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: An EA storybook for kids, published by Simon Newstead on November 27, 2022 on The Effective Altruism Forum.

Hi folks,

Sharing a new EA storybook for kids, designed to help introduce EA concepts to kids in a fun and age-appropriate way.Althea and the Generation Tree tells the story of a free-spirited girl and her trusty sidekick Hamster, who together make a fateful discovery from the distant past.

The goal of the project is to help inspire kindness and thoughtfulness in future generations. It's a non-profit project, the e-book will be free and any proceeds from a future hardcopy version will be donated to charity.

We'd like the book to be created in a collaborative way, and we're calling for beta readers from the community: - if you have any feedback on the concept or ideas or tips, feel free to share

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: The First Filter, published by adamShimi on November 26, 2022 on LessWrong.

Consistently optimizing for solving alignment (or any other difficult problem) is incredibly hard.

The first and most obvious obstacle is that you need to actually care about alignment and feel responsible for solving it. You cannot just ignore it or pass the buck; you need to aim for it.

If you care, you now have to go beyond the traditions you were raised in. Be willing to go beyond the tools that you were given, and to use them in inappropriate and weird ways. This is where most people who care about alignment tend to fail — they tackle it like a normal problem from a classical field of science and not an incredibly hard and epistemologically fraught problem.

If you manage to transcend your methodological upbringing, you might come up with a different, fitter approach to attack the problem — your own weird inside view. Yet beware becoming a slave to your own insight, a prisoner to your own frame; it’s far too easy to never look back and just settle in your new tradition.

If you cross all these obstacles, then whatever you do, even if it is not enough, you will be one of the few who adapt, who update, who course-correct again and again. Whatever the critics, you’ll actually be doing your best.

This is the first filter. This is the first hard and crucial step to solve alignment: actually optimizing for solving the problem.

When we criticize each other in good faith about our approaches to alignment, we are acknowledging that we are not wedded to any approach or tradition. That we’re both optimizing to solve the problem. This is a mutual acknowledgement that we have both passed the first filter.

Such criticism should thus be taken as a strong compliment: your interlocutor recognizes that you are actually trying to solve alignment and open to changing your ways.

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: Why square errors?, published by Aprillion (Peter Hozák) on November 26, 2022 on LessWrong.

Mean squared error (MSE) is a common metric to compare performance of models in linear regression or machine learning. But optimization based on the L2 norm metrics can exaggerate bias in our models in the name of lowering variance. So why do we square errors so often instead of using absolute value?

Follow my journey of deconfusion about the popularity of MSE.

My original intuition was that optimizing mean absolute error (MAE) ought to always create "better" models that are closer to the underlying reality, capture the "true meaning" of the data, be more robust to outliers. If that intuition were true, would it mean that I think most researchers are lazy and use an obviously incorrect metric purely for convenience?

MSE is differentiable, while MAE needs some obscure linear programming to find the (not necessarily unique) solutions. But in practice, we would just call a different method in our programming language. So the Levenberg–Marquardt algorithm shouldn't look like a mere convenience, there should be fundamental reasons why smart people want to use it.

Another explanation could be that MSE is easier to teach and to explain what's happening "inside", and thus more STEM students are more familiar with the benefits of MSE / L2. In toy examples, if you want to fit one line through an ambiguous set of points, minimizing MSE will give you that one intuitive line in the middle, while MAE will tell you that there are infinite number of equally good solutions to your linear regression problem:

I see this as an advantage of MAE or L1, that it reflects the ambiguousness of reality ("don't use point estimates pf linear regression parameters when multiple lines would fit the data equally well, use those damned error bars and/or collect more data"). But I also see that other people might understand this as an advantage of MSE or L2 ("I want to gain insights and some reasonable prediction, an over-simplified fast model is enough, without unnecessary complications").

Previously, I also blamed Stein's Paradox on L2, but that does not seem true, so other people probably formed their intuitions about L1 and L2 without imaginary evidence against one of them. If both norms are equally good starting points for simplifying reality into models, then computational performance will be a bigger factor in the consideration that L2 is "better".

However, Root mean square error (RMSE) or mean absolute error (MAE)? by Chai and Draxler (2014) captures the biggest gap in my previous understanding - we have to make an assumption (or measurement) about the error distribution and choose our metric(s) accordingly:

Condensing a set of error values into a single number, either the RMSE or the MAE, removes a lot of information. The best statistics metrics should provide not only a performance measure but also a representation of the error distribution. The MAE is suitable to describe uniformly distributed errors. Because model errors are likely to have a normal distribution rather than a uniform distribution, the RMSE is a better metric to present than the MAE for such a type of data.

