AI safety researcher
Thomas Kwa
Karma: 3,552
It is not clear to me that EA branding is net positive for the movement overall or if it’s been tarnished beyond repair by various scandals. Like, it might be that people should make a small personal sacrifice to be publicly EA, but it might also be that the pragmatic collective action is to completely rebrand and/or hope that EA provides a positive radical flank effect.
The reputation of EA at least in the news and on Twitter is pretty bad; something like 90% of the news articles mentioning EA are negative. I do not think it inherently compromises integrity to not publicly associate with EA even if you agree with most EA beliefs, because people who read opinion pieces will assume you agree with everything FTX did, or are a Luddite, or have some other strawman beliefs. I don’t know whether EAF readers calling themselves EAs would make others’ beliefs about their moral stances more or less accurate.
I don’t think this is currently true, but if the rate of scandals continues, anyone holding on to the EA label would be suffering from the toxoplasma of rage, where the EA meme survives by sounding slightly good to the ingroup but extremely negative to anyone else. Therefore, as someone who is disillusioned with the EA community but not various principles, I need to see some data before owning any sort of EA affiliation, to know I’m not making some anti-useful sacrifice.
Given the Guardian piece, inviting Hannania to Manifest seems like an unforced error on the part of Manifold and possibly Lightcone. This does not change because the article was a hitpiece with many inaccuracies. I might have more to say later.
I want to slightly push back against this post in two ways:
I do not think longtermism is any sort of higher form of care or empathy. Many longtermist EAs are motivated by empathy, but they are also driven by a desire for philosophical consistency, beneficentrism and scope-sensitivity that is uncommon among the general public. Many are also not motivated by empathy—I think empathy plays some role for me but is not the primary motivator? Cold utilitarianism is more important but not the primary motivator either [1]. I feel much more caring when I cook dinner for my friends than when I do CS research, and it is only because I internalize scope sensitivity more than >99% of people that I can turn empathy into any motivation whatsoever to work on longtermist projects. I think that for most longtermists, it is not more empathy, nor a better form of empathy, but the interaction of many normal (often non-empathy) altruistic motivators and other personality traits that makes them longtermists.
Longtermists make tradeoffs between other common values and helping vast future populations that most people disagree with, and without ideosyncratic EA values there is no reason that a caring person should make the same tradeoffs as longtermists. I think the EA value of “doing a lot more good matters a lot more” is really important, but it is still trading off against other values.
Helping people closer to you / in your community: many people think this has inherent value
Beneficentrism: most people think there is inherent value in being directly involved in helping people. Habitat for Humanity is extremely popular among caring and empathic people, and they would mostly not think it is better to make more of an overall difference by e.g. subsidizing eyeglasses in Bangladesh.
Justice: most people think it is more important to help one human trafficking victim than one tuberculosis victim or one victim of omnicidal AI if you create the same welfare, because they place inherent value on justice. Both longtermists and GiveWell think they’re similarly good modulo secondary consequences and decision theory.
Discount rate, risk aversion, etc.: There is no reason that having a 10% chance of saving 100 lives in 6,000 years is better than a 40% chance of saving 5 lives tomorrow, if you don’t already believe in zero-discount expected value as the metric to optimize. The reason to believe in zero-discount expected value is a thought experiment involving the veil of ignorance, or maybe the VNM theorem. It is not caring doing the work here because both can be very caring acts, it is your belief in the thought experiment connecting your caring to the expected value.
In conclusion, I think that while care and empathy can be an important motivator to longtermists, and it is valid for us to think of longtermist actions as the ultimate act of care, we are motivated by a conjunction of empathy/care and other attributes, and it is the other attributes that are by far more important. For someone who has empathy/care and values beneficentrism and scope-sensitivity, preventing an extinction-level pandemic is an important act of care; for someone like me or a utilitarian, pandemic prevention is also an important act. But for someone who values justice more, applying more care does not make them prioritize pandemic prevention over helping a sex trafficking victim, and in the larger altruistically-inclined population, I think a greater focus on care and empathy conflict with longtermist values more than they contribute.
[1] More important for me are: feeling moral obligation to make others’ lives better rather than worse, wanting to do my best when it matters, wanting future glory and social status for producing so much utility.
