Carl_Shulman
Karma: 1,775
NewTop
Hi Emanuele,
I saw your request for commentary on Facebook, so here are some off-the-cuff comments (about 1 hour’s worth so take with appropriate grains of salt, but summarizing prior thinking):
My prior take on metformin was that it seems promising for its space (albeit with mixed evidence, and prior longevity drug development efforts haven’t panned out, but the returns would be very high for medical research if true), although overall the space looks less promising than x-risk reduction to me; the following comments will be about details of the analysis where I would currently differ
The suggestion of this trial moving forward LEV by 3+ years through an icebreaker effect boosting research looks wildly implausible to me
LEV is not mainly bottlenecked on ‘research on aging,’ e.g. de Grey’s proposals require radical advances in generally medically applicable stem cell and genetic engineering technologies that already receive massive funding and are quite challenging; the ability to replace diseased cells with genetically engineered stem cell derived tissues is already a major priority, and curing cancer is a small subset of SENS
Much of the expected gain in biomedical technology is not driven by shifts within biology, and advances within a particular medical field are heavily driven by broader improvements (e.g. computers, CRISPR, genome sequencing, PCR, etc); if LEV is far off and heavily dependent on other areas, then developments in other fields will make it comparatively easy for aging research to benefit from ‘catch up growth’ reducing the expected value of immediate speedup (almost all of which would have washed away if LEV happens in the latter half of the century)
In particular, if automating R&D with AI is easier than LEV, and would moot prior biomedical research, then that adds an additional discount factor; I would bet that this happens before LEV through biomedical research
Getting approval to treat ‘aging’ isn’t actually particularly helpful relative to approval for ‘diseases of aging’ since all-cause mortality requires larger trials and we don’t have great aging biomarkers; and the NIH has taken steps in that direction regardless
Similar stories have been told about other developments and experiments, which haven’t had massive icebreaker effects
Combined, these effects look like they cost a couple orders of magnitude
From my current epistemic state the expected # of years added by metformin looks too high
Re the Guesstimate model the statistical power of the trial is tightly tied to effect size; the larger the effect size the fewer people you need to show results; that raises the returns of small trials, but means you have diminishing returns for larger ones (you are spending more money to detect smaller effects so marginal cost-effectiveness goes a lot lower than average cost-effectiveness, reflecting high VOI of testing the more extravagant possibility)
Likewise the proportion using metformin conditional on a positive result is also correlated with effect size (which raises average EV, but shifts marginal EV lower proportionate to average EV); also the proportion of users seems too low to me conditional on success
One issue I would add to your theoretical analysis: with assigning 1000+ QALYs to letting someone reach LEV is that people commonly don’t claim linear utility with lifespan, i.e. they would often prefer to live to 80 with certainty rather than die at 20 with 90% probability and live to 10,000 with 10% probability.
I agree it’s worth keeping the chance that people will be able to live much longer in the future in mind when assessing benefits to existing people (I would also add the possibility of drastic increases in quality of life through technology). I’d guess most of this comes from broader technological improvements (e.g. via AI) rather than reaching LEV through biomedical approaches), but not with extreme confidence.
However, I don’t think it has very radical implications for cause prioritization since, as you note, deaths for any reason (include malaria and global catastrophes) deny those people a chance at LEV. LEV-related issues are also mainly a concern for existing humans, so to the extent one gives a boost for enormous impacts on nonhuman animals and the existence of future generations, LEV speedup won’t reap much of those boosts.
Within the field of biomedical research, aging looks relatively promising, and I think on average the best-targeted biomedical research does well for current people compared to linear charity in support of deployment (e.g. gene drives vs bednets). But it’s not a slam dunk because the problems are so hard (including ones receiving massive investment). I don’t see it as strongly moving most people who prefer to support bednets over malaria gene drives, farmed animal welfare over gene drives, or GCR reduction over gene drives.
″ Oh, is the concern that they’re looking at a more biased subset of possible effects (by focusing primarily on effects that seem positive)? ”
Yes. It doesn’t mention other analyses that have come to opposite conclusions by considering effects on wild animals and long-term development.
If you’re going to select interventions specifically to reduce the human population and have downstream consequences, it seems absolutely essential to take a broader view of the empirical consequences than in the linked report. E.g. among others, effects on wild animals (not mentioned but most immediate animal effects of this change will be on wild animals), future technological advancement, and global catastrophic risks have good cases for being far larger and plausibly of opposite sign to the effects discussed in the report but are not mentioned even as areas for further investigation.
