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: Penn EA Residency Takeaways, published by Thomas Kwa on November 12, 2021 on The Effective Altruism Forum.

Thanks to Zack Dugue, Oliver Habryka, Adam Krivka, and the Penn EA team for useful comments.

In September 2021, sydv and I (Thomas Kwa) helped revive the University of Pennsylvania EA group through a residency. We think this went very well, and the ~300 hours invested during the 3 weeks we were there probably sped up the group by 4 months, and possibly made the group significantly better overall [3]. As of November 2021, Penn EA is currently a thriving group with six organizers, ~30 Intro Program participants and ~20 weekly dinner attendees, and potential to grow further into its huge 10,000 undergrad population, whereas I'd guess the counterfactual looks something like slowly building up to this size over the course of ~8 months with significant risk of the group dying again.

What is a residency?

In (our model of) a residency, one or two EA community builders travel to a large university at the start of the school year, and spend at least 1 FTE building a new or existing EA student group. At the end, the group is handed off to students, and there might be a retreat for new organizers. The primary goal is to build organizer capacity to rapidly make the group large and self-sustaining (at least 2 students with 10h/week each); a secondary goal is general community-building.

Penn is strong evidence that this model can work well. It's only one data point, but we think a lot of the success is generalizable. Note there were other residencies this fall that didn't go as well.

Misc comments

Sydney (a Stanford student) and I (a Caltech student) were able to invest this much time despite being full-time students with our own groups to run because Penn starts 3 weeks earlier than Caltech or Stanford. [1]

Our time was mostly spent building the website, tabling (sitting outside in a high-traffic area and advertising the club), having one-on-ones with prospective organizers and others who signed up through the website, organizing and speaking at events like intro sessions and dinners, other advertising (like flyers) and organizing a retreat (see below). Of these actions, we spent the most time tabling. But one-on-ones and the retreat were the most efficient use of time, as they build organizer capacity. The other actions are still roughly as good as standard university organizer time, given the importance of the first few weeks of uni.

Tabling: At Penn, it's really hard to send emails to the whole school. Emails are necessary to get attendance at events. We used the labor-intensive but very effective strategy of tabling to collect ~900 emails, and averaged ~20 emails and ~1 Intro Program application per hour (more on club fair days). Just as importantly, we met a couple more potential organizers. We highly recommend that group organizers read our Guide to Tabling, because following these practices more than doubled our tabling effectiveness.

One-on-ones: Sydney did most of these; the goal was to identify potential organizers / active members and what they might contribute to the club, and also just get to know people. I think there were about 20 one-hour 1-1s.

Retreat: We ran a weekend retreat for Mid-Atlantic region EA group organizers after the residency, with a total of ~30 people including ~8 from Penn. Retreats are out of scope of this post, but we think they're great and pair well with residencies; all the newly excited people you've had 1-1s with can talk to each other and learn more about EA and EA community building.

We did expensive messaging experiments which were inconclusive (we technically started 3 different clubs, and tabled once as Penn Rationality and a couple times as Penn Impactful Careers). In the end we stuck with neutral EA branding, roughly "we want to...

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: Weak point in "most important century": lock-in, published by Holden Karnofsky on November 11, 2021 on The Effective Altruism Forum.

This is the second of (for now) two posts covering what I see as the weakest points in the “most important century” series. (The first one is here.)

The weak point I’ll cover here is the discussion of “lock-in”: the idea that transformative AI could lead to societies that are stable for billions of years. If true, this means that how things go this century could affect what life is like in predictable, systematic ways for unfathomable amounts of time.

My main coverage of this topic is in a section of my piece on digital people. It’s pretty hand-wavy, not super thorough, and isn’t backed by an in-depth technical report (though I do link to some informal notes from physicist Jess Riedel that he made while working at Open Philanthropy). Overcoming Bias critiqued me on this point, leading to a brief exchange in the comments.

I’m not going to be dramatically more thorough or convincing here, but I will say a bit more about how the overall “most important century” argument is affected if we ignore this part of it, and a bit more about why I find “lock-in” plausible.

(Also note that "lock-in" will be discussed at some length in an upcoming book by Will MacAskill, What We Owe the Future.)

