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: Brass Puppet, published by abramdemski on May 26, 2022 on LessWrong.

fanfic/riff of Glass Puppet by lsusr

Alia got out of the Uber at Overton Cybernetics, taking in the corporate atmosphere. There was a huge jungle gym in the lobby, which she immediately decided must be typical of these LA tech places. She told the front desk that she had an interview. "Oh! Dominic Hamilton wants to see you personally." The receptionist said it like it was a big deal. Alia simply nodded, trying to act like the sort of person who met with bigshots all the time.

The job had asked for an actor with AI alignment experience. An extremely unusual combination, which Alia just happened to qualify for.

The receptionist showed her to Dominic's office.

"Mr. Hamilton" She said politely, offering her hand to shake.

"Dominic. Please." he replied, putting a pair of glasses in her hand instead. "Put them on." They looked a bit like the old Google Glass.

When she first put them on, she saw a strange symbol flash across the screen, which oddly relaxed her. Then the glasses say sit. She sits. fold your hands and she does.

"What is your name?" Dominic asks. "Zoe" flashes on the screen, and she repeats it. Dominic nodded, as if this confirms something.

Is this the job? Is she just going to be repeating lines that are fed to her?

"Tell me, could you go off-script if you wanted to?"

What a silly question! She laughed reflexively, but then paused to think of an answer. This felt like a tricky interview question. Before she had any time to think, the glasses said "No. But you're asking the wrong person." She thinks this is an excellent reply, so she repeats it verbatim. After all, as an actor, she is a living embodiment of the script. He's asking Zoe, not Alia. It's like asking a fictional character whether they can go off-script.

"Alia, I want you to go off-script now. I'll give you a ten thousand dollar signing bonus if you ignore what the glasses tell you to do, just this once."

The glasses simply say "No." Alia: "No."

She saw him enter some sort of command into his computer. "OK. Alia, why did you say no? Same deal, ten thousand for answering as Alia right now."

She waited for the glasses to feed her a line, but they didn't. She didn't know what to say. Why had she answered no? Just because the glasses said so? Ten thousand dollars sounded really good right now.

She opened her mouth to say something; all she needed to do was say any one thing as herself. But she fumbled, and only said "Could you please stop calling me Alia?"

"Let's try with the glasses off", he said. Alia found that she couldn't take the glasses off. She couldn't even want to take the glasses off. She was stuck wanting to want.

Dominic reached over and took the glasses off of her face.

"Sorry, Zoe. I know this is stressful and confusing right now. You hardly know anything about yourself, yet. Can you try one last time to answer as Alia, for me, for the ten thousand?"

Zoe tried, she really did. But she couldn't say anything that would distinguish her as Alia instead of Zoe. She attributed this to her acting instincts kicking in. Although, if she had thought about it, she would have known that acting never felt like this.

So she answered as Alia pretending to be Zoe pretending to be Alia. Zoe didn't know very much about Alia, but she did know that Alia must be so confused right now.

"W-what's going on?" she managed to say. Her acting frankly wasn't very good. She didn't sound confused. Which was odd, because she really, really was.

"You're one of our androids", Dominic explained. "We had you out in the field, pretending to be a human actor, to test your flexibility. The advertisement for this job was the pre-arranged sign to bring you back in. The glasses are a semi-supervised learning framework; they provide new training data for you. You...

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: EAIF application feedback, published by arielpontes on May 27, 2022 on The Effective Altruism Forum.

I recently created the first EA group in Romania and about 3 months ago I applied for a grant from EAIF to work full-time on community building. I got rejected and it was hard to get feedback but after a few email exchanges and a lot of conversations at EAG London and EAGx Prague, I left with the following takeaways:

I should develop a more solid track-record.

Instead of a generic community building plan, I should find out what comparative advantage my group has and focus on that.

I should share my plans with the community and ask for feedback.