What about a report along the lines of ‘I am donating in support of X, for highly illegible reasons relating to my intuition from looking at their work, and private information I have about them personally’?
This is a good point, and worth highlighting in discussion of reports (especially as we get more data on the effects of winning on donation patterns). On the other hand, the average depth and quality of investigation by winners (and the access they got) does seem higher than what they would otherwise have done, whilst less than expert donors.
I don’t think this is true. The probabilities and payouts are the same for any given participant, regardless of what others do, so people who are unlikely to write up a report don’t reduce the average number of reports produced by those who would.
Except that the pot size isn’t constrained by the participation of small donors: the CEA donor lottery has fixed pot sizes guaranteed by large donors, and the largest donors could be ~risk-neutral over lotteries with pots of many millions of donors. So there is no effect of this kind, and there is unlikely to ever be one except at ludicrously large scales (where one could use derivatives or the like to get similar effects).
Yes, the main effect balances out like that.
But insofar as the lottery enhances the effectiveness of donors (by letting them invest more in research if they win, amortized against a larger donation), then you want donors doing good to be enhanced and donors doing bad not to be enhanced. So you might want to try to avoid boosting pot size available to bad donors, and ensure good donors have large pots available. The CEA lottery is structured so that question doesn’t arise.
There is also the minor issue of correlation with other donors in the same block mentioned in the above comment, although you could ask CEA for a separate block if some unusual situation meant your donation plans would change a lot if you found out another block participant had won.
> but also in the other 80% of worlds you have a preference for your money being allocated by people who are more thoughtful.
For the CEA donor lottery, the pot size is fixed independent of one’s entry as the guarantor (Paul Christiano last year, the regranting pool I am administering this year) puts in funds for any unclaimed tickets. So the distribution of funding amounts for each entrant is unaffected by other entrants. It’s set up this way specifically so that people don’t even have to think about the sort of effect you discuss (the backstop fund has ~linear value of funds over the relevant range, so that isn’t an impact either).
The only thing that participating in the same lottery block as someone else matters for is correlations between your donations and theirs. E.g. if you would wind up choosing a different charity to give to depending on whether another participant won the lottery. But normally the behavior of one other donor wouldn’t change what you think is the best opportunity.
What happened in those cases?
I would love to see a canonical post making this argument, conflating EA with the benefits of maxing out personal warm fuzzies is one of my pet peeves.
I actually happen to think that the report was too dismissive of more leveraged climate change interventions that I expected could be a lot better than the estimates for Cool Earth (especially efficient angles on scientific research and political activity in the climate space), but the OP is suggesting that the original Cool Earth numbers (which indicate much lower cost-effectiveness than charities recommended by EAs in other areas with more robust data) were overstated, not understated (as the original report would suggest due to regression to the mean and measurement error).
One thing to emphasize more than that writeup did, is that in EA terms donating to such a lightly researched intervention (a few months work) is very likely dominated by donations to better research the area, finding higher expected value options and influencing others.
On the other hand, the point estimates in that report favored other charities like AMF over Cool Earth anyway, a conclusion strengthened by the OP critique (not that it excludes something else orders of magnitude better being found like unusual energy research, very effective political lobbying, geoengineering, etc; Open Philanthropy has made a few climate grants that look relatively leveraged).
And I agree with John Maxwell about it being oversold in some cases.
