Follow me on hauke.substack.com
I’m an independent researcher working on AI and other EA topics (Global Priorities Research and Economics).
Hauke Hillebrandt
Karma: 3,844
I did a shallow review of the evidence for ACT last year: “Anxiety defusion and acceptance (acceptance and commitment therapy) Mind Ease’s anxiety defusion exercise is based on acceptance and commitment therapy (ACT), which is backed by the following evidence: Traditional ACT with a therapist: A 2017 review of RCTs of ACT to treat anxiety and depression shows that ACT improves depression relative to no treatment up to 6-months follow-up. (ds = 0.32 to 1.18). Two studies compared ACT with minimally active comparison conditions (expressive writing and minimal support group) and found ACT outperformed comparison conditions on depression at post, but were equivalent at follow-up.[74] A 2020 meta-analysis of 18 studies with 1,088 participants showed that ACT significantly reduced depression as compared with the control group (d= 0.59, 95% CI [0.38, 0.81]).[75] Self-help: Traditionally face-to-face, ACT is also delivered in self-help formats. A meta-analysis shows that ACT self-help showed significant small effect sizes favoring intervention for depression (g=0.34; 95% CIs [0.07, 0.61]; Z=2.49, p=0.01) and anxiety (g=0.35; 95% CIs [0.09, 0.60]; Z=2.66, p=0.008). Higher levels of clinician guidance improved outcomes but intervention format (e.g. book/computer) was unlikely to moderate results.[76] Internet-based ACT (iACT): A systematic review of internet-delivered ACT (iACT) for anxiety[77] showed that 18 out 20 studies reported significant anxiety reduction after treatment. This was observed in studies that delivered iACT with (n=13) or without (n=5) therapist guidance. The average attrition rate during treatment was 19%. In 13 studies participants on average rated their iACT experience with above average to high treatment satisfaction. App-based ACT: A recent RCT of ACT in an app form showed that help-seeking individuals vs. waitlist increased well-being with moderate effect sizes.[78] In aggregate, anxiety defusion and acceptance (acceptance and commitment therapy) seems effective with small to medium effect sizes.”
Cause area: Developmental Cognitive Neuroepidemiology
The Global Catastrophic Risk Management Act of 2022 is a new bipartisan bill that was proposed recently and is going to be voted on in the US. There’s another bill on WMDs.
Thanks for the link—I think the economists surveyed were not unanimous in saying that it’s a slam dunk win, and as I wrote ‘might’ and ‘big, if true’ - also note that I’m citing a link from the very left-wing think tank associated with the German Green party.
Also see that while the case for immigration boosting the economy in the long-run is strong based on economic theory, there might still be upfront cost that could have bad effect such as displacing traditional aid:
https://www.givingwhatwecan.org/blog/using-aid-to-finance-the-refugee-crisis-a-worrying-trend
It could also be that, a la David Autor’s China shock literature, while the average economic effects of migration are positive some low-skilled domestic workers might have increased competition, which can cause populism. For instance, immigration can predict Brexit votes.
Again: big, if true and there should be more analysis. The main lesson here is if you’re dealing with trillion dollar numbers, it might be very important.
“star-manning is to not only engage with the most charitable version of your opponent’s argument, but also with the most charitable version of your opponent, by acknowledging their good intentions and your shared desires despite your disagreements. In our UBI example, star-manning would be to amend the steel man with something like, “…and you’re in favor of this because you think it will help people lead safer, freer, and more fulfilled lives—which we both want.” If used properly, star-manning can serve as an inoculant against our venomous discourse and a method for planting disputes on common ground rather than a fault line.”
https://centerforinquiry.org/blog/how-to-star-man-arguing-from-compassion/
Thanks for posting this.
“amounts matter” or “let’s actually do cost-benefit analysis” [...] To be clear, “amounts matter” is the usual EA stance already
I think EA is still overemphasizing high benefit-cost ratios, but it’s now better to find high benefit minus cost interventions. In other words, it used to be the case that we wanted to find a way to fill a $1m funding gap to save 1000 lives, and save a life for $1k, but now even though these small funding gaps still exist it’s quite hard to find them at scale, and we might rather want to find a billion dollar funding gap that saves lives at only $10k, but then save 100k lives which is better since amounts matter as you say.