Throughout this piece, I’ll be using “lock-in” to mean “key things about society, such as who is in power or which religions/ideologies are dominant, are locked into place indefinitely, plausibly for billions of years,” and “dynamism” or “instability” to mean the opposite: “such things change on much shorter time horizons, as in decades/centuries/millennia.” As noted previously, I consider "lock-in" to be a scary possibility by default, though it's imaginable that certain kinds of lock-in (e.g., of human rights protections) could be good.

“Most important century” minus “lock-in”

First, let’s just see what happens if we throw out this entire part of the argument and assume that “lock-in” isn’t a possibility at all, but accept the rest of the claims. In other words, we assume that:

Something like PASTA (advanced AI that automates scientific and technological advancement) is likely to be developed this century.

That, in turn, would lead to explosive scientific and technological advancement, resulting in a world run by digital people or misaligned AI or something else that would make it fair to say we have "transitioned to a state in which humans as we know them are no longer the main force in world events."

But it would not lead to any particular aspect of the world being permanently set in stone. There would remain billions of years full of unpredictable developments.

In this case, I think there is still an important sense in which this would be the “most important century for humanity”: it would be our last chance to shape the transition from a world run by humans to a world run by something very much unlike humans. This is one of the two definitions of “most important century” given here.

More broadly, in this case, I think there’s an important sense in which the “most important century” series should be thought of as “Pointing to a drastically underrated issue; correct in its most consequential, controversial implications, if not in every detail.” When people talk about the most significant issues of our time (in fact, even when they are specifically talking about likely consequences of advanced AI), they rarely include much discussion of the sorts of issues emphasized in this series; and they should, whether or not this series is correct about the possibility of “lock-in.”

As noted here, I ultimately care more about whether the “most important century” series is correct in this sense - pointing at drastically underappreciated issues - than about how likely 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: Weak point in "most important century": lock-in, published by Holden Karnofsky on November 11, 2021 on The Effective Altruism Forum.

This is the second of (for now) two posts covering what I see as the weakest points in the “most important century” series. (The first one is here.)

The weak point I’ll cover here is the discussion of “lock-in”: the idea that transformative AI could lead to societies that are stable for billions of years. If true, this means that how things go this century could affect what life is like in predictable, systematic ways for unfathomable amounts of time.

My main coverage of this topic is in a section of my piece on digital people. It’s pretty hand-wavy, not super thorough, and isn’t backed by an in-depth technical report (though I do link to some informal notes from physicist Jess Riedel that he made while working at Open Philanthropy). Overcoming Bias critiqued me on this point, leading to a brief exchange in the comments.

I’m not going to be dramatically more thorough or convincing here, but I will say a bit more about how the overall “most important century” argument is affected if we ignore this part of it, and a bit more about why I find “lock-in” plausible.

(Also note that "lock-in" will be discussed at some length in an upcoming book by Will MacAskill, What We Owe the Future.)

Throughout this piece, I’ll be using “lock-in” to mean “key things about society, such as who is in power or which religions/ideologies are dominant, are locked into place indefinitely, plausibly for billions of years,” and “dynamism” or “instability” to mean the opposite: “such things change on much shorter time horizons, as in decades/centuries/millennia.” As noted previously, I consider "lock-in" to be a scary possibility by default, though it's imaginable that certain kinds of lock-in (e.g., of human rights protections) could be good.

“Most important century” minus “lock-in”

First, let’s just see what happens if we throw out this entire part of the argument and assume that “lock-in” isn’t a possibility at all, but accept the rest of the claims. In other words, we assume that:

Something like PASTA (advanced AI that automates scientific and technological advancement) is likely to be developed this century.

That, in turn, would lead to explosive scientific and technological advancement, resulting in a world run by digital people or misaligned AI or something else that would make it fair to say we have "transitioned to a state in which humans as we know them are no longer the main force in world events."

But it would not lead to any particular aspect of the world being permanently set in stone. There would remain billions of years full of unpredictable developments.

In this case, I think there is still an important sense in which this would be the “most important century for humanity”: it would be our last chance to shape the transition from a world run by humans to a world run by something very much unlike humans. This is one of the two definitions of “most important century” given here.