″ I’d like to hear if there’s been any relevant work done on this topic (either within EA organizations or within general academia). Increasing returns is a fairly common topic within economics, so I figure there is plenty of relevant research out there on this. ”
These are my key reasons (with links to academic EA and other discussions) for seeing diminishing returns as the relevant situation on average for EA as a whole, and in particular the most effective causes:
If problems can be solved, and vary in difficulty over multiple orders of magnitude (in required inputs), you will tend to see diminishing returns as you plot the number of problems solved with increasing resources; see this series of posts by Owen Cotton-Barratt and others
Empirically, we do see systematic diminishing returns to R&D inputs across fields of scientific and technological innovation, and for global total factor productivity; but historically the greatest successes of philanthropy, reductions in poverty, and increased prosperity have stemmed from innovation, and many EA priorities involve research and development
In politics and public policy the literatures on lobbying and campaign finance suggest diminishing returns
In growing new movements, there is an element of compounding returns, as new participants carry forward work (including further growth), and so influencing; this topic has been the subject of a fair amount of EA attention
When there are varied possible expenditures with widely varying cost-effectiveness and some limits on room for more funding (eventually, there may be increasing returns before that), then working one’s way from the most effective options to the least produces a diminishing returns curve at a scale large enough to encompass multiple interventions; Toby Ord discusses the landscape of global health interventions having this property
Elaborating on the idea of limits to funding and scaling: an extremely cost-effective intervention with linear or increasing returns that scaled to very large expenditures would often imply impossibly large effects; there can be cheap opportunities to save a human life today for $100 under special circumstances, but there can’t be trillions of dollars worth of such opportunities, since you would be saving more than the world population; likewise the probability of premature extinction cannot fall below 0, etc
So far EA is still small and unusual relative to the world, and much of its activity is harvesting low-hanging fruit from areas with diminishing returns (a consequence of those fruit) that couldn’t be scaled to extremes (this is least true for linear aid interventions added to the already large global aid and local programs, and in particular GiveDirectly, but holds for what I would consider more promising, in cost-effectiveness, EA global health interventions such as gene drive R&D for malaria eradication); as EA activity expands more currently underfunded areas will see returns diminish to the point of falling behind interventions with more linear or increasing returns but worse current cost-effectiveness
Experience with successes using neglectedness (which in prioritization practice does involve looking at the reasons for neglect) thus far, at least on dimensions for which feedback has yet arrived
″ Ideally, we would like to not simply select causes that are neglected, but to select causes that are neglected for reasons other than their impact. ”
Agreed.
The examples in the post have expected utilities assigned using inconsistent methodologies. If it’s possible to have long-run effects on future generations, then many actions will have such effects (elections can cause human extinction sometimes, an additional person saved from malaria could go on to cause or prevent extinction). If ludicrously vast universes and influence over them are subjectively possible, then we should likewise consider being less likely to get ludicrous returns if we are extinct or badly-governed (see ‘empirical stabilization assumptions’ in Nick Bostrom’s infinite ethics paper). We might have infinite impact (under certain decision theories) when we make a decision to eat a sandwich if there are infinite physically identical beings in the universe who will make the same decision as us.
Any argument of the form “consider type of consequence X, which is larger than consequences you had previously considered, as it applies to option A” calls for application of X to analyzing other options. When you do that you don’t get any 10^100 differences in expected utility of this sort, without an overwhelming amount of evidence to indicate that A has 10^100+ times the impact on X as option B or C (or your prior over other and unknown alternatives you may find later).
A strictly positive critical level that is low enough such that it would not result in the choice for that counter-intuitive situation, is still posiible.
As a matter of mathematics this appears impossible. For any critical level c that you pick where c>0, there is some level of positive welfare w where c>w>0, with relative utility u, 0>u, u=c-w.
There will then be some quantity of expected negative utility and relative utility people with relative utility between 0 and u that variable CLU would prefer to the existence of you with c and w. You can use gambles (with arbitrarily divisible probabilities) or aggregation across similar people to get arbitrarily close to zero. So either c<=0 or CLU will recommend creation of negative utility and relative utility people to prevent your existence for some positive welfare levels.
This objection to fixed critical level utilitarianism can be easily avoided with variable critical level utilitarianism. Suppose there is someone with a positive utility (a very happy person), who sets his critical level so high that a situation should be chosen where he does not exist, and where extra people with negative utilities exist. Why would he set such a high critical level? He cannot want that. This is even more counter-intuitive than the repugnant sadistic conclusion. With fixed critical level utilitarianism, such counter-intuitive conclusion can occur because everyone would have to accept the high critical level. But variable critical level utilitarianism can easily avoid it by taking lower critical levels.
Such situations exist for any critical level above zero, since any critical level above zero means treating people with positive welfare as a bad thing, to be avoided even at the expense of some amount of negative welfare.
If you think the idea of people with negative utility being created to prevent your happy existence is even more counterintuitive than people having negative welfare to produce your happy existence, it would seem your view would demand that you set a critical value of 0 for yourself.
For example: I have a happy life with a positive utility. But if one could choose another situation where I did not exist and everyone else was maximally happy and satisfied, I would prefer (if that would still be an option) that second situation, even if I don’t exist in that situation.