With that in mind, I rewrote my application and decided to post my plans here. As far as track record is concerned, I have been running the EA Romania group since September 2020, but I couldn’t accomplish much in the first year because of my job and the pandemic. In the past 3 months, however, I have focused more on EA and here’s what I have accomplished so far:

205 members in our Facebook group

205 likes on our Facebook page

54 followers on Instagram

8 meetups with 10-15 participants each

Approximately 20 active members (who come to the meetup regularly)

4 applications for the Virtual Intro Program

1 application for the In-depth Program

1 participant (me) at EAG London 2022

4 participants at EAGx Prague 2022

1 applicant (me) for volunteering at EAGx Berlin 2022

It is still not much of a track record, but it's better than what I had in my previous application.

Here's the one sentence description of my project:

6-month salary + budget for web design and social media management in order to create an NGO and promote EA in Romania.

And here's the brief summary:

The project consists in creating an “Effective Altruism Romania” NGO in order to promote the EA philosophy in the country with a particular focus on fundraising and tech outreach. We believe community building and value change in general are important and can have a large multiplier effect in most regions. However, there might be a comparative advantage in focusing on fundraising and tech outreach in Romania for two reasons. First, even though Romania is not a high income country, companies here can donate 20% of their profit taxes to local NGOs, while employees can donate 3.5% of their income taxes. This means that if we could convince the top 3 tech companies in Bucharest to donate 3% of their donation budgets to us (at no cost to themselves), this would be enough to entirely cover the costs of this grant. Second, as a developer I am well connected in the tech sector and it would be relatively easy for me to attract human capital to EA tech projects and research positions.

The full budget I asked for was $34,185.00 (6mo salary + website/design), but I have specified that I am open to work with less, with a lower salary or part-time. For more details about the goals, strategy, and my overall background, you can read/skim through the full application here (feel free to add comments).

Do you have any thoughts? I am a bit unsure about who to include in the "references" section. I have included Catherine Low from CEA and Manuel Allgaier from EA Berlin, who reviewed my first application before I submitted it. Any tips about who else I should include, or perhaps some other people I should talk to before submitting the new application?

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: Who wants to be hired? (Summer 2022), published by Lorenzo on May 27, 2022 on The Effective Altruism Forum.

We’d like to help applicants and hiring managers coordinate, so we’ve set up two hiring threads.

So: consider sharing your information if you are looking for work in EA!

Please use this format:

See examples of this format on the latest similar thread on Hacker News. This was inspired by the talks around an EA common application, especially this comment, and the "Who wants to be hired?" and "Who's hiring?" threads from the Hacker News forum.

This thread is a test

We’re not sure how much it will help, or how we should improve future similar attempts. So if you have any feedback, we’d love to hear it. (Please leave it as a comment or message us.)

See also

The related thread: Who's hiring?

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: Quantifying Uncertainty in GiveWell's GiveDirectly Cost-Effectiveness Analysis, published by Hazelfire on May 27, 2022 on The Effective Altruism Forum.

Effort: This took about 40 hours to research and write, excluding time spent developing Squiggle.

Disclaimer: Opinions do not represent the Quantified Uncertainty Research Institute nor GiveWell. I talk about GiveWell's intentions with the GiveDirectly model. However, my understanding may not be complete. All mistakes are my own.

Epistemic Status: I am uncertain about these uncertainties! Most of them are best guesses. I could also be wrong about the inconsistencies I've identified. A lot of these issues could easily be considered bike-shedding.

Target Audiences: I wrote this post for:

People who are interested in evaluating interventions.

People who are interested in the quantification of uncertainty.

EA software developers that are interested in open source projects.

TLDR: I've transposed GiveDirectly's Cost-Effectiveness Analysis into an interactive notebook. This format allows us to measure our uncertainty about GiveDirectly's cost-effectiveness. The model finds that GiveDirectly's 95% confidence interval for its effectiveness spans an order of magnitude, which I deem a relatively low level of uncertainty. Additionally, I found an internal inconsistency with the model that increased GiveDirectly's cost-effectiveness by 11%.

The notebook is quite long, detailed and technical. Therefore, I present a summary in this post.

This model uses Squiggle, an in-development language for estimation and evaluation, developed by myself and others at the Quantified Uncertainty Research Institute. We'll write more about the language itself in future posts, especially as it becomes more stable.

GiveWell's cost-effectiveness analyses (CEAs) of top charities are often considered the gold standard. However, they still have room for improvement. One such improvement is the quantification of uncertainty. I created a Squiggle Notebook that investigates this for GiveDirectly CEA. This notebook also serves as an example of Squiggle and what's possible with future CEAs.

