
What this shows is that people are inconsistent in a certain way. If you ask them the same question in two different ways (packed vs. unpacked) you get different answers. Is there any indication of which is the better way to ask the question, or whether asking it some other way is better still? Without an answer to this question, it’s unclear to me whether we should talk about an “unpacking fallacy” or a “failure to unpack fallacy”.
Here’s a handy example discussion of related conjunction issues from the Project Cyclops report:
Regarding the “unpacking fallacy”: I don’t think you’ve pointed to a fallacy here. You have pointed to a particular causal pathway which seems to be quite specific, and I’ve claimed that this particular causal pathway has a tiny expected effect by virtue of its unlikeliness. The negation of this sequence of events simply can’t be unpacked as a conjunction in any natural way, it really is fundamentally a disjunction. You might point out that the competing arguments are weak, but they can be much stronger in the cases were they aren’t predicated on detailed stories about the future.
As you say, even events that actually happened can also be made to look quite unlikely. But those events were, for the most part, unlikely ex ante. This is like saying “This argument can suggest that any lottery number probably wouldn’t win the lottery, even the lottery numbers that actually won!”
If you had a track record of successful predictions, or if anyone who embraced this view had a track record of successful predictions, maybe you could say “all of these successful predictions could be unpacked, so you shouldn’t be so skeptical of unpackable arguments.” But I don’t know of anyone with a reasonably good predictive record who takes this view, and most smart people seem to find it ridiculous.
I don’t understand your argument here. Yes, future civilization builds AI. It doesn’t follow that the value of the future is first determined by what type of AI they build (they also build nanotech, but the value of the future isn’t determined by the type of nanotech they build, and you haven’t offered a substantial argument that discriminates between the cases). There could be any number of important events beforehand or afterwards; there could be any number of other important characteristics surrounding how they build AI which influence whether the outcome is positive or negative.
Do you think the main effects of economic progress in 1600 were on the degree of parallelization in AI work? 1800? The magnitude of the direct effects of economic progress on AI work depends on how close the economic progress is to the AI work; as the time involved gets larger, indirect effects come to dominate.
You have a specific view, that there is a set of problems which need to be solved in order to make AI friendly, and that these problems have some kind of principled relationship to the problems that seem important to you now. This is as opposed to e.g. “there are two random approaches to AI, one of which leads to good outcomes and one of which leads to bad outcomes,” or “there are many approaches to AI, and you have to think about it in advance to figure out which lead to good outcomes” or “there is a specific problem that you can’t have solved by the time you get to AI if you want to to have a positive outcome” or an incredible variety of alternative models. The “parallelization is bad” argument doesn’t apply to most of these models, and in some you have “parallelization is good.”
Even granting that your picture of AI vs. FAI is correct, and there are these particular theoretical problems that need to be solved, it is completely unclear that more people working in the field makes things worse. I don’t know why you think this follows from 3 or can be sensibly lumped with 3, and you don’t provide an argument. Suppose I said “The most important thing about dam safety is whether you have a good theoretical understanding the dam before building it” and you said “Yes, and if you increase the number of people working on the dam you are less likely to understand it by the time it gets built, because someone will stumble across an ad hoc way to build a dam.” This seems ridiculous both a priori and based on the empirical evidence. There are many possible models for the way that important problems in AI get solved, and you seem to be assuming a particular one.
Suppose that I airdrop in a million knowledge workers this year and they leave next year, corresponding to an exogenous boost in productivity this year. You are claiming that this obviously increases the degree of parallelization on relevant AI work. This isn’t obvious, unless a big part of the relevant work is being done today (which seems unlikely, casually?)
I agree that I’ve only argued that your argument has a tiny impact; it could still dominate if there was literally nothing else going on. But even granting 1-5 there seem to be other big effects from economic growth.
The case in favor of growth seems to be pretty straightforward; I linked to a blog post in the last comment. Let me try to make the point more clearly:
Increasing economic activity speeds up a lot of things. Speeding up everything is neutral, so the important point is the difference between what it speeds up and what it doesn’t speed up. Most things people are actually trying to do get sped up, while a bunch of random things (aging and disease, natural disasters, mood changes) don’t get sped up. Lots of other things get sped up but significantly less than 1-for-1, because they have some inputs that get sped up and some that don’t (accidents of all kinds, conflicts of all kinds, resource depletion). Given that things people are trying to do get sped up, and the things that happen which they aren’t trying to do get sped up less, we should expect the effect to be to positive, as long as people are trying to do good things.
What’s a specific relevant example of something people are trying to speed up / not speed up besides AGI (= UFAI) and FAI? You pick out aging, disease, and natural disasters as not-sped-up but these seem very loosely coupled to astronomical benefits.