And for many problems, small errors are more likely than large errors, without an inherent bias in the direction of the error (or we can control for bias), so normal distribution is a reasonable assumption.

When we are not sure (which we never are), we can also use more than one metric to guide our understanding, evaluating our models by one over-simplified number shouldn't be enough.

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: What videos should Rational Animations make?, published by Writer on November 26, 2022 on LessWrong.

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: Don't align agents to evaluations of plans, published by Alex Turner on November 26, 2022 on The AI Alignment Forum.

Another stab at explaining Don't design agents which exploit adversarial inputs. This is not the follow-up post mentioned therein. That post will, presumably, come next.

After asking several readers for their understandings, I think that I didn't successfully communicate my points to many readers. I'm now trying again, because I think these points are deeply important. In particular, I think that my arguments rule out many target AI motivational structures, including approval-directed agents (over a rich action space), approval-based amplification (if the trained agent is supposed to be terminally motivated by the amplified overseer's ratings), and some kinds of indirect normativity.

Background material

One motif in some AI alignment proposals is:

An actor which proposes plans, and

A grader which evaluates them.

For simplicity, imagine we want the AI to find a plan where it makes an enormous number of diamonds. We train an actor to propose plans which the grading procedure predicts lead to lots of diamonds.

In this setting, here's one way of slicing up the problem:

Outer alignment: Find a sufficiently good grader.

Inner alignment: Train the actor to propose plans which the grader rates as highly possible (ideally argmaxing on grader output, but possibly just intent alignment with high grader output).

This "grader optimization" paradigm ordains that the AI find plans which make the grader output good evaluations. An inner-aligned actor is singlemindedly motivated to find plans which are graded maximally well by the grader. Therefore, for any goal by which the grader may grade, an inner-aligned actor is positively searching for adversarial inputs which fool the grader into spitting out a high number!

In the diamond case, if the actor is inner-aligned to the grading procedure, then the actor isn't actually aligned towards diamond-production. The actor is aligned towards diamond-production as quoted via the grader's evaluations. In the end, the actor is aligned to the evaluations.

Clarifications

Grader-optimization is about the intended agent motivational structure. It's about a trained agent which is trying to find plans which grade highly according to some criterion.

Grader-optimization is not about grading agents when you give them reward during training. EG "We watch the agent bump around and grade it on whether it touches a diamond; when it does, we give it +1 reward." This process involves the agent's cognition getting reshaped by policy gradients, e.g. upon receipt of +1 reward.

In policy gradient methods, reward chisels cognitive circuits into the agent. Therefore, the agent is being optimized by the reward signals, but the agent is not necessarily optimizing for the reward signals or for any grader function which computes those signals.

Grader-optimization is not about the actor physically tampering with e.g. the plan-diamondness calculator. The grading rule can be, "How highly would Albert Einstein rate this plan if he thought about it for a while?". Albert Einstein doesn't have to be alive in reality for that.

These will be elaborated later in the essay.

Grader-optimization doesn't seem sensible

I'm going to try saying things, hoping to make something land. While I'll mostly discuss grader-optimization, I'll sometimes discuss related issues with argmaxing over all plans.

An agent which desperately and monomaniacally wants to optimize the mathematical (plan/state/trajectory) ↦ (evaluation) "grader" function is not aligned to the goals we had in mind when specifying/training the grader (e.g. "make diamonds"), the agent is aligned to the evaluations of the grader (e.g. "a smart person's best guess as to how many diamonds a plan leads to").

Don't al...

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: Respecting your Local Preferences, published by Scott Garrabrant on November 26, 2022 on LessWrong.

In this post, I give a application of geometric rationality to a toy version of a real problem.

A Conflicted Agent

Let's say you are an agent with two partially conflicting goals. Part of you wants to play a video game, and part of you wants to save the world and tile the multiverse with computronium shaped exactly the way you like it (not paperclips). How should these conflicting interests figure out what to do? (Assume you have not yet had the idea of starting a video game company to save the world.)

We will assume that 2/3 of you wants to save the world, and 1/3 of you wants to play video games.