Not sure how to post these two thoughts so I might as well combine them.
In an ideal world, SBF should have been sentenced to thousands of years in prison. This is partially due to the enormous harm done to both FTX depositors and EA, but mainly for basic deterrence reasons; a risk-neutral person will not mind 25 years in prison if the ex ante upside was becoming a trillionaire.However, I also think many lessons from SBF’s personal statements e.g. his interview on 80k are still as valid as ever. Just off the top of my head:
Startup-to-give as a high EV career path. Entrepreneurship is why we have OP and SFF! Perhaps also the importance of keeping as much equity as possible, although in the process one should not lie to investors or employees more than is standard.
Ambition and working really hard as success multipliers in entrepreneurship.
A career decision algorithm that includes doing a BOTEC and rejecting options that are 10x worse than others.
It is probably okay to work in an industry that is slightly bad for the world if you do lots of good by donating. [1] (But fraud is still bad, of course.)
Just because SBF stole billions of dollars does not mean he has fewer virtuous personality traits than the average person. He hits at least as many multipliers than the average reader of this forum. But importantly, maximization is perilous; some particular qualities like integrity and good decision-making are absolutely essential, and if you lack them your impact could be multiplied by minus 20.
[1] The unregulated nature of crypto may have allowed the FTX fraud, but things like the zero-sum zero-NPV nature of many cryptoassets, or its negative climate impacts, seem unrelated. Many industries are about this bad for the world, like HFT or some kinds of social media. I do not think people who criticized FTX on these grounds score many points. However, perhaps it was (weak) evidence towards FTX being willing to do harm in general for a perceived greater good, which is maybe plausible especially if Ben Delo also did market manipulation or otherwise acted immorally.
Also note that in the interview, SBF didn’t claim his donations offset a negative direct impact; he said the impact was likely positive, which seems dubious.
This seems right, thanks. I don’t think we have positive evidence that Trabucco was not EA, though.
[Warning: long comment] Thanks for the pushback. I think converting to lives is good in other cases, especially if it’s (a) useful for judging effectiveness, and (b) not used as a misleading rhetorical device [1].
The basic point I want to make is that all interventions have to pencil out. When donating, we are trying to maximize the good we create, not decide which superficially sounds better between the different strategies “empower beneficiaries to invest in their communities’ infrastructure” and “use RCTs to choose lifesaving interventions” [2]. Lives are at stake, and I don’t think those lives are less important simply because it’s harder to put names and faces to the ~60 lives that were saved from a 0.04% chance of reduction of malaria deaths from a malaria net. Of course this applies equally to the Wytham Abbey purchase or anything else. But to point (a), we actually can compare the welfare gain from 61 lives saved to the economic security produced by this project. GiveWell has weights for doubling of consumption, partly based on interviews from Africans [3]. With other projects, this might be intractable due to entirely different cause areas or different moral preferences e.g. longtermism.
Imagine that we have a cost-effectiveness analysis made by a person with knowledge of local conditions and local moral preferences, domain expertise in East African agricultural markets, and the quantitative expertise of GiveWell analysts. If it comes out that one intervention is 5 or 10 times better than the other, as is very common, we need a very compelling reason why some consideration was missed to justify funding the other one. Compare this to our currently almost complete state of ignorance as to the value of building this plant, and you see the value of numbers. We might not get a CEA this good, but we should get close as we have all the pieces.
As to point (b), I am largely pro making these comparisons in most cases just to remind people of the value of our resources. But I feel like the Wytham and HPMOR cases, depending on phrasing, could exploit peoples’ tendency to think of projects that save lives in emotionally salient ways as better than projects that save lives via less direct methods. It will always sound bad to say that intervention A is funded rather than saving X lives, and we should generally not shut down discussion of A by creating indignation. This kind of misleading rhetoric is not at all my intention; we all understand that allowing a large enough number of farmers access to sorghum markets can produce more welfare than preventing 61 deaths from malaria. We have the choice between saving 61 of someones’ sons and daughters, and allowing X extremely poor people to perhaps buy metal roofs, send their children to school, and generally have some chance of escaping a millennia-long poverty trap. We should think: “I really want to know how large X is”.