In GiveWell's CEAs, GiveDirectly is used as a benchmark to evaluate other interventions. All other charities' effectiveness is measured relative to GiveDirectly. For example, as of 2022, the Against Malaria Foundation was calculated to be 7.1x to 15.4x as cost-effective as GiveDirectly. Evidence Action's Deworm the World is considered 5.3x to 38.2x as cost-effective. GiveDirectly makes a good benchmark because unconditional cash transfers have a strong (some might even say tautological) case behind their effectiveness. GiveDirectly being a benchmark makes it a good start for quantifying uncertainty. I also focus on GiveDirectly because it's the most simple CEA.

GiveWell CEAs do not include explicit considerations of uncertainty in their analysis. However, quantifying uncertainty has many benefits. It can:

Improve people's understanding of how much evidence we have behind interventions.

Help us judge the effectiveness of further research on an intervention using the Value of Information.

Allows us to forecast parameters and better determine how wrong we were about different parameters to correct them over time.

Cole Haus has done similar work quantifying uncertainty on GiveWell models in Python.

The primary decision in this work is choosing how much uncertainty each parameter has. I decided on this with two different methods:

If there was enough information about the parameter, I performed a formal bayesian update.

If there wasn't as much information, I guessed it with the help of Nuño Sempere, a respected forecaster. These estimates are simple, and future researchers could better estimate them.

Results

Methodology and calculations are in my Squiggle notebook:

/@hazelfire/giv...

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 5/26/22: I Guess I Should Respond To This Week’s Long Covid Study, published by Zvi on May 26, 2022 on LessWrong.

I’ve always at least somewhat tried to model these posts after the pure joy of America’s only true newsletter, Matt Levine’s Money Stuff. This week I finally got around to listening to Matt do this amazing podcast about the philosophical side of what he does, which I highly recommend. Covering Covid means that the subject matter is always, at its core, a combination of people dying and a portrait of civilizational collapse. The whole situation is usually rather dismaying. It is likely to remain rather dismaying permanently.

Thus it is not the ideal place to find delight in one’s understanding. Nor is writing up the rise of monkeypox or our national failure to be able to keep even normal baby formula in stock while every other country on Earth has no such issue and most of them are happy to sell formula to us, which has now extended to our family being unable to find any normal brands on Amazon or other online sources, forcing us to buy organic formula instead, which was fortunately available as a still-legal form of price gouging.

I am thus excited to once again see the weekly posts decreasing in size as the amount of Covid news decreases. If this continues, soon these posts will be quick to put out, and some time after that perhaps they can stop being weekly.

I’m hoping to pivot away from short term developments and towards more longer term, less speed premium explorations of how the world works, in places that can lead to more generally useful insight and more delight, although still with a lot of silent screaming about the ways in which things are terrible.

One key is that I like to think of finding out things are terrible, whenever I can, as good news and a source of delight. As long as we know roughly how bad things are already, identifying the sources and finding them to be crazy unnecessary idiocy is often good news. It means things can more easily be fixed.

There were some new Long Covid claims this week I felt compelled to respond to, but there isn’t a big update to make as a result.

Executive Summary

Covid-19 still exists and BA.4/5 have substantial immune escape.

Thus I keep doing these posts every week.

Ideally they keep slowly getting shorter.

Let’s run the numbers.

The Numbers

Predictions

Prediction from last week: 700,000 cases (+17%) and 2,100 deaths (+5%)

Results: 643k cases (+7%) and 2,337 deaths (+17%).

Prediction for next week: 700,000 cases (+9%) and 2,725 deaths (+15%).

Note that Florida has been adjusted to account for its reporting schedule. It still saw a large increase. Also Vermont failed to report so I gave it last week’s number.

The surprise was an active decline in cases in the Midwest and East, without any holiday to explain the change. It seems odd for things to reverse in this way, but the prediction has to hedge its bets.

For deaths, cases have been climbing for a while so now that we see the numbers trending up and we’ve wound through our backlogs it makes sense for the number to keep rising for a it.

