Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Introducing the Center for Effective Aid Policy (CEAP), published by MathiasKB on December 23, 2022 on The Effective Altruism Forum.

We are incredibly excited to announce the launch of the Center for Effective Aid Policy (CEAP), a new non-profit incubated by Charity Entrepreneurship. Our mission is to improve the effectiveness of development aid. We will do so through policy advocacy to governments and INGOs.

If you are unfamiliar with development aid and want to learn more, we wrote an introduction to the field which was posted to the forum last week. In short:

Development aid represents one fourth of all charitable giving worldwide and is not known for its immaculate efficiency and clandestine operations. The cost-effectiveness of many aid projects can be vastly improved and we believe there are many opportunities to do so.

In this post we will go over:

Our near term plans.

Speculative plans for the long term.

How you can help!

We additionally hope sharing our tentative plans can be a step towards greater organizational transparency in the effective altruism community. Some, both inside and outside this community, will disagree that our organization is a good use of resources. Our funding would most likely have gone to highly effective charities counterfactually.

Being held accountable and scrutinized for our decisions, might hurt us in the short run but benefits everyone in the long run.

Many parliaments are surrounded by ‘think-tanks’ who seek to influence policy in directions that just so happen to benefit the industries which are funding them. Decision makers should be free to evaluate our organization’s priorities and decide for themselves if they agree.

Interventions we are excited about

We are entering a well-established field and stand on the shoulders of giants.

There are many organizations with decades of experience doing great work to improve aid effectiveness - from research institutes doing academic research, to organizations solely focused on advocacy.

Our work builds upon the collective research of thousands of academics, practitioners, and policy makers who have worked tirelessly for decades to improve the quality of aid.

As a new organization we have the ability to move fast and break things. Taking risks on one's own behalf is all well and good, but mistakes we make might hurt the efforts of other organizations advocating for cost-effective aid spending.

When we set out to prioritize between the many possible interventions, we looked for aid policies that experts in development and policymakers alike were excited about when interviewed.

We have reached three of interventions that we think look especially promising:

Interventions we want to address in our first year:

Advocacy for cash-benchmarking of aid projects

Cash-benchmarking advocacy was the intervention with the best reception among policymakers and academics we interviewed. Despite this, there is very little information available on cash-benchmarking. It doesn’t even have a wikipedia page! Google’s top result for cash-benchmarking is a one page report by USAID, describing a recent successful experiment they did.

The discrepancy between the possible impact of improved benchmarking, the excitement of decision-makers, and lack of high quality public material is larger than for any other improvement to aid we reviewed. To change this state of affairs we are producing a comprehensive report, which can serve as a safe point of referral for policymakers and advocates.

Advocacy to affect aid cuts

A recent trend in western countries is for governments to cut aid spending. In 2020, the UK government made the decision to cut its aid spending from 0.7 to 0.5% of GNI. In 2022 the newly elected Swedish government cut future spending from 1% to 0.7%.

Governments are also classifying previously non-aid bud...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Read The Sequences, published by Quadratic Reciprocity on December 23, 2022 on The Effective Altruism Forum.

There is heavy overlap among the effective altruism and rationality communities but they are not the same thing. Within the effective altruism community, especially among those who are newer to the movement and were introduced to it through a university group, I've noticed some tension between the two. I often sense the vibe that sometimes people into effective altruism who haven’t read much of the canonical LessWrong content write off the rationalist stuff as weird or unimportant.

I think this is a pretty big mistake.

Lots of people doing very valuable work within effective altruism got interested in it via first interacting with rationalist content, in particular The Sequences and Harry Potter and the Methods of Rationality. I think that is for good reason. If you haven’t come across those writings before, here’s a nudge to give The Sequences a read. The Sequences are a (really long) collection of blog posts written by Eliezer Yudkowsky on the science and philosophy of human rationality. They are divided into sequences - a list of posts on a similar topic. Most of the posts would have been pretty useful to me on their own but I also got more value from reading posts in a particular sequence to better internalise the concepts. There are slightly fewer posts in The Sequences than there are days in the year so reading the whole thing is a very doable thing to do in the coming year! You can also read Highlights from the Sequences which cover 50 of the best essays.