[1] and maybe (c) not bad for your mental health?
[2] Unless you believe empowering people is inherently better regardless of the relative cost, which I strongly disagree with.
[3] This is important—Westerners may be biased here because we place different values on life compared to doubling consumption. But these interviews were from Kenya and Ghana, so maybe Uganda’s weights slightly differ.
Just to remind everyone, 339,000 GBP in malaria nets is estimated by GiveWell to save around 61 lives, mostly young children. Therefore a 25% difference in effectiveness either way is 15 lives. A cost-effectiveness analysis is definitely required given what is at stake, even if the complexities of this project mean it is not taken as final.
Thanks. In addition to lots of general information about FTX, this helps answer some of my questions about FTX: it seems likely that FTX/Alameda were never massively profitable except for large bets on unsellable assets (anyone have better information on this?); even though they had large revenues maybe much of it was spent dubiously by SBF. And the various actions needed to maintain a web of lies indicate that Caroline Ellison and Nishad Singh, and very likely Gary Wang and Sam Trabucco (who dropped off the face of the earth at the time of the bankruptcy [1]) were definitely complicit in fraud severe and obvious enough that any moral person, (possibly even a hardcore utilitarian, if it was true that FTX was consistently losing money), should have quit or leaked evidence of said fraud.
Four or five people is very different from a single bad actor, and this almost confirms for me that FTX belongs on the list of ways EA and rationalist organizations can basically go insane in harmful ways, alongside Leverage, Zizians and possibly others. It is not clear that FTX experienced a specifically EA failure mode, rather than the very common one in which power corrupts.
I think someone should do an investigation much wider in scope than what happened at FTX, covering the entire causal chain from SBF first talking to EAs at MIT to the damage done to EA. Here are some questions I’m particularly curious about:
Did SBF show signs of dishonesty early on at MIT? If so, why did he not have a negative reputation among the EAs there?
To what extent did EA “create SBF”—influence the values of SBF and others at FTX? Could a version of EA that placed more emphasis on integrity, diminishing returns to altruistic donations, or something else have prevented FTX?
Alameda was started by various traders from Jane Street, especially EAs. Did they do this despite concerns about how the company would be run, and were they correct to leave at the time?
[edited to add] I have heard that Tara Mac Aulay and others left Alameda in 2018. Mac Aulay claims this was “in part due to concerns over risk management and business ethics”. Do they get a bunch of points for this? Why did this warning not spread, and can we even spread such warnings without overloading the community with gossip even more than it is?
Were Alameda/FTX ever highly profitable controlling for the price of crypto? (edit: ruling out that FTX’s market share was due to artificially tight spreads due to money-losing trades from Alameda). How should we update on the overall competence of companies with lots of EAs?
SBF believed in linear returns to altruistic donations (I think he said this on the 80k podcast), unlike most EAs. Did this cause him to take on undue risk, or would fraud have happened if FTX had a view on altruistic returns similar to that of OP or SFF but linear moral views?
What is the cause of the exceptionally poor media perception of EA after FTX? When i search for “effective altruism news”, around 90% of articles I could find negative and none positive, including many with extremely negative opinions unrelated to FTX. One would expect at least some article saying “Here’s why donating to effective causes is still good”. (In no way do I want to diminish the harms done to customers whose money was gambled away, but it seems prudent to investigate the harms to EA per se)
My guess is that this hasn’t been done simply because it’s a lot of work (perhaps 100 interviews and one person-year of work), no one thinks it’s their job, and conducting such an investigation would somewhat entail someone both speaking for the entire EA movement and criticizing powerful people and organizations.
See also: Ryan Carey’s comment
2-year update on infant outreach
To our knowledge, there have been no significant infant outreach efforts in the past two years. We are deeply saddened by this development, because by now there could have been two full generations of babies, including community builders who would go on to attract even more talent. However, one silver lining is that no large-scale financial fraud has been committed by EA infants.
We think the importance of infant outreach is higher than ever, and still largely endorse this post. However, given FTX events, there are a few changes we would make, including a decreased focus on galactic-scale ambition and especially some way to select against sociopathic and risk-seeking infants. We tentatively propose that future programs favor infants who share their toys, are wary of infants who take others’ toys without giving them back, and never support infants who, when playing with blocks, try to construct tall towers that have high risk of collapse.