In contrast to a movement trying to push high B/C interventions like EA, an ‘amounts matter’ / B-C movement would have much higher popular appeal as it would directly and personally affect many more people.
A few things that this movement might highlight (all ‘big, if true’):
Covid: As you say, global GDP losses from COVID are massive already $12T and 3 billion annual vaccine courses have a global benefit of $17T—a benefit of ~$576 to $989 per course. at a cost of $6 to $40 per course—and so the benefit cost ratios might not actually be as low as on the order of 10x, yet still then total benefits would be enormous.
Ukraine war might also cost the global economy $0.5T (or more)
2008 financial crisis: $6-14T just for the U.S. ($50k-$120k for every U.S. household) and caused 500,000 cancer deaths worldwide, so macroprudential and generally macroeconomic policy might have huge effects
There are some public health things probably roughly on a similar level (smoking, obesity, etc.).
On some level, politics is already doing this, but I think there’s still a lot of scope insensitivity and not concrete focus on these issues.
These are good references—I’m especially interested in arxiv-vanity, are you talking about Ben Firshman? I’ll reach out once I start working there
Yes
Congrats on the job! Seems really high impact. A few thoughts:
Scientometrics
Scientometrics is the field of study which concerns itself with measuring and analysing scientific literature such as the impact of research papers and academic journals. “Of course, such tools cannot substitute for substantive knowledge of human experts, but they can be used as powerful decision support systems to structure humans’ effort and augment their capabilities/efficiency to handle the enormous volume of data on research input and output” Finding rising stars in bibliometric networks | SpringerLink Scientometric indicators and machine learning-based models can be used to predict the ‘rising stars’ in academia , —identifying these junior researchers and awarding them prizes would greatly improve research output. https://ieeexplore.ieee.org/abstract/document/8843686/.
2. https://allenai.org/ has both semantic scholar and an NLP AI team—I think there are overlaps wrt using the arxiv corpus for language models.
3. The creator of https://www.arxiv-vanity.com/ is also really interested in EA- maybe get in tocuh
Very cool—law seems underexplored.
Gates has a legal fund to help countries fight big tobacco—towards legal advice for nations whose health measures are challenged by tobacco industry, as in Uruguay and Australia. The mere existence of such funds might have be a credible threat or deterrence.
There might be other examples of high-impact law e.g. scientists are sometimes sued by companies for publishing the truth: e.g. “In the 1980s, various interests tried to suppress the work of Dr. Herbert Needleman on the effects of lead exposure.′ c.f. EA project arising as a result
Or: One of the company selling stimultants sued a Harvard Medical School Prof for “$200m in damages for libel, alleging that statements in the peer review article, and subsequent interviews with the media, were false.” more here see ‘Why a Lot of Important Research Is Not Being Done’
As a layperson, common law seems generally much more elegant legal system and well-suited for EA and unknown emerging risks.
My naive simplistic view of civil law is: rules are written down quite explicitly and it’s a bit more deontological (and that’s why you have it in say Germany where Kant came from) e.g. the law says specifically ‘You aren’t allowed to use an algorithm that discriminates based on age’. But if you use an black box algorithm that discriminates based on something else, or outsource your hiring to a foreign firm that does the discriminating for you, then you’re off the hook. When that behavior gets out of hand, the law has to be painfully rewritten, they try to generalize but it’s hard, resulting in a crazy complicated legal corpus, and EVEN if you’re found guilty you get a fine must have been stipulated in advance in the law, which is often not proportional to the crime and there’s little deterrent effect. For instance, German courts don’t use punitive damages and people seem confused by them and their usefulness (see McDonald’s lawsuit).
Whereas in common law seems more consequentialist / utilitarian (and comes from the UK where Bentham comes from): if you show that there’s precedent of someone having done ~similar harm to before, then there’ll often be punitive damages in proportion to the crime for deterrence.
Similar to large settlements in Big Pharma, Big Tech has been fined >$30bn in recent years. Consider that the EU has fined Google $10bn. The EU seems to use ~case law and there’s the Brussel’s Effect, which might be very high leverage (A new UK regulator is said that it will also have “the power” to fine tech companies up to 10% of their global turnover if they fail to comply). This is interesting for slow take-off / ~prosaic AI safety / risks from malevolent actors reasons.