Below, I’ll list some of the parts that I have found especially helpful and that I often try to point to when talking to people into effective altruism (things I wish they had read too).

Fake Beliefs is an excellent sequence if you already know a bit about biases in human thinking. The key insight there is about making beliefs pay rent (“don’t ask what to believe—ask what to anticipate”) and that sometimes your expectations can come apart from your professed beliefs (fake beliefs). The ideas were helpful for me noticing when that happens, for example when I believe I believe something but actually do not. It happens a bunch when I start talking about abstract, wordy things but forget to ask myself what I would actually expect to see in the world if the things I am saying were true.

Noticing Confusion is a cool sequence that talks about things like:

What is evidence? (“For an event to be evidence about a target of inquiry, it has to happen differently in a way that’s entangled with the different possible states of the target”)

Your strength as a rationalist is your ability to be more confused by fiction than by reality - noticing confusion when something doesn’t check out and going EITHER MY MODEL IS FALSE OR THIS STORY IS WRONG

Absence of evidence is evidence of absence, and conservation of expected evidence (“If you expect a strong probability of seeing weak evidence in one direction, it must be balanced by a weak expectation of seeing strong evidence in the other direction”)

I am often surrounded by people who are very smart and say convincing-sounding things all the time. The ideas mentioned above have helped me better recognise when I'm confused and when a smooth-sounding argument doesn't match up with how I think the world actually works.

Against rationalisation has things that are useful to remember:

Knowing about biases can hurt people. Exposing subjects to an apparently balanced set of pro and con arguments will exaggerate their initial polarisation. Politically knowledgeable subjects, because they possess greater ammunition with which to counter-argue incongruent facts and arguments, will be more prone to some biases.

Not to avoid your belief's real weak points. “Ask yourself what smart people...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Katja Grace: Let's think about slowing down AI, published by peterhartree on December 23, 2022 on The Effective Altruism Forum.

On Twitter, Katja Grace wrote:

I think people should think more about trying to slow down AI progress, if they believe it's going to destroy the world soon. I know people have like eighteen reasons to dismiss this idea out of hand, but I dispute them.

The introduction to the post is below. Do read the whole thing.

Consider reading alongside:

Holden Karnofsky's recent posts on Cold Takes, especially "Racing Through a Minefield: the AI Deployment Problem".

Scott Alexander's recent "Perhaps It Is A Bad Thing That The World's Leading AI Companies Cannot Control Their AIs".

Averting doom by not building the doom machine

If you fear that someone will build a machine that will seize control of the world and annihilate humanity, then one kind of response is to try to build further machines that will seize control of the world even earlier without destroying it, forestalling the ruinous machine’s conquest. An alternative or complementary kind of response is to try to avert such machines being built at all, at least while the degree of their apocalyptic tendencies is ambiguous.

The latter approach seems to me like the kind of basic and obvious thing worthy of at least consideration, and also in its favor, fits nicely in the genre ‘stuff that it isn’t that hard to imagine happening in the real world’. Yet my impression is that for people worried about extinction risk from artificial intelligence, strategies under the heading ‘actively slow down AI progress’ have historically been dismissed and ignored (though ‘don’t actively speed up AI progress’ is popular).

The conversation near me over the years has felt a bit like this:

Some people: AI might kill everyone. We should design a godlike super-AI of perfect goodness to prevent that.

Others: wow that sounds extremely ambitious

Some people: yeah but it’s very important and also we are extremely smart so idk it could work

[Work on it for a decade and a half]

Some people: ok that’s pretty hard, we give up

Others: oh huh shouldn’t we maybe try to stop the building of this dangerous AI?

Some people: hmm, that would involve coordinating numerous people—we may be arrogant enough to think that we might build a god-machine that can take over the world and remake it as a paradise, but we aren’t delusional

This seems like an error to me. (And lately, to a bunch of other people.)

I don’t have a strong view on whether anything in the space of ‘try to slow down some AI research’ should be done. But I think a) the naive first-pass guess should be a strong ‘probably’, and b) a decent amount of thinking should happen before writing off everything in this large space of interventions. Whereas customarily the tentative answer seems to be, ‘of course not’ and then the topic seems to be avoided for further thinking. (At least in my experience—the AI safety community is large, and for most things I say here, different experiences are probably had in different bits of it.)