This post is important and I agree with almost everything it says, but I do want to nitpick one crucial sentence:
There may well come a day when humanity would tear apart a thousand suns in order to prevent a single untimely death.
I think it is unlikely that we should ever pay the price of a thousand suns to prevent one death, because tradeoffs will always exist. The same resources used to prevent that death could support trillions upon trillions of sentient beings at utopic living standards for billions of years, either biologically or in simulation. The only circumstances where I think such a decision would be acceptable are things like
The “person” we’re trying to save is actually a single astronomically vast hivemind/AI/etc that runs on a star-sized computer and is worth that many resources.
Our moral views at the time dictate that preventing one death now is at least fifteen orders of magnitude worse than extending another being’s life by a billion years.
The action is symbolic, like how in The Martian billions of dollars were spent to save Mark Watney, rather than driven by cause prioritization.
Otherwise, we are always in triage and always will be, and while prices may fluctuate, we will never be rich enough to get everything we want.
My study of the monkeys and infants, i.e. my analysis of past wars, suggested an annual extinction risk from wars of 6.36*10^-14, which is still 1.07 % (= 5.93*10^-12/(5.53*10^-10)) of my best guess.
The fact that one model of one process gives a low number doesn’t mean the true number is within a couple orders of magnitude of that. Modeling mortgage-backed security risk in 2007 using a Gaussian copula gives an astronomically low estimate of something like 10^-200, even though they did in fact default and cause the financial crisis. If the bankers adjusted their estimate upward to 10^-198 it would still be wrong.
IMO it is not really surprising for very near 100% of the risk of something to come from unmodeled risks, if the modeled risk is extremely low. Like say I write some code to generate random digits, and the first 200 outputs are zeros. One might estimate this at 10^-200 probability or adjust upwards to 10^-198, but the probability of this happening is way more than 10^-200 due to bugs.
Don’t have time to reply in depth, but here are some thoughts:
If a risk estimate is used for EA cause prio, it should be our betting odds / subjectie probabilities, that is, average over our epistemic uncertainty. If from our point of view a risk is 10% likely to be >0.001%, and 90% likely to be ~0%, this lower bounds our betting odds at 0.0001%. It doesn’t matter that it’s more likely to be 0%.
Statistics of human height are much better understood than nuclear war because we have billions of humans but no nuclear wars. The situation is more analogous to finding the probability of a 10 meter tall adult human having only ever observed a few thousand monkeys (conventional wars), plus one human infant (WWII) and also knowing that every few individuals humans mutate into an entirely new species (technological progress).
It would be difficult to create a model suggesting a much higher risk because most of the risk comes from black swan events. Maybe one could upper bound the probability by considering huge numbers of possible mechanisms for extinction and ruling them out, but I don’t see how you could get anywhere near 10^-12.
Any probability as low as 5.93*10^-12 about something as difficult to model as the effects of nuclear war on human society seems extremely overconfident to me. Can you really make 1⁄5.93*10^-12 (170 billion) predictions about independent topics and expect to be wrong only once? Are you 99.99% [edit: fixed this number] sure that there is no unmodeled set of conditions under which civilizational collapse occurs quickly, which a nuclear war is at least 0.001% likely to cause? I think the minimum probabilities that one should have given these considerations is not much lower than the superforecasters’ numbers.
There was likely no FTX polycule (a Manifold question resolved 15%) but I was aware that the FTX and Alameda CEOs were dating. I had gone to a couple of FTX events but try to avoid gossip, so my guess is that half of the well-connected EAs had heard gossip about this.
Being attention-getting and obnoxious probably paid off with slavery because abolition was tractable. But animal advocacy is different. I think a big question is whether he was being strategic, or just obnoxious by nature? If we put Benjamin Lay in 2000, would he start cage-free campaigns or become PETA? Or perhaps find some angle we’re overlooking?
My comment is not an ad hominem. An ad hominem attack would be if someone is arguing point X and you distract from X by attacking their character. I was questioning only Remmelt’s ability to distinguish good research from crankery, which is directly relevant to the job of an AISC organizer, especially because some AISC streams are about the work in question by Forrest Landry. I apologize if I was unintentionally making some broader character attack. Whether it’s obnoxious is up to you to judge.