It’s more elegant as it makes people and corporations generally be more on guard about misbehaving for fear of being sued. Punitive damages are also theoretically equivalent to specific kind of Pigovian tax on externalities (which seems much better than traditional corporate tax and I’m against tort reform arguments and I’d hate to see caps on damages “Many state statutes are the result of insurance industry lobbying to impose “caps” on punitive damages; however, several state courts have struck down these statutory caps as unconstitutional.”)
I guess currently this is disbursed to go into the general budget. But maybe one could fine corporations in stock and use the dividends to fund the regulators, so they’re incentivized to reduce negative externalities (through fining companies), but then also they’d be held back to completely wreck industries or companies, because they’re financed by the overall health of the industry after the fine.
The following is not financial advice.
Sam Bowman who’s EA adjacent has written about this.
Wise actually has a bank account that you can invest in an index for UK customers only w/ 0.5% fees.
Note that investing in the stock market is generally seen as riskier than holding a major country’s currency.
Also inflation is unusually high and markets expect it to come down and currencies are unusually volatile currently:
“Rock-bottom inflation and interest rates over the past decade helped smother swings in exchange rates. Deutsche Bank’s cvix index, a gauge of forex volatility, has been above its current level more than 90% of the time over the past 20 years. By contrast, the vix, which measures expected volatility for America’s s&p 500 index of stocks and is often used as a measure of overall market sentiment, has so far spent October at roughly its long-term average.” [src]
Policy makers use both fiscal and monetary policy to improve domestic welfare:
Fiscal policy: An example is increasing tax on the rich people who earn >$100k/y, and then redistribute it to the poor people who earn $10k/y as welfare payment (or earned income tax breaks). At a first approximation, the cost of this policy might be $100B, but the utility costs are less as of the diminishing returns to utility and the benefits are larger. This is crazy scalable—you can easily send every person a stimulus check and spend a lot of money that way. However, generally, the fiscal multiplier is small, because people can’t use capital as effectively as firms. Transfer payments multiplier are usually smaller than government spending multipliers. Money can also be spent on public goods like health or education with higher sROI, but less scalability and steeper diminishing returns.
Monetary policy: An example is printing $1T, then loaning it out to private banks at a below-market interest rate (say 0% though it could be negative). If the government would invest in an index, it could get a risk-adjusted return of 5% per year, which it is losing out of. Thus, the cost of the policy is $50B. The direct cost being distributed across the whole population due to inflation, which hits the poorest disproportionately, and thus having high cost in terms of utility. However, the fiscal multiplier of such a policy might be very high: banks are incentivized to recoup their investment and will loan to firms that succeed. The fiscal multiplier is generally higher, with higher social surplus, which creates jobs, and increases growth. Also, lower interest rates generally lower unemployment and increase wages. However, people who are not very rich only benefit slowly through trickle down effects.
Why would the state make concessional loans and lose money? Why not loan at market rates? For this would have no additionality, the state would just be another lender. These concessional loans pay for the service that banks provide to the state. What is the service? Banks find companies that will perform well, so that the state is not in the business of ‘picking winners’.
Analogously, policy makers use both aid and development finance to improve foreign welfare:
Aid: An example of UNICEF’s humanitarian aid (~$26B/y) like disaster relief in cash or in kind (e.g. food aid, short-term reconstruction relief). Or they can spend on public goods like social infrastructure and services ($65B/y) to develop the human resource potential and improve living conditions. Health aid like malaria nets is$12.9bn, $6.2bn was spent on perhaps higher leveragetechnical assistance. Half of all $180B/y in aid is spent like this.[1] But the fiscal multiplier is generally considered small (see General equilibrium effects of cash transfers).
Development finance: An example of development finance is giving out loans through development banks. For instance, the World Bank’s International Finance Corporation,[2] loans to foreign private firms and public private partnerships. Development finance institutions give out ~$50B/y.[3],[4]
Like central banks, development banks also give loans to poor countries’ governments, like the World Bank’s International Development Association, which offers concessional loans to poor countries.[5] These concessional loans have the benefit of feeling altruistic, creating strategic alliances, or spurring (global) growth.