Maybe my strongest view is that one shouldn’t apply such different standards of ambition to these different classes of intervention. Like: yes, there appear to be substantial difficulties in slowing down AI progress to good effect. But in technical alignment, mountainous challenges are met with enthusiasm for mountainous efforts. And it is very non-obvious that the scale of difficulty here is much larger than that involved in designing acceptably safe versions of machines capable of taking over the world before anyone else in the world designs dangerous versions.

I’ve been talking about this with people over the past many months, and have accumulated an abundance of reasons for not trying to slow down AI, most of which I’d like to argue about at least a bit....

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Lead in Chocolate?, published by jefftk on December 22, 2022 on LessWrong.

Consumer Reports recently tested a bunch of dark chocolate for lead and cadmium, and found that:

For 23 of the bars, eating just an ounce a day would put an adult over a level that public health authorities and CR's experts say may be harmful for at least one of those heavy metals. Five of the bars were above those levels for both cadmium and lead.

If you regularly eat chocolate, how worried about this should you be?

Consumer Reports used the limits from California's Prop

65, which you may recall from the "this thing contains chemicals known to the state of California to cause cancer, birth defects, or other reproductive harm" warnings that are all over everything in

CA. For both lead and cadmium Consumer Reports used the reproductive harm "maximum allowable daily level" (MADL), which are 0.5 µg/d and

4.1µg/d respectively.

CA calculates these limits by looking at studies that try to find the highest dose that has no observable effect (NOEL, "No Observable

Effect Level") or the lowest dose that does have one (LOEL, "Lowest

Observable Effect Level"), and estimating an effect threshold. Then they divide by 1,000 (as required by the original proposition text) to get the MADL limit. For example, here's the process they used for cadmium (pdf).

A safety factor of 1,000 is huge, so it's not surprising that CA's limits are tighter than you see elsewhere. For example, the FDA's limits for lead are 3µg/d for kids and

12.5µg/d for women who might become pregnant (pdf), 6x and 25x the Prop 65 lead MADL. The FDA translates this into a limit of 0.1 ppm for candy likely to be consumed by small children, 5.7x the level Consumer Reports used. [1]

Similarly, for cadmium the CDC reports that the EPA limit (which I can't find on epa.gov) is for food is

1µg/kg/d. CA used a pregnant adult weight of 58 kg in their cadmium calculation, so the equivalent from the EPA limit would be

58µg/d, 14x the CA Prop 65 level. Using 25kg for a child weight would give 25µg/d, 6.1x the Prop 65 cadmium MADL.

Putting these all on the same footing, assuming you're eating 1oz of chocolate daily:

µg/d ppm (µg/g) CA Prop 65 lead MADL 0.5 0.018 FDA kid lead limit 3 0.1 FDA pregnant adult lead limit 12.5 0.44

and:

µg/d ppm (µg/g) CA Prop 65 cadmium MADL 4.1 0.15 EPA child cadmium limit 25 0.88 EPA adult cadmium limit 58 2.0

(sheet)

My overall interpretation is that while it's better to eat chocolate with lower heavy metal content, all things being equal, these levels are basically fine and it's not worth avoiding chocolate over. The Consumer Reports findings seem alarming only because dividing by

1,000 in determining CA Prop 65 levels gives a super conservative safety factor.

[1] Consumer Reports used a limit of 0.5µg/oz, which is

0.018µg/g (the same as ppm).

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: FTX’s collapse mirrors an infamous 18th century British financial scandal, published by Michael Huang on December 22, 2022 on The Effective Altruism Forum.

Amy Froide, Professor of History at the University of Maryland, discusses the remarkable similarities between the recent FTX collapse and the Charitable Corporation Scandal of 1732.

...the management of both ventures was centralized in the hands of just a few people. The Charitable Corporation got into trouble when it reduced its directors from 12 to five and when it consolidated most of its loan business in the hands of one employee – namely, Thomson. FTX’s example is even more extreme, with founder Sam Bankman-Fried calling all the shots...

In both cases, the key fraud was using the assets of one company to prop up another company managed by the same people...