At first your world-saving-self has the bright idea that 2/3 is bigger that 1/3, and you should therefore devote all your time to saving the world. However, this proposal doesn't stick. Eventually, you end up Nash bargaining with your time, and devoting 2/3 of your time to saving the world and 1/3 of your time to playing video games. This works well for a while, but then your world-saving-self has a new bright idea:

"Let's look around at the world, and see how much ability we expect to have to save it. If it feels like we are in the top 60 percentile of worlds ordered by how much control we have, then we will try to save the world. If we are in the bottom 40 percentile, we will play video games!"

(The 60 percentile is arbitrarily rounding down from 2/3, so that you can both play more video games and save more worlds.)

A Nash Bargaining Model

Let's model this more carefully. Let's say there are five different types of worlds: 1, 2, 3, 4, and 5. In each world, you have two buttons in front of you. The video game button, and the save the world button. Each time step, pressing the video game button lets you play some video games, and pressing the save the world button saves the world with probability εi in the world of type i.

We have five degrees of freedom. For each i∈{1,2,3,4,5}, we have pi, which represents the proportion of our time in world i that we spend pressing the save the world button. The rest of the time is spent playing video games.

The part of you that wants to save the world has power 23, and utility equal to 1010100ε(p1+2p2+3p3+4p4+5p5). However, since we are Nash bargaining, the coefficient out in front does not matter.

The part of you that wants to play video games has power 13 and utility equal to 5−p1−p2−p3−p4−p5.

We are trying to maximize the weighted geometric mean, 3√(p1+2p2+3p3+4p4+5p5)2(5−p1−p2−p3−p4−p5), on the cube [0,1]5.

This achieves a maximum when p1=p2=0, and p3=p4=p5=1. Simple enough.

(Linear is not the best model for the distribution of how much control you have. I am just trying to keep things simple.)

Local Preferences

The main problem with the above analysis, according to me, is that it is not respecting the locality of some of your preferences.

Maybe your desire to save the world is nonlocal. Maybe you care equally about whether this world is saved and whether some hypothetical other world in which you have more or less control over saving the world is saved. Why should you care about this Everett branch more than the other ones? I will grant that your altruistic self thinks that way, but I am guessing your desire to play video games probably doesn't.

The part of me that wants to play video games wants to play video games in this world. If I ask it what it wants in other Everett branches, it doesn't really care much. This does not mean it is making a mistake. Indeed, if I were to go back in time before I observed what world I am in and ask it whether it wanted to play more video games, concentrated in a small number of futures, or fewer total video games equally distributed, it chooses the equal distribution. The (video game part of a) ver...

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 Frontiers in Mojibake, published by Adam Scherlis on November 26, 2022 on LessWrong.

Fun with mismatched encodings

Mojibake is the garbled text that result from character-encoding errors.

If you’ve seen text that looks like this — and I’m sure you have — then you’ve seen mojibake.

(You should be seeing something like this:

If you see something else, this post may be a little confusing and you need a new web browser.)

Computers represent text as a sequence of bytes, and “text encodings” are dictionaries that turn characters (i.e. symbols: letters, punctuation, etc.) into bytes and vice-versa.

The garbled text above is a pretty common species of mojibake. It’s what happens when em-dashes and curly apostrophes are encoded as bytes with UTF-8 (the now-nearly-universal text encoding) and decoded back to characters with Windows-1252 (an obsolete encoding that is still pretty widespread).

Windows-1252 is pretty straightforward: each character gets one byte, and there are only 256 characters so this works out.

UTF-8 is one of several character encodings based on Unicode, which (for our purposes) is a massive numbered list of over 100,000 characters. Unicode includes nearly every language in the world and a bunch of other nonsense like emojis.

UTF-8 turns each character into a sequence of up to four bytes based on its Unicode “codepoint” (position in the list). Codepoints are bigger than bytes, so you still need this translation step.

(I’m simplifying a little here, and you should be grateful for that.)

Specifically, an em-dash gets mangled like this:

EM DASH (—) is Unicode character #8,212, usually written (in hex) as U+2014.

UTF-8 encodes the number 8,212, which is too big to fit in a single byte, as the sequence of bytes 0xE2, 0x80, 0x94

Windows-1252 looks at each byte in turn and decodes them directly as the characters â, €, ” respectively.