More broadly, in this case, I think there’s an important sense in which the “most important century” series should be thought of as “Pointing to a drastically underrated issue; correct in its most consequential, controversial implications, if not in every detail.” When people talk about the most significant issues of our time (in fact, even when they are specifically talking about likely consequences of advanced AI), they rarely include much discussion of the sorts of issues emphasized in this series; and they should, whether or not this series is correct about the possibility of “lock-in.”

As noted here, I ultimately care more about whether the “most important century” series is correct in this sense - pointing at drastically underappreciated issues - than about how likely 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: Improve delegation abilities today, delegating heavily tomorrow, published by Ozzie Gooen on November 11, 2021 on The Effective Altruism Forum.

TLDR: Eventually, we want higher and higher levels of worthwhile delegation. This will take work.

Rigor: Very quickly written rant. I want to experiment with getting more pieces like these out there, as opposed to focusing more on polish. Feedback is appreciated.

Example 1: Bricks

Low delegation: You assemble bricks for your house

Medium: You hire a team to put together your house

High: You tell an assistant that you want a house. They go about choosing a plan and selecting a team.

Very high: You tell an assistant to “make your life better”, they decide that a house is the best approach, and then they go about figuring it out.

Example 2: Charity

Low delegation: You give food to a homeless person

Medium: You donate to a nonprofit feeding the homeless

High: You donate to a nonprofit fund that re-grants to nonprofits that help the homeless

Very high: You donate to a fund to “do useful stuff”, and they sometimes decide that helping the homeless is ideal.

If delegation can be well managed, it’s incredibly useful. There’s always more and more stuff to do and be concerned about. We clearly want to hand off as much work as possible to others and then for those tasks to achieve economies of scale.

Those who can delegate well, often win. This is exactly the job of CEOs/entrepreneurs/managers. Companies with great teams that can be given bold projects with little oversight, or nations with very low political corruption or infighting.

Right now in the US, we don't trust the government much, so it's relegated to a low delegation level. This means it just simply can't do that much.

The higher level you delegate, the more agency you (typically) sacrifice. When you choose a house construction crew, you’ll have a bit less control over your brick types. When you choose some highly-meta-nonprofit-fund, you give them incredibly broad authority to handle your money.

Typically delegation breaks down somewhere between levels 1-4. Sometimes this is due to undertrust; you expect the authority will be corrupt or poor, even if they won’t be. Sometimes it’s overtrust; you mistakenly expect things to go well, then they abuse or misuse the agency.

To work well:

There needs to be significant amounts of justified trust between client and provider

The costs of delegation must be lower than the benefits

Most altruists seem to be in either the “medium” levels, or the unjustified “very high levels” (for example, donating to a single politician or cult to fix all of one’s problems). Some are in the low levels and just go out and try to help people personally (with widely mixed results).

On one hand, you have literal cult leaders, and on the other, you have millionaire altruists literally laying bricks to build houses.

Right now in EA, we have some “high” levels of funding delegation (EA Funds, GiveWell), but it's limited.

There’s a ton of value at hand on moving towards more delegation and increasing the amount of justified trust. “Everyone trying to do all of the necessary analysis on all things” themselves clearly doesn’t scale, and it should be clear that very few people are strong at this.

I don’t see that much literature at this level of abstraction, but I often see arguments like,

“We can’t trust the authorities”

“It’s important to think for yourself”

I get that, but it seems like it’s arguing on the wrong axis. We really want to push forward the Pareto frontier of delegation potential. That’s much more important, in the long run, than small decisions of where on that frontier we want to be right now. The frontier right now is bad and I think we could do much better.

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

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: Improve delegation abilities today, delegating heavily tomorrow, published by Ozzie Gooen on November 11, 2021 on The Effective Altruism Forum.

TLDR: Eventually, we want higher and higher levels of worthwhile delegation. This will take work.

Rigor: Very quickly written rant. I want to experiment with getting more pieces like these out there, as opposed to focusing more on polish. Feedback is appreciated.

Example 1: Bricks

Low delegation: You assemble bricks for your house

Medium: You hire a team to put together your house

High: You tell an assistant that you want a house. They go about choosing a plan and selecting a team.

Very high: You tell an assistant to “make your life better”, they decide that a house is the best approach, and then they go about figuring it out.