News of both frauds also came as a surprise, with little advance warning. Part of this is due to the ways in which managers were well respected and well connected to both politicians and the financial world...

I would also argue that in both cases the company’s connection to philanthropy lent it another level of cover. The Charitable Corporation’s very name announced its altruism. And even after the scandal subsided, commentators pointed out that the original business of microlending was useful. FTX’s founder Bankman-Fried is an advocate of effective altruism and has argued that it was useful for him and his companies to make lots of money so he could give it away to what he deemed effective causes.

It's worth noting that the Charitable Corporation's demise did not lead to the end of microlending. Pioneers in micro-credit, Muhammad Yunus and Grameen Bank, won the Nobel Peace Prize in 2006.

In the same way, FTX's demise does not necessarily lead to the end of cryptocurrency, or the end of effective altruism.

It does support the case for more oversight and stronger regulations:

After the Charitable Corporation’s collapse in 1732, Parliament didn’t institute any regulation that would prevent such a fraud from happening again.

A tradition of loose oversight and regulations has been the hallmark of Anglo-American capitalism. If the response to the 2008 financial crash is any indication of what will come in the wake of FTX’s collapse, it’s possible that some bad actors, like Bankman-Fried, will be punished. But any regulation will be undone at the first opportunity–or never put in place to begin with.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: December 2022 updates and fundraising, published by AI Impacts on December 22, 2022 on LessWrong.

Harlan Stewart and Katja Grace, 22 December, 2022

News

New Hires and role changes

In 2022, the AI Impacts team has grown from two to seven full time staff. Out of more than 250 applicants, we hired Elizabeth Santos as Operations Lead, Harlan Stewart as Research Assistant, and three Research Analysts: Zach Stein-Perlman, Aysja Johnson, and (are in the process of hiring) Jeffrey Heninger. We’re excited to have them all, and you can learn more about them on our about page.

Rick and Katja have traded some responsibilities: Rick is now Director of AI Impacts, and Katja is Lead Researcher. This means Rick is generally in charge of making decisions about running the org, though Katja has veto power. Katja is responsible for doing research, as well as directing and overseeing it.

Summer Internship Program

We ran an internship program during the summer. Between May and September, six interns worked on various research projects on topics such as international coordination, explanations of historic human success, case studies in risk mitigation, R&D funding in AI, our new survey of Machine Learning researchers, current AI capabilities, technologies that are strategically-relevant to AI, and the scale of machine learning models.

AI Impacts Wiki

We intend to replace our pages with an AI Impacts Wiki. Our pages have always been functionally something like a wiki, so hopefully this new format will make it clearer how to interact with them (as distinct from our blog posts), as well as easier to navigate for readers and easier to update for researchers. The AI Impacts Wiki will launch soon and can be previewed here. . We’ll say more about other minor changes when we launch it, but AI Impacts’ past and future public research will be either detailed on the wiki or findable through the wiki. You can let us know what you think using our feedback form as well as comments on this blog post.

Research

Finished this year

This year, our main new pages and research-heavy blog posts are:

A survey of 738 machine learning experts, about progress in AI. This survey was a rerun of the one conducted by AI Impacts in 2016, and a blog post on the tentative conclusions (Katja and Zach in collaboration with Ben Weinstein-Raun)

Detailed arguments answering the question, ‘Will Superhuman AI be created?’ with a tentative ‘yes’ (Katja)

Review of US public opinion surveys on AI (Zach)

A database of inducement prizes (Elizabeth)

A literature review of notable cognitive abilities of honeybees (Aysja)

An analysis of discontinuities in historic trends in manned altitude (Jeffrey)

A list of counterarguments to the basic AI x-risk case (Katja)

A list of possible incentives to create AI that is known to pose extinction risks (Katja)

Lists of sources arguing for and against existential risk from AI (Katja)

AI Impacts is in large part a set of pages that are intended to get updated over time, so our research should not necessarily show up as new pages, and is generally a bit harder to measure than in more standard research institutions. On this occasion, the above pages and posts probably represent most of our finished research output this year.