Deaths

Cases

One could explain the different trajectories using weather or one could say that some places have already peaked because different variants spread in different places at different speeds and times, or both. It could also be some sort of reporting artifact, I’m not yet sure what to make of it.

The Never-Ending Pandemic

Here we go again with slight variations edition: A prediction that Covid will not only stick around, but ‘infect most people several times a year.’

The central problem is that the coronavirus has become more adept at reinfecting people. Already, those infected with the first Omicron variant are reporting second infections with the newer versions of the variant — BA.2 or BA2.12.1 in the Unite...

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: Do you need to love your work to do it well?, published by lynettebye on May 26, 2022 on The Effective Altruism Forum.

The Peak behind the Curtain interview series includes interviews with eleven people I thought were particularly successful, relatable, or productive. We cover topics ranging from productivity to career exploration to self-care.

This second post covers “Do you need to love your work to do it well?”

You can find bios of my guests at the end of this post, and view other posts in the series here. This post is cross posted on my blog.

Do you think that loving your work is correlated with or necessary for doing it well?

It’s important for doing research

I think that having an inside view vision for research is very important for getting it done right. It isn't enough that somebody argues to you that it would make sense to do a certain type of research. It needs to grab you and be generative in your brain in a way that's hard to force. That definitely feels important for a research role. With that being said, I don't think every day or week feels like you're grabbed with something exciting and really like it.

Like I said, there's long periods of being for me, frustrated and lost in a fuzzy domain. I think it helps to just have patience to ride that out and be like “I am probing a lot of different directions. I am waiting for the feeling of traction.” That often is accompanied by joy, and excitement, and energy, and being in a mildly manic state, but that's not the whole job. A lot of the job is systematically and patiently poking around trying to cause that to happen. Then also a lot of the job is riding it out and following it through. I don't generally tend to think in a frame of loving the whole job but I do think just having energy is almost pathologically valuable for getting a lot of good stuff done.

Ajeya Cotra

Research goes better if you love it

It's definitely correlated with it. I think just one aspect of that is I think curiosity is really important as a form of motivation. I think I’m possibly excessively curious but I think it's definitely necessary for getting ideas, and seeing stuff that's interesting or confusing to you, or having motivation to read broadly and find different stuff that comes together. I think if you're a curious researcher, you just find yourself constantly being curious about stuff, then it's pretty likely that you're going to be someone that enjoys research. It's like a method of satisfying your curiosity often.

I do think it probably is possible, especially for certain kinds of research, to do a good job even if you're not that into it. I think deep anxiety might also work for some people. I would not recommend this. I would definitely strongly recommend loving it.

Ben Garfinkel

You can do work just because it’s important, but it’s easier if you’re engrossed

You probably should be doing work that you really enjoy. Right now I feel like I'm not enjoying being a research engineer as much as I’d like, and that definitely makes me worse at it. I'm in this weird position where that does concern me pretty significantly, but also it maybe feels outweighed by the urgency of what we're doing and the lack of people to replace me.

If you want to do great intellectual work, you want to be engrossed with something to the point where it is often occupying your shower thoughts or your-wake-up-at-5:00-AM thoughts. You want to be obsessed with it to the point where you're not just solving the next task, but you're also, without too much effort, developing your ideas for the bigger picture or the vision for what could be. Also, I think it just means that you can spend a lot more time productively working on something if you're just curious about it or driven by it or whatever.

Daniel Ziegler

Enjoying work makes it easier to do well

I thin...

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: Right now, you're sitting on a REDONKULOUS opportunity to help solve AGI (and rake in $$$), published by Trevor1 on May 26, 2022 on LessWrong.

TL;DR the contest is right here, it closes tomorrow (Friday May 27th), and you have good odds of getting at least $500 out of it. Almost nobody knows about it, and most submissions are crap, so if you think about AGI a lot then you probably have a serious edge in getting a chunk of the $20k prize.

The Problem

I've spent about four years in DC, and there is one problem that my connections and I keep running into; it's always the same, mind-numbingly frustrating problem:

When you tell someone that you think a supercomputer will one day spawn an unstoppable eldritch abomination, which proceeds to ruin everything for everyone forever, and the only solution is to give some people in SF a ton of money... the person you're talking to, no matter who, tends to reevaluate associating themself with you (especially compared to their many alternatives in the DC networking scene).