Example 2: Charity

Low delegation: You give food to a homeless person

Medium: You donate to a nonprofit feeding the homeless

High: You donate to a nonprofit fund that re-grants to nonprofits that help the homeless

Very high: You donate to a fund to “do useful stuff”, and they sometimes decide that helping the homeless is ideal.

If delegation can be well managed, it’s incredibly useful. There’s always more and more stuff to do and be concerned about. We clearly want to hand off as much work as possible to others and then for those tasks to achieve economies of scale.

Those who can delegate well, often win. This is exactly the job of CEOs/entrepreneurs/managers. Companies with great teams that can be given bold projects with little oversight, or nations with very low political corruption or infighting.

Right now in the US, we don't trust the government much, so it's relegated to a low delegation level. This means it just simply can't do that much.

The higher level you delegate, the more agency you (typically) sacrifice. When you choose a house construction crew, you’ll have a bit less control over your brick types. When you choose some highly-meta-nonprofit-fund, you give them incredibly broad authority to handle your money.

Typically delegation breaks down somewhere between levels 1-4. Sometimes this is due to undertrust; you expect the authority will be corrupt or poor, even if they won’t be. Sometimes it’s overtrust; you mistakenly expect things to go well, then they abuse or misuse the agency.

To work well:

There needs to be significant amounts of justified trust between client and provider

The costs of delegation must be lower than the benefits

Most altruists seem to be in either the “medium” levels, or the unjustified “very high levels” (for example, donating to a single politician or cult to fix all of one’s problems). Some are in the low levels and just go out and try to help people personally (with widely mixed results).

On one hand, you have literal cult leaders, and on the other, you have millionaire altruists literally laying bricks to build houses.

Right now in EA, we have some “high” levels of funding delegation (EA Funds, GiveWell), but it's limited.

There’s a ton of value at hand on moving towards more delegation and increasing the amount of justified trust. “Everyone trying to do all of the necessary analysis on all things” themselves clearly doesn’t scale, and it should be clear that very few people are strong at this.

I don’t see that much literature at this level of abstraction, but I often see arguments like,

“We can’t trust the authorities”

“It’s important to think for yourself”

I get that, but it seems like it’s arguing on the wrong axis. We really want to push forward the Pareto frontier of delegation potential. That’s much more important, in the long run, than small decisions of where on that frontier we want to be right now. The frontier right now is bad and I think we could do much better.

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

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: Covid 11/11: Winter and Effective Treatments Are Coming, published by Zvi on November 11, 2021 on LessWrong.

I was considering splitting off my analysis of the effectiveness of Pfizer’s Paxlovid, a new Covid-19 treatment pill, but it’s so straightforward there’s no reason to do that. It works, with our best guess being it is 89% effective at preventing hospitalization and death, and that’s that. Trial was halted because medical ‘ethics’ is crazy and decided we couldn’t study the pill because it was too effective. Paxlovid still needs to be approved and production needs to be scaled up, but both of these seem like near certainties.

In the meantime, it looks like cases are heading back upwards again. There was high uncertainty (I said I’d be more surprised by no change than by some change, but it wasn’t obvious which direction was more likely) so this was definitely in the possibility space, but it’s near the top end and rather disheartening.

There’s the usual kinds of developments on other fronts, but none of it should come as a meaningful surprise.

Executive Summary

Pfizer’s new Covid-19 treatment pill Paxlovid is ~89% effective at preventing hospitalization and death, but it will take several months to ramp up production and gain official approvals before it is widely available. Effective treatment is coming.

Cases are starting to increase again outside of the South. Winter is coming.

Everything else about as you’d expect.

Let’s run the numbers.

The Numbers

Predictions

Prediction from last week: 442k cases (no change) and 7500 deaths (-11%)

Results: 470k cases (+7%) and 8092 deaths (-4%).

Prediction for next week:

The miss on deaths is due to a large increase in the West. On reflection I likely should have baked some of that in, -11% was a lousy prediction and -7% or so would have been a better prediction. I do think that this is still a disappointing number despite that. For cases the miss is the same size, but this seems more like uncertainty rather than mostly being a mistake in prediction. Cases rising this much this quickly was definitely surprising, and a large update on what’s likely to happen this winter.