A situation where you don’t exist but uncounted trillions of others are made maximally happy is going to be better in utilitarian terms (normal, critical-level, variable, whatever), regardless of your critical level (or theirs, for that matter). A change in your personal critical level only changes the actions recommended by your variable CLU when it changes the rankings of actions in terms of relative utilities, such that the actions were close to within a distance on the scale of one life.
In other words, that’s a result of the summing up of (relative) welfare, not a reason to misstate your valuation of your own existence.
I have several issues with the internal consistency of this argument:
If individuals are allowed to select their own critical levels to respect their autonomy and preferences in any meaningful sense, that seems to imply respecting those people who value their existence and so would set a low critical level; then you get an approximately total view with regards to those sorts of creatures, and so a future populated with such beings can still be astronomically great
The treatment of zero levels seems inconsistent: if it is contradictory to set a critical level below the level one would prefer to exist, it seems likewise nonsensical to set it above that level
You suggest that people set their critical levels based on their personal preferences about their own lives, but then you make claims about their choices based on your intuitions about global properties like the Repugnant Conclusion, with no link between the two
The article makes much about avoiding repugnant sadistic conclusion, but the view you seem to endorse at the end would support creating arbitrary numbers of lives consisting of nothing but intense suffering to prevent the existence of happy people with no suffering who set their critical level to an even higher level than the actual one
On the first point, you suggest that that individuals get to set their own critical levels based on their preferences about their own lives. E.g.
The lowest preferred critical level is zero: if a person would choose a negative critical level, that person would accept a situation where he or she can have a negative utility, such as a life not worth living. Accepting a situation that one would not prefer, is basically a contradiction.
So if my desires and attitudes are such that I set a critical level well below the maximum, then my life can add substantial global value. E.g. if A has utility +5 and sets critical value 0, B has utility +5 and chooses critical value 10, and C has utility −5 and critical value 10, then 3 lives like A will offset one life like C, and you can get most of the implications of the total view, and in particular an overwhelmingly high value of the future if the future is mostly populated with beings who favor existing and set low critical levels for themselves (which one could expect from people choosing features of their descendants or selection).
On the second point, returning to this quote:
The lowest preferred critical level is zero: if a person would choose a negative critical level, that person would accept a situation where he or she can have a negative utility, such as a life not worth living. Accepting a situation that one would not prefer, is basically a contradiction.
I would note that utility in the sense of preferences over choices, or a utility function, need not correspond to pleasure or pain. The article is unclear on the concept of utility it is using but the above quote seems to require a preference base, i.e. zero utility is defined as the point at which the person would prefer to be alive rather than not. But then if 0 is the level at which one would prefer to exist, isn’t it equally contradictory to have a higher critical level and reject lives that you would prefer? Perhaps you are imagining someone who thinks ‘given that I am alive I would rather live than die, but I dislike having coming into existence in the first place, which death would not change.’ But in this framework that would just be negative utility part of the assessment of the overall life (and people without that attitude can be unbothered).
Regarding the third point, if each of us choose our own critical level autonomously, I do not get to decree a level for others. But the article makes several arguments that seem to conflate individual and global choice by talking about everyone choosing a certain level, e.g.:
If people want to move safely away from the sadistic repugnant conclusion and other problems of rigid critical level utilitarianism, they should choose a critical level infinitesimally close to (but still below) their highest preferred levels.
But if I set a very high critical level for myself, that doesn’t lower the critical levels of others, and so the repugnant conclusion can proceed just fine with the mildly good lives of those who choose low critical levels for themselves. Having the individuals choose for themselves based on prior prejudices about global population ethics also defeats the role of the individual choice as a way to derive the global conclusion. I don’t need to be a total utilitarian in general to approve of my existence in cases in which I would prefer to exist.
Lastly, a standard objection to critical level views is that they treat lives below the critical level (but better than nothing by the person’s own lights and containing happiness but not pain) as negative, and so will endorse creating lives of intense suffering by people who wish they had never existed to prevent the creation of multiple mildly good lives. With the variable critical level account all those cases would still go through using people who choose high critical levels (with the quasi-negative view, it would favor creating suicidal lives of torment to offset the creation of blissful beings a bit below the maximum). I don’t see that addressed in the article.
I don’t think you can define aging research so narrowly and get the same expected impact. E.g. De Grey’s SENS includes curing cancer as one of many subgoals, and radical advances in stem cell biology and genetic engineering, massive fields that don’t fall under ‘aging research.’ The more dependent progress in an area is advances from outside that field, the less reliable this sort of projection will be.