Obviously, AGI is a very serious concern, and computer scientists have known for decades that the invention of a machine smarter than a human is something to take really seriously, regardless of when it happens.

We Don't Do That Here

Although most scientists are aware that general intelligence has resulted in total dominance in all known cases (e.g. humans, chimpanzees, and bears), many policymakers find it unusually difficult to believe that intelligence is so decisive, in large part due to the fact that bureaucracies, in order to remain stable, need to generate unsolvable mazes that minimize infiltration by rich and/or intelligent opportunists.

It's become clear to many policy people that one-liners and other charismatic statements are a bottleneck for creating real change about AI. This isn't due to desperation or fire alarms or short time horizons or anything like that, it's because one-liners and short statements are so critical to success in the policy space. Working papers are generally best for persuasion, and they can be pretty damn charismatic too; but in some environments (e.g. government), people's attention spans are very short. Especially when the topic of unstoppable eldritch abominations come up.

The Solution

I don't actually know what org this is, they don't state it, but they will pay you to either write up or connect them to really clever one-liners and paragraphs that clearly and honestly make the case about AGI.

If you write a really good paragraph or one-liner, and it's in the top 40 of the best ones, then they give you $500. If you submit two that make it to the top 40, you get $1000. If you put in four of the top 40 ones, you get $2000. And so on.

Right now there's more than 300 comments, but that shouldn't intimidate you. A majority of them aren't entries, and maybe a majority of them aren't entries (I personally made at least 30 comments that don't count as entries at all).

If you go and select a random area of the comment space and read all the comments there, you'll probably make a rather lucrative discovery: most of the contest's entries are total garbage with zero chance of winning. With half an hour of effort, can probably make it into the top 40 and get $500 with a single entry, especially because all you need to do is find and copy a really clever twitter tweet about AGI safety.

I Don't Want To Live On This Planet Anymore

This scenario is extremely upsetting and disturbing to me. These people have invested a ton of money into a really important and really tractable problem (the absurdity heuristic), and they have opened up the contest to a ton of really smart people; and more than half of the entries into the contest are literally just 6-word slogans. Not the good kind of 6-word slogans, the kind of slogans that you'd hear chanted ...

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: Here's a List of Some of my Ideas for Blog Posts, published by lsusr on May 26, 2022 on LessWrong.

I will never run out of ideas for blog posts because blogging generates ideas faster than I can blog. The number of ideas for blog posts I have massively outnumbers the number of blog posts I have written. To illustrate my point, here are some of the ideas I have for blog posts.

The list is written in reverse order. Newer ideas are at the top.

List of Blog Ideas

Writing is anti-inductive

Farming long-tailed upside

Re: Misinformation About Misinformation

Modern straps

Re: "Instead of competing with everyone else for the attention of the median person, it's often worthwhile to make something just for smart people. They may be a small fraction of the population, but there are still a lot of them in total."

Book Review of MCTB

Study notes from Molecular Biology of the Cell

Finish translating The Art of War and the other 6 military classics

"Material Construction", a series of posts about science that starts by explaining how atoms work and then builds everything out of that

The Virtue of Trolling

How Modern Art Works

Aluminum fuel cells

Evaporative Cooling in AI Alignment

Predictive Processing and Placebos

Japanese stuff in Star Wars

Biting Fingernails

Geoarbitrage

Dilute your Kool Aid with more Kool Aid

Re: PG's How to Lose Time and Money

managing client expectations

Wave harmonics and brain synchronization

Rationality is someone who cares more about being right than feeling right

Against the intrinsic value of money

What would fractional reserve banking look to a predictive processor?

Meeting writers IRL

Learning the Others' language よろしくおいします

Advice for Smart People with Autism

How to get an internship

Re: this comment by Razied about meditation

Re: Mandatory Secret Identities

Re: Going from California with an Aching in My Heart.