Deaths

The jump in the West is not centered on one particular state, with several reporting the highest levels in a while. I’d like to say it is a systematic reporting pattern somehow, and I presume that some deaths from last week out West got shifted into this week. I don’t have any gears for that explanation and given it’s not focused in one state I don’t feel great about not having any gears to explain it. Still, it seems a lot more likely than this being a real decline followed by a real increase.

Cases

The increase looks fully real. The question is not whether cases will continue increasing, but rather whether the increase will accelerate and by how much. Child vaccinations will help a bit once they come online, but that will take a while and the total size is likely small, so we are probably looking at another winter wave.

Vaccinations

[Vaccinations graphs]

Vaccine Effectiveness

Half doses of Pfizer produce strong immune response (MR), presumably with fewer short term side effects, and conserving supply. We could have vaccinated far more people far quicker, here and around the world, with smaller doses. Oh well. At least kids are already going to get lower doses. This will almost certainly have exactly zero impact on the doses given out in the United States, and at this point it’s too late for most of the gains from this anyway.

Did you know the vaccines are super effective against all-cause mortality? I mean, no, that’s crazy and any such finding is super duper suspicious and presumed hopelessly confounded but it would be remiss not to holy check out this chart from this video:

The effect is even stronger for Moderna than Pfizer which is stronger th...

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: Covid 11/11: Winter and Effective Treatments Are Coming, published by Zvi on November 11, 2021 on LessWrong.

I was considering splitting off my analysis of the effectiveness of Pfizer’s Paxlovid, a new Covid-19 treatment pill, but it’s so straightforward there’s no reason to do that. It works, with our best guess being it is 89% effective at preventing hospitalization and death, and that’s that. Trial was halted because medical ‘ethics’ is crazy and decided we couldn’t study the pill because it was too effective. Paxlovid still needs to be approved and production needs to be scaled up, but both of these seem like near certainties.

In the meantime, it looks like cases are heading back upwards again. There was high uncertainty (I said I’d be more surprised by no change than by some change, but it wasn’t obvious which direction was more likely) so this was definitely in the possibility space, but it’s near the top end and rather disheartening.

There’s the usual kinds of developments on other fronts, but none of it should come as a meaningful surprise.

Executive Summary

Pfizer’s new Covid-19 treatment pill Paxlovid is ~89% effective at preventing hospitalization and death, but it will take several months to ramp up production and gain official approvals before it is widely available. Effective treatment is coming.

Cases are starting to increase again outside of the South. Winter is coming.

Everything else about as you’d expect.

Let’s run the numbers.

The Numbers

Predictions

Prediction from last week: 442k cases (no change) and 7500 deaths (-11%)

Results: 470k cases (+7%) and 8092 deaths (-4%).

Prediction for next week:

The miss on deaths is due to a large increase in the West. On reflection I likely should have baked some of that in, -11% was a lousy prediction and -7% or so would have been a better prediction. I do think that this is still a disappointing number despite that. For cases the miss is the same size, but this seems more like uncertainty rather than mostly being a mistake in prediction. Cases rising this much this quickly was definitely surprising, and a large update on what’s likely to happen this winter.

Deaths

The jump in the West is not centered on one particular state, with several reporting the highest levels in a while. I’d like to say it is a systematic reporting pattern somehow, and I presume that some deaths from last week out West got shifted into this week. I don’t have any gears for that explanation and given it’s not focused in one state I don’t feel great about not having any gears to explain it. Still, it seems a lot more likely than this being a real decline followed by a real increase.

Cases

The increase looks fully real. The question is not whether cases will continue increasing, but rather whether the increase will accelerate and by how much. Child vaccinations will help a bit once they come online, but that will take a while and the total size is likely small, so we are probably looking at another winter wave.

Vaccinations

[Vaccinations graphs]

Vaccine Effectiveness

Half doses of Pfizer produce strong immune response (MR), presumably with fewer short term side effects, and conserving supply. We could have vaccinated far more people far quicker, here and around the world, with smaller doses. Oh well. At least kids are already going to get lower doses. This will almost certainly have exactly zero impact on the doses given out in the United States, and at this point it’s too late for most of the gains from this anyway.

Did you know the vaccines are super effective against all-cause mortality? I mean, no, that’s crazy and any such finding is super duper suspicious and presumed hopelessly confounded but it would be remiss not to holy check out this chart from this video:

The effect is even stronger for Moderna than Pfizer which is stronger th...