Increasing capital stocks, improving manufacturing, improving education, improving methodologies for discourse, figuring out important considerations. Making charity more efficient, ending poverty. Improving collective decision-making and governance. All of the social sciences. All of the hard sciences. Math and philosophy and computer science. Everything that everyone is working on, everywhere in the world.
I picked out conflict, accidents, and resource depletion as not being sped up 1-for-1, i.e. such that a 1% boost in economic activity corresponds to a <1% boost in those processes. Most people would say that war and accidents account for many bad things that happen. War is basically defined by people making decisions that are unusually misaligned with aggregate welfare. Accidents are basically defined by people not getting what they want. I could have lumped in terrorism, and then accounted for basically all of the ways that we can see things going really badly in the present day.
You have a particular story about how a bad thing might happen in the future. Maybe that’s enough to conclude the future will be entirely unlike the present. But it seems like (1) that’s a really brittle way to reason, however much you want to accuse its detractors of the “unpacking fallacy,” and most smart people take this view, and (2) even granting almost all of your assumptions, it’s pretty easy to think of scenarios where war, terrorism, or accidents are inputs into AI going badly, or where better education, more social stability, or better decision-making are inputs into AI going well. People promoting these positive changes are also working against forces that wouldn’t be accelerated, like people growing old and dying and thereby throwing away their accumulated human capital, or infrastructure being stressed to keep people alive, etc. etc.
How is an increased capital stock supposed to improve our x-risk / astronomical benefit profile except by being an input into something else? Yes, computer science benefits, that’s putatively the problem. We need certain types of math for FAI but does math benefit more from increased capital stocks compared to, say, computing power? Which of these other things are supposed to save the world faster than computer science destroys it, and how? How the heck would terrorism be a plausible input into AI going badly? Terrorists are not going to be the most-funded organizations with the smartest researchers working on AGI (= UFAI) as opposed to MIT, Google or Goldman Sachs.
Does your argument primarily reduce to “If there’s no local FOOM then economic growth is a good thing, and I believe much less than you do in local FOOM”? Or do you also think that in local FOOM scenarios higher economic growth now expectedly results in a better local FOOM? And if so is there at least one plausible specific scenario that we can sketch out now for how that works, as opposed to general hopes that a higher economic growth exponent has vague nice effects which will outweigh the shortening of time until the local FOOM with a correspondingly reduced opportunity to get FAI research done in time? When you sketch out a specific scenario, this makes it possible to point out fragile links which conjunctively decrease the probability of that scenario, and often these fragile links generalize, which is why it’s a bad idea to keep things vague and not sketch out any concrete scenarios for fear of the conjunction fallacy.
It seems to me that a lot of your reply, going by the mention of things like terrorism and poverty, must be either prioritizing near-term benefits over the astronomical future, or else being predicated on a very different model from local FOOM. We already have a known persistent disagreement on local FOOM. This is an important modular part of the disagreement on which other MIRIfolk do not all line up on one side or another. Thus I would like to know how much we disagree about expected goodness of higher econ growth exponents given local FOOM, and whether there’s a big left over factor where “Paul Christiano thinks you’re just being silly even assuming that a FOOM is local”, especially if this factor is not further traceable to a persistent disagreement about competence of elites. It would then be helpful to sketch out a concrete scenario corresponding to this disagreement to see if it looks even more fragile and conjunctive.
(Note that e.g. Wei Dai also thought it was obviously true that faster econ growth exponents had a negative-sign effect on FAI, though, like me, this debate made him question (but not yet reject) the ‘obvious’ conclusion.)
I’m confused by the logic of this sentence (in particular how the ‘though’ and ‘like me’ fit together). Are you saying that you and Wei both at first accepted that faster econ growth meant less chance of FAI, but then were both caused to doubt this conclusion by the fact that others debated the claim?
Yep.
This was one of those cases where precisely stating the question helps you get to the answer. Thanks for the confirmation!
Even given a very fast local foom (to which I do assign a pretty small probability, especially as we make the situation more detailed and conclude that fewer things are relevant), I would still expect higher education and better discourse to improve the probability that people handle the situation well. It’s weird to cash this out as a concrete scenario, because that just doesn’t seem like how reasonable reasoning works.
But trying anyway: someone is deciding whether to run an AI or delay, and they correctly choose to delay. Someone is arguing that research direction X is safer than research direction Y, and others are more likely to respond selectively to correct arguments. Someone is more likely to notice there is a problem with a particular approach and they should do something differently, etc. etc.