Crossposted from LessWrong.
Maybe I’m being cynical, but I’d give >30% that funders have declined to fund AI Safety Camp in its current form for some good reason. Has anyone written the case against? I know that AISC used to be good by talking to various colleagues, but I have no particular reason to believe in its current quality.
MATS has steadily increased in quality over the past two years, and is now more prestigious than AISC. We also have Astra, and people who go directly to residencies at OpenAI, Anthropic, etc. One should expect that AISC doesn’t attract the best talent.
If so, AISC might not make efficient use of mentor / PI time, which is a key goal of MATS and one of the reasons it’s been successful.
Why does the founder, Remmelt Ellen, keep linkposting writing by Forrest Landry which I’m 90% sure is obvious crankery? It’s not just my opinion; Paul Christiano said “the entire scientific community would probably consider this writing to be crankery”, one post was so obviously flawed it gets −46 karma, and generally the community response has been extremely negative. Some AISC work is directly about the content in question. This seems like a concern especially given the philosophical/conceptual focus of AISC projects, and the historical difficulty in choosing useful AI alignment directions without empirical grounding. [Edit: To clarify, this is not meant to be a character attack. I am concerned that Remmelt does not have the skill of distinguishing crankery from good research, even if he has substantially contributed to AISC’s success in the past.]
All but 2 of the papers listed on Manifund as coming from AISC projects are from 2021 or earlier. Because I’m interested in the current quality in the presence of competing programs, I looked at the two from 2022 or later: this in a second-tier journal and this in a NeurIPS workshop, with no top conference papers. I count 52 participants in the last AISC so this seems like a pretty poor rate, especially given that 2022 and 2023 cohorts (#7 and #8) could both have published by now. (though see this reply from Linda on why most of AISC’s impact is from upskilling)
The impact assessment was commissioned by AISC, not independent. They also use the number of AI alignment researchers created as an important metric. But impact is heavy-tailed, so the better metric is value of total research produced. Because there seems to be little direct research, to estimate the impact we should count the research that AISC alums from the last two years go on to produce. Unfortunately I don’t have time to do this.
Surely as their gold standard “career change” pin-up story, they could find a higher EV career change.
You’re assuming that the EV of switches from global health to biosecurity is lower than the EV of switching from something else to biosecurity. Even though global health is better than most cause areas, this could be false in practice for at least two reasons
If the impact of biosecurity careers is many times higher than the impact of global health, and people currently in global health are slightly more talented, altruistic, or hardworking.
If people currently in global health are not doing the most effective global health interventions.
Suppose that the EA community were transported to the UK and US in 1776. How fast would slavery have been abolished? Recall that the slave trade ended in 1807 in the UK and 1808 in the US, and abolition happened between 1838-1843 in the British Empire and 1865 in the US.
Assumptions:
Not sure how to define “EA community”, but some groups that should definitely be included are the entire staff of OpenPhil and CEA, anyone who dedicates their career choices or donates more than 10% along EA principles, and anyone with >5k EA forum karma.
EAs have the same proportion of the population as they do now, as well as the same relative levels of wealth, political power, intelligence, and drive.
EAs forget all our post-1776 historical knowledge, including the historical paths to abolition.
EA attention is split among other top causes of the day, like infectious disease and crop yields. I can’t think of a reason why antislavery would be totally ignored by EAs though, as it seems huge in scope and highly morally salient to people like Bentham.
I’m also interested in speculating on other causes, I’ve just been thinking about abolition recently due to the 80k podcast with Prof. Christopher Brown.
Note that (according to ChatGPT) Quakers were more dedicated to abolition than EAs are to animal advocacy, have a much larger population, and deserve lots of moral credit for abolition in real life. But my guess would be that EAs could find some angles the Quakers wouldn’t due to the consequentialist principles of EA. Maybe more evangelism and growth (Quaker population declined in the early 1800s), pragmatism about compensating slaveholders in the US as was done in the UK, or direct political action. Could EAs have gotten the Fugitive Slave Clause out of the Constitution?