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: Why boring writing is unethical: the case for it being high impact to make EA writing entertaining, published by Kat Woods on November 11, 2021 on The Effective Altruism Forum.

EA content is often dry, which leads to fewer people reading it, which leads to the content having less impact.

Dry writing → ↓ People reading → ↓ Impact

EA writing can’t have an impact if the right people don’t read it, and the right people are more likely to read it if it’s interesting. Therefore, if you want to maximize impact (which I imagine you do since you’re reading this), you shouldn’t make your writing dry.

In the rest of this essay I'll explain why it's a problem, potential objections, and practical tips for how to make writing more engaging.

Is this really a problem?

Of course, there’s a spectrum of how entertaining EA writing is and it’s a wide distribution, with some very engaging writing being produced. Joe Carlsmith’s “Against neutrality about creating happy lives” and Nate Soares’ Replacing Guilt series come to mind.

However, I think we can all agree that a large amount of the writing is a bit. . . well, mostly sticking to a dry, unobjectionable list of claims and arguments. The percentage of posts that have a single joke in them is probably than half, perhaps as low as 10%.

The numbers are similarly bad for most other ways that an article can be spiced up, such as with clever turns of phrase, images, or anything that might cause you to feel any sort of emotion.

The focus is almost entirely on accuracy with little consideration to other possible metrics, such as being engaging or beautiful.

The writing is, in perhaps unsurprising news, rather utilitarian.

Why this matters: dry writing leads to less utility for the world on average

This is a problem because if nobody reads an article, it has no impact. Research and writing have an impact through other people, so other people need to have a way of being affected by the research, which is typically by reading or listening to the content. Here’s a quick excerpt from a previous post I wrote about why increasing (impact-adjusted) readership is important:

Here are some purely hypothetical numbers just to illustrate this way of thinking:

Imagine that you, a researcher, have spent 100 hours producing outstanding research that is relevant to 1,000 out of a total of 10,000 EAs.

Each relevant EA who reads your research will generate $1,000 of positive impact. So, if all 1,000 relevant EAs read your research, you will generate $1 million of impact.

You post it to the EA Forum, where posts receive 500 views on average. Let’s say, because your report is long, only 20% read the whole thing - that’s 100 readers. So you’ve created 1001,000 = $100,000 of impact. Since you spent 100 hours and created $100,000 of impact, that’s $1,000 per hour - pretty good!

But if you were to spend, say 1 hour, promoting your report - for example, by posting links on EA-related Facebook groups - to generate another 100 readers, that would produce another $100,000 of impact. That’s $100,000 per marginal hour or ~$2,000 per hour taking into account the fixed cost of doing the original research.

Likewise, if you spend a bit of time while writing your essay to make it interesting and fun, you could potentially 2-100x the readership and thus impact of your research. In this example, that could lead to another tens of thousands to millions of dollars worth of value per marginal hour. This is an extremely good investment of your effort.

Perhaps the most compelling example of this effect is Eliezer Yudkowsky. I remember one day deciding that I would read decision theory books instead of relying on LessWrong. To my surprise, I realized that pretty much everything in the sequences is in the intro to decision theory textbooks. However, they present it in the most boring, th...

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: Why boring writing is unethical: the case for it being high impact to make EA writing entertaining, published by Kat Woods on November 11, 2021 on The Effective Altruism Forum.

EA content is often dry, which leads to fewer people reading it, which leads to the content having less impact.

Dry writing → ↓ People reading → ↓ Impact

EA writing can’t have an impact if the right people don’t read it, and the right people are more likely to read it if it’s interesting. Therefore, if you want to maximize impact (which I imagine you do since you’re reading this), you shouldn’t make your writing dry.

In the rest of this essay I'll explain why it's a problem, potential objections, and practical tips for how to make writing more engaging.

Is this really a problem?

Of course, there’s a spectrum of how entertaining EA writing is and it’s a wide distribution, with some very engaging writing being produced. Joe Carlsmith’s “Against neutrality about creating happy lives” and Nate Soares’ Replacing Guilt series come to mind.

However, I think we can all agree that a large amount of the writing is a bit. . . well, mostly sticking to a dry, unobjectionable list of claims and arguments. The percentage of posts that have a single joke in them is probably than half, perhaps as low as 10%.