In progress

Things people are working on lately:

Noteworthy capabilities and limitations of state-of-the-art AI (Zach, Harlan)

A case study of Alexander Fleming’s efforts to warn the world about antibiotic resistance (Harlan)

A literature review of notable cognitive abilities of ants (Aysja)

Review and analysis of AI forecasting methods (Zach)

Case studies of actors deciding not to pursue technologies, despite apparent incentives to do so. (Jeffrey, Aysja)

Strategically significant narrow AI capabilities (Zach)

The implications of the fermi paradox and anthropics for AI (Zach)

Wha...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: China Covid #4, published by Zvi on December 22, 2022 on LessWrong.

China has now decided to let Covid take its course, so it makes sense to spin developments there off into their own posts for a bit. It is hard to get a good read on what is happening, because China’s core strategy was to stop controlling the spread of Covid and instead to put tight controls on its Covid numbers.

You can’t do this while maintaining any credibility. China decided not to let that be a problem.

Please Disregard the Following

The official case numbers in China are now complete nonsense.

Please disregard the following, except to note the need to disregard it after its peak:

You have to love Reuters reporting ‘WHO says China may be struggling to keep a tally of Covid.’ There is no ‘may’ here, nor is there is a struggle. China is choosing not to keep a tally. Or at least, choosing not share it with the rest of us. If China wanted a count, they’d have a count.

China’s official seven day rolling average of deaths at the moment, according to OurWorldInData? One.

Reuters has an article asking how accurate these numbers are, and the correct response is that it is a headline with a question and the answer is simply ‘no,’ and elsewhere they estimate more than 5,000 deaths per day.

Bloomberg agrees and also estimates over a million daily cases, saying this could get ‘even worse’ and rise to 3.7 million cases per day in January. This is the usual ‘oh no exponential growth could get worse, we could have one or even two doublings ahead of us’ style projection, and I’d be worried about numbers closer to 10 million per day, perhaps even somewhat higher.

So how is China reporting one death? Other than ‘with a straight face’?

They did that the same way they solved their reported infections problem.

It’s simple. They’re redefining Covid deaths the same way they redefined cases.

I interpret this as ‘we will look for any way we can to attribute your death to something else, regardless of whether you had Covid at the time.’

Thus:

Letting it Rip

We went very quickly from Zero Covid to Let It Rip.

China’s public health officials say that possibly 800 million people could be infected with the coronavirus over the next few months. And several models predict that a half million people could die, possibly more.

“Recently, the deputy director of China CDC, Xiaofeng Liang, who’ s a good friend of mine, was announcing through the public media that the first COVID wave may, in fact, infect around 60% of the population,” says Xi Chen, who’s a global health researcher at Yale University and an expert on China’s health-care system.

Scientists at the China National Health Commission estimate the R number is currently a whopping 16 in China durng this surge. “This is a really high level of transmissibility,” Cowling says. “That’s why China couldn’t keep their zero-COVID policy going. The virus is just too transmissible even for them.”

Last winter, cases doubled in the U.S. every three days or so. “Now in China, the doubling time is like hours,” Cowling says.

But Cowling thinks that ultimately China will still fare much better against COVID than America has.

“China has done really well to hold back the virus for three years, and ultimately, I think, the mortality rate will still be much lower than elsewhere in the world,” he says, because the country has vaccinated such a high percentage of its population overall. In other words, the death toll will likely be high, given the sheer number of people infected, but it could have been much worse without the vaccinations, he explains.

Almost no one’s brain is capable of understanding what it means for infections to double every few hours, or to react rationally to that situation. I hope that their vaccine, combined with Omicron being milder, softens the blow.

I am skeptical that i...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: EA Global 2023 applications are now open!, published by Ivan Burduk on December 22, 2022 on The Effective Altruism Forum.

Apply Now

Upcoming conferences

EA Global: Bay Area 2023

Dates: 24–26 February 2023Location: Oakland Marriott City CenterApplication deadline: 8 February 2023 (8:00am UTC)Early bird registration deadline: 13 January 2023 (8:00am UTC)

EA Global: London 2023

Dates: 19–21 May 2023Location: Tobacco DockApplication deadline: 28 April 2023 (8:00am UTC)Early bird registration deadline: 31 March 2023 (8:00am UTC)

What is an EA Global conference like?