The later (growth-friendly spending) is mostly through IMF and World Bank loans for countries that they need to pay back (analogous to impact investing), whereas aid is analogous to donating to nonprofits. The great thing about loans is that you need to pay them back with interest and so only if you’re certain you’ll create growth and a multiplier, you will take them, loans are the most direct way to subsidize business activity. They’re also very scalable. Sure you can subsidize the latest thing like chickens or therapy or whatever, but even Blattman admits money for entrepreneurs is best, so why not add skin in the game and make it a loan and go directly to the source of the problem (lack of growth)?
There is an allure to policy coherence and optimizing for several objectives at once, finding an intervention that has the best of both worlds, but it would be a suspicious convergence if the best poverty reduction methods happened to be the effective at creating growth and entrepreneurship as well. Giving directly to the poorest at scale AND having a high fiscal multiplier i.e. making people productive to create growth. The Tinbergen Rule is a basic principle of effective policy, which states that to achieve n independent policy targets you need at at least n independent policy instruments. And so people want microfinance to work at scale for poor people in poor countries, and think UBI will create many entrepreneurs and you get massive productivity at scale and a massive fiscal multiplier, but some European welfare states have de facto UBI and cuddly capitalism hasn’t created crazy growth. Not saying it’s bad for welfare reasons.
If you want anything outside this continuum, then I’d be more excited about subsidizing highly scalable, zero marginal cost (global) public goods. These could even be for entrepreneurship, like Moskovitz & Collison subsidizing Asana and Stripe Atlas for poor countries—that’d be really effective.
Seems like the opposite of crewing where “each person would bring some problem they were working on to the group, and receive 90 minutes of undivided attention from their peers”—maybe it could be combined.
You allocate 1⁄4 to new interventions (presumably higher-risk/higher-reward) - do you agree with OpenPhil’s post you link to and are you going to fund those preferably given the funding shortfall?
Do you agree OpenPhil that Givewell’s interventions have ~the same CBA during economic crises? For instance, AMF is now expanding to Nigeria, where GDP/capita gone down to 2008 levels and generally Malaria deaths are up ~10% due to Covid. Does this increase the CBA of your core interventions? Relatedly: $150m seems quite a large reduction—OpenPhil considered using mission hedging at the last year’s EAG—have you considered this for your assets?
Cf. I always recommend this excellent philosophy paper “On the aptness of anger”.
I blog using Google Docs + Google sites—I just write a Google doc, then push it into the public folder:
removing the trivial inconvenience of reformatting and hitting a publish button has definitely made me publish more posts.
related: Imagen replicating DALL-E very well, seems like good evidence that there’s healthy competition between big tech companies, which drives down profits.
One thing that might push against this are economies of scope and if data really does become the new oil and become more relevant over time.
Will China launch a full scale invasion of Taiwan before 2025? Currently at 9%
Will China launch a full-scale invasion of Taiwan before 2030? Currently at 25%
Will China launch a full-scale invasion of Taiwan before 2035? Currently at 39%
If China launches a full-scale invasion of Taiwan before 2035, will the US respond militarily? Currently at 66%
If China launches an invasion of Taiwan before 2035, and the US intervenes, will China attack the United States? Currently at 60%
If China launches a full-scale invasion of Taiwan before 2035, will they successfully control Taiwan within three years? Currently at 56% (I’m at 75% personally)
Multiplying this out, the joint probabilities for hot US-Sino war in 3, 7, and 12 years, are thus, 4%, 10%, and 15% respectively.
You could submit it as question to his podcast!
Thanks! Similarly, I’m enjoying https://simpl.fyi/ which simplifies gmails design.
I’m not an expert on this but here are a few links:
https://www.cambridge.org/core/journals/proceedings-of-the-nutrition-society/article/is-iodine-deficiency-still-a-problem-in-subsaharan-africa-a-review/6C87E944AF05DEE3B7821D986D2F1B77
https://www.givewell.org/international/technical/programs/salt-iodization
https://www.openphilanthropy.org/grants/iodine-global-network-general-support-december-2020/