The numbers are similarly bad for most other ways that an article can be spiced up, such as with clever turns of phrase, images, or anything that might cause you to feel any sort of emotion.

The focus is almost entirely on accuracy with little consideration to other possible metrics, such as being engaging or beautiful.

The writing is, in perhaps unsurprising news, rather utilitarian.

Why this matters: dry writing leads to less utility for the world on average

This is a problem because if nobody reads an article, it has no impact. Research and writing have an impact through other people, so other people need to have a way of being affected by the research, which is typically by reading or listening to the content. Here’s a quick excerpt from a previous post I wrote about why increasing (impact-adjusted) readership is important:

Here are some purely hypothetical numbers just to illustrate this way of thinking:

Imagine that you, a researcher, have spent 100 hours producing outstanding research that is relevant to 1,000 out of a total of 10,000 EAs.

Each relevant EA who reads your research will generate $1,000 of positive impact. So, if all 1,000 relevant EAs read your research, you will generate $1 million of impact.

You post it to the EA Forum, where posts receive 500 views on average. Let’s say, because your report is long, only 20% read the whole thing - that’s 100 readers. So you’ve created 1001,000 = $100,000 of impact. Since you spent 100 hours and created $100,000 of impact, that’s $1,000 per hour - pretty good!

But if you were to spend, say 1 hour, promoting your report - for example, by posting links on EA-related Facebook groups - to generate another 100 readers, that would produce another $100,000 of impact. That’s $100,000 per marginal hour or ~$2,000 per hour taking into account the fixed cost of doing the original research.

Likewise, if you spend a bit of time while writing your essay to make it interesting and fun, you could potentially 2-100x the readership and thus impact of your research. In this example, that could lead to another tens of thousands to millions of dollars worth of value per marginal hour. This is an extremely good investment of your effort.

Perhaps the most compelling example of this effect is Eliezer Yudkowsky. I remember one day deciding that I would read decision theory books instead of relying on LessWrong. To my surprise, I realized that pretty much everything in the sequences is in the intro to decision theory textbooks. However, they present it in the most boring, th...

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: TikTok EA Book Giveaway Intermediate Results, published by Ben West on November 11, 2021 on The Effective Altruism Forum.

Myself and Lacey (@benthamite) and Emma Williamson (@dailydosenews) piloted giving away copies of Doing Good Better on TikTok. This is a short write up of intermediate results.

Executive summary

We gave away 359 books across 5 videos

320 of those came from the most successful video

Each video took a couple of hours of work

I want to do a follow-up survey to see how many of these people actually read the book and got involved in EA. My guess is that a few of them did, which would make this a fairly cost-effective outreach strategy. I plan to publish a longer write up once those survey results are in.

Process

We created videos advertising various EA things, with a call to action for them to get a free book by accessing a link in our bio

People filled out a form with their contact information to receive the book. We did not require a credit card etc.

The form let them opt out of receiving a follow-up survey

Advertising

I tried purchasing promotion for one video. In TikTok, this means that the video was shown to more people than usual with small text indicating that it's being shown to the user via paid promotion.

As expected, this dramatically increased the number of people viewing the video, but very few of them signed up to get a book.

Probably someone should try this with a larger sample size though, and also with video which is inherently more compelling.

(I am unable to promote the most popular video as the sound is copyrighted.)

Thoughts on the most popular video

I filmed Sam doing a dance (as opposed to e.g. talking into the camera) because this let me decide later what text to put on top of it instead of having to decide that before filming.

I think this was the right call – my first attempt was not very successful, but the second one was.

Sam obviously has starpower but the fact that the first video was unsuccessful indicates that people aren't just watching because of his celebrity. This makes me hopeful that at least some of the people who got books did so because they were intrigued by the message.

The success of the video seems to have mostly been driven by having a higher watch percentage (as opposed to engagement through likes/comments etc.)

Thanks to the EABD team and the Open Philanthropy Project for making this project possible

16

1,543

136

5

5

21.0%

3-10

1,090

69

9

1

17.0%

4-7

14,200

247

2

0

23

2,580

84

18

7

31.7%

9

125,800

1,652

182

390

43.9%

320

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