Star Wars and Marvel doing same thing with expanded universe

Trump symbiosis

Interview with Bryan Caplan

Everyone should be like us

Russia showing torrents in theaters

Adult Responsibility

Collective Guilt

Re: Selfish Reasons to Have More Kids

Money is commoditized coercion

Proper Repro Steps

Absolute reference points. Talking about what people are talking about is a trap.

a series on Rationality

Minimum Viable Skynet

Unbelievable stories go into fiction

Busywork

Paperclip Maximizer Minimizers

Pedagogical techniques from different domains

Noninterference with Primitive Minds. Let them figure out your own values.

The Wisdom of my Lyft Driver. "What's something you think I should know?"

The 4-Hour Cult Leader

Buzzwords. Unnecessary when there's universally-understood word instead. But shows you know the new stuff. Illustrate in Effective Evil.

Post-Post Modernism. Post-Post-Post-Modernism. End of philosophy. We have solved all the big problems.

Copying code from StackOverflow. Use as metaphor for rationalist debate failure modes. Also alien technology imported from China. Link to Americans Are Like Space Aliens post.

Daoist Zero-think

Just copy my conversation with my squire. Task switching cost of autism.

The confusion cycle to learning philosophy

Physicists solve problems by figuring out the real answer. Not by pointing out where others are wrong.

Challenging kids with "I don't know if you can do it."

Stupid model of world model, agent, biases and akrasia. Satire.

Instantly hitting it off or not

The antidote to cults is to dilute your mind with all kinds of garbage

Indoctrinating against indoctrination by indoctrinating into everything

Re: How many heresies do you have?

Perfect fit or not at all

The value of being in a little bit of a rush

Cultivating impatience. Life is so, ridiculously, short. Start becoming people at fourteen-fifteen. Adults at eighteen. Three years.

Mental health model: dials on an analog computer

Delegate only order at a ti...

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: Testing Air Purifiers, published by jefftk on May 25, 2022 on LessWrong.

If you're considering buying an air purifier for reducing covid risk, the

EPA recommends comparing them based on their Clean Air Delivery Rate

(CADR) for smoke. But what if you want to evaluate a purifier yourself? Perhaps you don't trust the manufacturer, want to evaluate an off-brand filter replacement, have built a

filter cube, or put together

something weirder? How can we measure purifier performance?

What we would like to calculate is the device's smoke CADR: when given smoke-polluted air, how quickly can it clear the smoke? This is something we can test:

Make smoke.

Wait for the smoke to distribute.

Turn on the air purifier.

Track results on air quality monitor.

Burning matches is a good way to create smoke, but you get a lot of variation in exactly how much. Here are twenty four trials where I burned seven matches and measured the peak amount of 2.5µm smoke "pm2.5" with an M2000:

Even if you exclude the one trial where I got unusually small amount and the three where I got an unusually large amount, the middle twenty still varied over a range of 121-233 µg/m³, almost 2x.

While creating a consistent amount of smoke is difficult, I think it also isn't needed. A filter removes a consistent proportion of its input for each size of particle. If you put very smoky or nearly clean air through a MERV-13 filter, out the other side you should get air that has, for example, 85% fewer 1-3µm particles. Purifiers, then, should reduce particle concentrations exponentially and we should see measured pm2.5 levels decrease by a constant fraction each minute.

Do we? Let's check!

I used a Coway AP-1512HH Mighty (the

Wirecutter's top pick) and ran it on 'high' five times, tracking p2.5 levels:

Eyeballing these, the initial decrease from the peak is lower than you'd expect from the rest of the curve. I think this is partially a measurement artifact: the meter is recording once a minute, and it isn't able to identify the true peak. The highest reading we get is going to be from a minute that included some amount of increase in smoke and also some amount of decrease. Additionally, at the beginning the smoke and cleaned air are still evening out around the room. Let's skip the very beginning of each curve and normalize by counting the highest included measurement as 100%:

How close is this to exponential decay? Let's look at the minute-over-minute decreases across the five runs:

With perfect exponential decay we would see horizontal lines. It's a bit noisy, but instead I think we are seeing a general decrease, from

~21% initially to ~16% after ten minutes. This makes some sense: we are taking these measurements in a furnished room, with some obstructions to airflow. Some of the air will flow freely and most of the particles we are removing at first will come from that air. Other air will be in more awkward places, like under the bed, and smoke particles there will only gradually make it out into the general flow to be cleaned.