EA Globals bring together a network of people who are using the principles of effective altruism to take significant action to have a positive impact on the world. For each event, we’re expecting around 1500 people to come together for a weekend of talks, workshops, and other opportunities. We put a special emphasis on networking and 1:1 meetings, and attendees frequently meet future collaborators, find new jobs, or make new friends.

You can check out content from previous conferences here, and reflections from past attendees here.

Is EA Global the right conference for me?

EA Global is mostly aimed at people who have a solid understanding of the core ideas of effective altruism and who are already taking significant actions based on those ideas. This often looks like professionally working on effective-altruism-inspired projects or working out how best to work on such projects.

EAGx’s are community-run conferences and tend to have broader admissions criteria. They are primarily for people familiar with the core ideas of effective altruism, who want to learn more about what they can do, and want to meet the community, especially in their region. You can see an up to date list of confirmed conferences and their dates here.

If you want to attend but are unsure about whether to apply, please err on the side of applying. You can read more about admissions for EA Global on our FAQ page.

How do I apply to an EA Global?

Click here to go to our (new!) application form.

There's one application form for all EA Globals in 2023, and if your EA Global application is accepted, you can then register for any EA Global in 2023. Though please note that for each event you will have to apply before the application deadline, and we will not be reviewing applications submitted in the window between this deadline and the upcoming conference. Additionally, please note that EAGx applications are separate as these are community run events with different admissions criteria. You can apply for EAGx events via the corresponding page on our website.

This change is part of a new application system which will require you to login using a www.effectivealtruism.org account to apply (similar to how you login to the forum). This has allowed us to enable a variety of features that we’re hoping will allow for a smoother user experience.

How much does it cost to attend EA Global?

Default ticket price: £200 GBP (with the option of a free ticket for students)

For an automatic 25% ticket discount, register by the "early bird" application deadline (approved application required)

Discounts are available — you can select from a range of ticket price options during checkout.

Updated travel expense policy

We (the CEA events team) are planning to reduce spending on events in the coming year, which includes reducing our travel grants budget. Previously we’ve been able to fund people to travel to any EA conference they’ve been accepted to, but we will likely be significantly more conservative in the coming year. For example, we can still provide visa support letters, but we likely won’t be able to fund visa expenses. You can check our updated travel support policy for more details.

While we cannot guarantee travel funding, please do still apply for support if ...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: New intervention: paying farmers to not burn crops, published by Karthik Tadepalli on December 22, 2022 on The Effective Altruism Forum.

I summarize a recent paper evaluating an intervention to reduce air pollution: paying farmers not to burn their crop residue.

A pure conditional contract is ineffective, but when farmers are paid some fraction upfront, they reduce crop burning significantly.

The authors calculate that this contract saves a life for $5,000, comparable to GiveWell's top charities; I calculate an even lower cost of $1,500 per life saved using a broader measurement of health benefits, beating GiveWell's top charities.

I also calculate promising climate co-benefits; this intervention reduces GHG emissions for roughly $36 a ton, which costs less than the social cost of carbon, but more than the best climate mitigation strategies.

I strongly recommend further research and piloting as a way to build on this single study, especially given the lack of scalable air pollution interventions in EA.

Air pollution causes at least 7 million premature deaths each year. Despite this, it has only recently surfaced as a top cause for effective altruists, with Open Philanthropy announcing South Asian air quality as its newest focus area last year. Even with this recent focus, grants have focused mostly on research rather than on interventions to actually improve air quality. The problem is that we don't yet have shovel-ready air pollution interventions, demonstrably cost-effective interventions that are feasible and scalable for charities to implement themselves (rather than relying on uncertain advocacy). I think one recent paper offers a cost-effective, feasible and scalable intervention to reduce air pollution, and we should investigate it much more closely.

Overview

One pernicious source of air pollution in developing countries[1] is crop residue burning, where farmers burn the remnants of their summer crop to quickly clear the fields for their winter crop. The pollution from this burning causes 66,000 premature deaths a year in India alone. Both bans on crop burning and subsidizing alternatives have failed, leaving crop burning in desperate need of a solution.