Finally, your computer looks up the characters â, €, ” in some font file and draws the specific glyph for each of those characters in that font. (A “glyph” is the actual picture on your screen; a “character” is the abstract concept of the euro symbol or whatever.)

(I made this happen deliberately with python: '\u2014'.encode('utf8').decode('1252').)

This sometimes happens to the same text multiple times, and special characters turn into exponentially-growing strings of nonsense that overwhelm the text:

An astrological mystery

I saw this today on a used-book website:

Astrological symbols??

My first thought was that this was something like the above, but with a different output text encoding in place of Windows-1252. But this isn’t really plausible; since UTF-8 encodes dashes and apostrophes as three bytes apiece, the other encoding would have to have a multi-byte sequence for ☋, the “descending node” astrological symbol. The main problems with this are that UTF-8 is the only multi-byte encoding in widespread use, and that (AFAIK) Unicode is the only character set in widespread use that includes ☋.

Maybe it’s reversed? Text goes into a legacy encoding and comes out of UTF-8 looking like ☋? This has the same problem: UTF-8 encodes ☋ as 0xE2, 0x98, 0x8B, another three-byte sequence. No other encoding is going to use three bytes for an em dash.

A rogue font

But then I remembered something wacky about my computer.

A bunch of Unicode characters are “control codes” rather than text characters in the usual sense, like U+000A NEW LINE (inserts a line break) and U+200F RIGHT-TO-LEFT MARK (makes subsequent text appear right-to-left, like Hebrew or Arabic). The first 32 characters from U+0000 to U+001F are a set of control codes inherited from ASCII, which was designed for teletype machines. They’re mostly garbage like “END OF TRANSMISSION BLOCK” and “DEVICE CONTROL FOUR” that make no sense in a digital context.

Typically, these legac...

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: The Least Controversial Application of Geometric Rationality, published by Scott Garrabrant on November 25, 2022 on LessWrong.

I have been posting a lot on instrumental geometric rationality, with Nash bargaining, Kelly betting, and Thompson sampling. I feel some duty to also post about epistemic geometric rationality, especially since information theory is filled with geometric maximization. The problem is that epistemic geometric rationality is kind of obvious.

A Silly Toy Model

Let's say you have some prior beliefs P0 at time 0, and you have to choose some new beliefs P1 for time 1. P0,P1∈ΔW are both distributions over worlds. For now, let's say you don't make any observations. What should your new beliefs be?

The answer is obvious, you should set P1=P0. However, if we want to phrase this as a geometric maximization, we can say

P1=argmaxP∈ΔWGw∼P0P(w).

This is saying, imagine the true world is sampled according to P0, and geometrically maximize the probability you assign to the true world. I feel silly recommending this because it is much more complicated that P1=P0. However, it gives us a new lens that we can use to generalize and consider alternatives.

For example, we can consider the corresponding arithmetic maximization,

P1=argmaxP∈ΔWEw∼P0P(w).

What would happen if we were to do this? We would find the world with the highest probability, and put all our probability mass on that world. We would anticipate that world, and ignore all the others.

This is a stupid way to manage our anticipation. Nobody is going around saying we should arithmetically maximize the probability we assign to the true world. (However, people are going around saying we should arithmetically maximize average utility, or arithmetically maximize our wealth.)

Not only does arithmetic maximization put all our anticipatory eggs in one basket, it also opens us up to all sorts of internal politics. If we take a world and add some extra features to it to split it up into multiple different worlds, this changes the evaluation of which world is most efficient to believe in. This is illustrating two of the biggest virtues of geometric rationality: proportional representation, and gerrymander resistance.

A Slightly Less Silly Toy Model

Now, lets assume we make some observation X⊆W, so we know that the real world is in X. We will restrict our attention to probability distributions that assign probability 1 to X. However, if we try to set

P1=argmaxP∈ΔW, P(X)=1Gw∼P0P(w),

we run into a problem. We are geometrically maximizing the same quantity as before, subject to the constraint that P(X)=1, but the problem is that the geometric expectation is 0 no matter what we do, because P(w)=0 for any w∉X.

However, this is easy to fix. Instead of requiring that P(X)=1, we can require that P(X)≥b, and take a limit as b approaches 1 from below. Thus, we get

P1=limb1−argmaxP∈ΔW, P(X)≥bGw∼P0P(w).

Turns out this limit exists, and P1 corresponds exactly to Bayesian updating on the event X. Again, this is more complicated that just defining Bayesian updating like a normal person, so I am being a little obnoxious by treating this as an application of geometric rationality.