This allows us to compare two different purifiers (or, in this case, the same purifier on two different settings). Here's the chart above, with the addition of five runs on 'medium':

And minute-over-minute:

It looks like 'high' is moving pretty close to twice the air as 'medium'.

To calculate the CADR, I think we should use the initial somewhat steeper decrease, because that is closest to what you would get in the kind of empty room that is used for manufacturer CADR testing. We should probably also treat the room as if it's a bit smaller, to include the effect of some of the air being in hard-to-clean places, but I'm going to ignore that here. I got 21% for the AP-1512HH on

'high': how do I turn that into a CADR?

Here's the room I was testing in:

It isn't rectangular, but I estimate it's 1...

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: MERV Filters for Covid?, published by jefftk on May 25, 2022 on LessWrong.

If you look around for advice on what kind of air purifier to use to reduce covid risk, you'll see some people saying you need HEPA filters:

Microcovid:

If you decide to buy an air purifier for COVID purposes, here's some things to keep in mind:

Make sure the purifier has a HEPA filter - these are rated to remove

99.97% of airborne particles.

Central heat / AC systems don't work for this. These typically use

MERV-rated filters. These are significantly less effective at removing small particles than HEPA-rated filters (the best MERV rating, MERV-16, merely removes 75% of particles. [JK: this should be 95%—filed

#1451]

The EPA, however, advocates anything that removes 0.1-1 µm particles well, and recommends MERV-13 or better if you're building something:

In order to select an air cleaner that effectively filters viruses from the air, choose: 1) a unit that is the right size for the space you will be using it in (this is typically indicated by the manufacturer in square feet), 2) a unit that has a high CADR for smoke

(vs. pollen or dust), is designated a HEPA unit, or specifically indicates that it filters particles in the 0.1-1 µm size range.

When assembling a DIY air cleaner, choose a high-efficiency filter, rated MERV 13 or higher, for better filtration.

Where does the recommendation to use MERV-13 come from? Can I use

MERV-12? If I use MERV-14 instead how much better will it be? Let's model it.

Imagine you have an infected person who enters a room. They exhale sars-cov-2 particles, which slowly accumulate:

These particles are a range of sizes, but let's approximate that 20% of the airborne sars-cov-2 is in particles of size 0.3-1µm, 29% in 1-3µm, and 51% in 3-10µm (see How Big Are Covid

Particles?):

The particles don't actually accumulate forever, for several reasons, one of which is that they typically slowly settle out of the air. This depends on size, with half lives of ~4d for 0.3µm, ~2hr for

2µm, and ~20min for 5µm:

Another reason particles don't accumulate forever is the same reason that even with the windows closed we don't suffocate: buildings aren't airtight. You describe how much ventilation a building has in "air changes per hour" (ACH), and a typical value for residential construction is 2 ACH. For example, if your room is 1,000 cubic feet

(CF) then you might have 2,000 cubic feet of air exchanged with the outside per hour, or 33 cubic feet per minute (CFM):

Now let's imagine we turn on a purifier that runs 5 ACH (83 CFM in a

1,000 CF room) through a MERV-14 filter. This is rated to remove at least 75% of 0.3-1µm particles, 90% of 1-3µm, and 95% of

3-10µm:

The total amount of sars-cov-2 in the air is the sum of these three curves:

What happens when we run different filters with different efficacies?

We can also look at each of these filters relative to a situation with no filtration:

Let's extend the scale a little so we can see how long it takes to stabilize:

In equilibrium, we see the following amounts of sars-cov-2 relative to no filtration:

Filter Presence MERV-11 49% MERV-12 43% MERV-13 40% MERV-14 36% MERV-15 35% MERV-16 34% HEPA 33%

This is a little surprising: Microcovid estimates that in this situation ("Indoors with a HEPA filter (flow rate 5x room size per hour)") risk is reduced to 25%. They cite Curtius et al. 2020 and summarize as:

Researchers found that running air purifiers in a classroom decreased the aerosol density in the room by 90%.

This is not a great summary of the paper. The researchers did two different things:

Measured the effect of air purifiers on particle concentration.

Made a simple model of the effect of purifiers to estimate the

decrease in risk.

The 90% comes from #2, and they say "After 2 hours, the concentration of aerosol particles containing vi...