Similarly, I expect war or external stressors to make things worse, but it seems silly to try and break this down as very specific situations. In general, people are making decisions about what to do, and if they have big alternative motivations (like winning a war, or avoiding social collapse, or what have you), I expect them to make decisions that are less aligned with aggregate welfare. They choose to run a less safe AI, they pursue a research direction that is less safe, etc. Similarly, I expect competent behavior by policy-makers to improve the situation across a broad distribution of scenarios, and I think that is less likely given other pressing issues. We nationalize AI projects, we effectively encourage coordination of AI researchers, we fund more safety-conscious research, etc. Similarly, I expect that an improved understanding of forecasting and decision-making would improve outcomes, and improved understanding of social sciences would play a small role in this. And so on.
But at any rate, my main question is how you can be so confident of local foom that you think this tiny effect given local foom scenarios dominates the effect given business as usual? I don’t understand where you are coming from there. The secondary objection is to your epistemic framework. I have no idea how you would have thought about the future if you lived in 1800 or even 1900; it seems almost certain that this framework reasoning would have led you to crazy conclusions, and I’m afraid that the same thing is true in 2000. You just shouldn’t expect to be able to think of detailed situations that determine the whole value of the universe, unless you are in an anomalous situation, but that doesn’t mean that your actions have no effect and that you should condition on being in an anomalous situation.
How did this happen as a result of economic growth having a marginally greater exponent? Doesn’t that just take us to this point faster and give less time for serial thought, less time for deep theories, less time for the EA movement to spread faster than the exponent on economic growth, etcetera? This decision would ceteris paribus need to be made at some particular cumulative level of scientific development, which will involve relatively more parallel work and relatively less serial work if the exponent of econ growth is higher. How does that help it be made correctly?
Exposing (and potentially answering) questions like this is very much the point of making the scenario concrete, and I have always held rather firmly on meta-level epistemic grounds that visualizing things out concretely is almost always a good idea in math, science, futurology and anywhere. You don’t have to make all your predictions based on that example but you have to generate at least one concrete example and question it. I have espoused this principle widely and held to it myself in many cases apart from this particular dispute.
Procedurally, we’re not likely to resolve that particular persistent disagreement in this comment thread which is why I want to factor it out.
I could make analogies about smart-people-will-then-decide and don’t-worry-the-elite-wouldn’t-be-that-stupid reasoning to various historical projections that failed, but I don’t think we can get very much mileage out of nonspecifically arguing which of us would have been more wrong about 2000 if we had tried to project it out while living in 1800. I mean, obviously a major reason I don’t trust your style of reasoning is that I think it wouldn’t have worked historically, not that I think your reasoning mode would have worked well historically but I’ve decided to reject it because I’m stubborn. (If I were to be more specific, when I listen to your projections of future events they don’t sound very much like recollections of past events as I have read about them in history books, where jaw-dropping stupidity usually plays a much stronger role.)
I think an important thing to keep in mind throughout is that we’re not asking whether this present world would be stronger and wiser if it were economically poorer. I think it’s much better to frame the question as whether we would be in a marginally better or worse position with respect to FAI today if we had the present level of economic development but the past century from 1913-2013 had taken ten fewer years to get there so that the current date were 2003. This seems a lot more subtle.
How sure are you that this isn’t hindsight bias, that if various involved historical figures had been smarter they would have understood the situation and not done things that look unbelievably stupid looking back?
Do you have particular historical events in mind?
We are discussing the relative value of two different things: the stuff people do intentionally (and the byproducts thereof), and everything else.
In the case of the negative scenarios I outlined this is hopefully clear: wars aren’t sped up 1-for-1, so there will be fewer wars between here and any relevant technological milestones. And similarly for other stressors, etc.
Regarding education: Suppose you made everything 1% more efficient. The amount of education a person gets over their life is 1% higher (because you didn’t increase the pace of aging / turnover between people, which is the thing people were struggling against, and so people do better at getting what they want).
Other cases seem to be similar: some things are a wash, but more things get better than worse, because systematically people are pushing on the positive direction.
This discussion was useful for getting a more precise sense of what exactly it is you assign high probability to.
I wish you two had the time for a full-blown adversarial collaboration on this topic, or perhaps on some sub-problem within the topic, with Carl Shulman as moderator.
Please do this. I really, really want to read that post. Also I think writing it would save you time, since you could then link to it instead of re-explaining it in comments. (I think this is the third time I’ve seen you say something about that post, and I don’t read everything you write.)
If there’s anything I can do to help make this happen (such as digging through your old comments for previous explanations of this point, copyediting, or collecting a petition of people who want to see the post to provide motivation), please please please let me know.
My experience has been that asking people “let me know if I can help” doesn’t result in requests for help. I’d suggest just going ahead and compiling a list of relevant comments (like this one) and sending them along.
(If Eliezer doesn’t end up writing the post, well, you now have a bunch of comments you could use to get started on a post yourself.)