A new working paper by Kelsey Jack, Seema Jayachandran, Namrata Kala and Rohini Pande proposes one solution: paying farmers directly not to burn crop residue, with a partial payment upfront and the remainder conditional on not burning. They show this contract reduces crop burning with an RCT, while purely conditional payments have no effect. Importantly, they also calculate that this approach saves a life for around $5,000, which is competitive with GiveWell's top charities.

I believe their approach may actually underestimate the benefits of reduced crop burning by focusing solely on premature deaths; using an alternative approach based on DALYs lost, I estimate that this intervention saves a life for $1,500, which beats GiveWell's top charities. In addition, it has important climate co-benefits that traditional health interventions do not.

I think this study poses an important line of research for EA organizations to pursue, and I conclude with some ways for us to use this research.

The paper

One of the authors has summarized the paper already.

In brief: We use a randomized trial to evaluate a program that financially rewarded farmers if they avoided burning their rice stubble. We tried out a standard incentive contract that paid the farmer after we verified that he’d complied with the contract terms. That approach had no impact. However, when the contract was tweaked so that some of the payment was made upfront, the financial rewards program became a very cost-effective way to reduce burning, saving a life for <$5000. [author's emphasis]

I recommend reading her full summary, but here is a self-contained over...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Let’s think about slowing down AI, published by KatjaGrace on December 22, 2022 on LessWrong.

Averting doom by not building the doom machine

If you fear that someone will build a machine that will seize control of the world and annihilate humanity, then one kind of response is to try to build further machines that will seize control of the world even earlier without destroying it, forestalling the ruinous machine’s conquest. An alternative or complementary kind of response is to try to avert such machines being built at all, at least while the degree of their apocalyptic tendencies is ambiguous.

The latter approach seems to me like the kind of basic and obvious thing worthy of at least consideration, and also in its favor, fits nicely in the genre ‘stuff that it isn’t that hard to imagine happening in the real world’. Yet my impression is that for people worried about extinction risk from artificial intelligence, strategies under the heading ‘actively slow down AI progress’ have historically been dismissed and ignored (though ‘don’t actively speed up AI progress’ is popular).

The conversation near me over the years has felt a bit like this:

Some people: AI might kill everyone. We should design a godlike super-AI of perfect goodness to prevent that.

Others: wow that sounds extremely ambitious

Some people: yeah but it’s very important and also we are extremely smart so idk it could work

[Work on it for a decade and a half]

Some people: ok that’s pretty hard, we give up

Others: oh huh shouldn’t we maybe try to stop the building of this dangerous AI?

Some people: hmm, that would involve coordinating numerous people—we may be arrogant enough to think that we might build a god-machine that can take over the world and remake it as a paradise, but we aren’t delusional

This seems like an error to me. (And lately, to a bunch of other people.)

I don’t have a strong view on whether anything in the space of ‘try to slow down some AI research’ should be done. But I think a) the naive first-pass guess should be a strong ‘probably’, and b) a decent amount of thinking should happen before writing off everything in this large space of interventions. Whereas customarily the tentative answer seems to be, ‘of course not’ and then the topic seems to be avoided for further thinking. (At least in my experience—the AI safety community is large, and for most things I say here, different experiences are probably had in different bits of it.)

Maybe my strongest view is that one shouldn’t apply such different standards of ambition to these different classes of intervention. Like: yes, there appear to be substantial difficulties in slowing down AI progress to good effect. But in technical alignment, mountainous challenges are met with enthusiasm for mountainous efforts. And it is very non-obvious that the scale of difficulty here is much larger than that involved in designing acceptably safe versions of machines capable of taking over the world before anyone else in the world designs dangerous versions.

I’ve been talking about this with people over the past many months, and have accumulated an abundance of reasons for not trying to slow down AI, most of which I’d like to argue about at least a bit. My impression is that arguing in real life has coincided with people moving toward my views.

Quick clarifications

First, to fend off misunderstanding

I take ‘slowing down dangerous AI’ to include any of: (So in particular, I’m including both actions whose direct aim is slowness in general, and actions whose aim is requiring safety before specific developments, which implies slower progress.)

reducing the speed at which AI progress is made in general, e.g. as would occur if general funding for AI declined.

shifting AI efforts from work leading more directly to risky outcomes to other work, e.g. as might...