However, I think you can see the connection to geometric rationality in Bayesian updating directly. When I assign probability 1 to an event X that I used to assign probability 12, I have all this extra probability mass in X worlds. How should I distribute this probability mass across the X worlds? Bayesian updating recommends that I geometrically scale all the probabilities up by the same amount. Why are we scaling probabilities up rather than e.g. arithmetically increasing all the probabilities by the same amount?

Because we actually care about our probabilities in a geometric (multiplicative) way!

It is much worse to decrease your probability of the true...

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, 'EA Principles', and 'The (Longtermist) EA Community', published by Violet Hour on November 25, 2022 on The Effective Altruism Forum.

1. Intro

Two weeks ago, I was in the process of writing a different essay. Instead, I’ll touch on FTX, and make the following claims.

I think ‘the principles of EA’ are, at the community level, indeterminate in important ways. This makes me feel uncertain about the degree to which we can legitimately make statements of the form: “SBF violated EA principles”.

The longtermist community — despite not having a set of explicit, widely agreed upon, and determinate set of deontic norms — nevertheless contains a distinctive set of more implicit norms, which I believe are worth preserving at the community level. I thus suggest an alternative self-conception for the longtermist community, centered on striving towards a certain set of moral-cum-epistemic virtues.

Section 2 discusses the first claim, and Section 3 discusses the second. Each chunk can probably be read independently, though I’d like it if you read them both.

2. 'EA Principles’

This section will criticize some of the comments in Will’s tweet thread, published in the aftermath of FTX’s collapse.

I want to say that, while I’ll criticize some of Will’s remarks, I recognize that expressing yourself well under conditions of emotional stress is really, really hard. Despite this difficulty, I imagine that Will nevertheless felt he had to say something, and quickly. So, while I stand behind my criticism, I hope that my criticism can be viewed as an attempt to live up to ideals I think Will and I both share — of frank intellectual honesty, in service of a better world.

2.1.

From Will’s response:

“If those involved deceived others and engaged in fraud (whether illegal or not) that may cost many thousands of people their savings, they entirely abandoned the principles of the effective altruism community.” (emphasis mine)

Overall, I’m not convinced. In his tweet thread, Will cites various sources — one of which is Holden’s post on the dangers of maximization, in which Holden makes the following claim:

“I think “do the most good possible” is an . important idea . but it’s also a perilous idea if taken too far . Fortunately, I think EA mostly resists [the perils of maximization] – but that’s due to the good judgment and general anti-radicalism of the human beings involved, not because the ideas/themes/memes themselves offer enough guidance on how to avoid the pitfalls.”

According to Holden, one of EA’s “core ideas” is a concern with maximization. And he thinks that the primary way in which EA avoids the pitfalls of their code ideas is through being tempered by moderating forces external to the core ideas themselves. If we weren’t tempered by moderating forces, Holden claims that:

We’d have a community full of low-integrity people, and “bad people” as most people define it.

Here’s one (to me natural) reading of Holden’s post, in light of the FTX debacle. SBF was a risk-neutral Benthamite, who describes his own motivations in founding FTX as the result of a risky, but positive expected value bet done in service of the greater good. And, indeed, there are other examples of Sam being really quite unusually committed to this risk-neutral, Benthamite way of approaching decisions. In light of this, one may think that Sam’s decision to deceive and commit fraud may well have been more in keeping with an attempt to meet the core EA idea of explicit maximization, even if his attempt was poorly executed. On this reading, Sam’s fault may not have consisted in abandoning the principles of the EA community. Instead, his failings may have arisen from the absence of normal moderating forces, which are external to EA ideas themselves.

Recall Will’s statement: he claimed that, conditional on Sam committing frau...

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: Does putting kids in school now put money in their pockets later? Revisiting a natural experiment in Indonesia, published by droodman on November 25, 2022 on The Effective Altruism Forum.

Open Philanthropy’s Global Health and Wellbeing team continues to investigate potential areas for grantmaking. One of those is education in poorer countries. These countries have massively expanded schooling in the last half century. but many of their students lack minimal numeracy and literacy.

To support the team’s assessment of the scope for doing good through education, I reviewed prominent research on the effect of schooling on how much children earn after they grow up. Here, I will describe my reanalysis of a study published by Esther Duflo in 2001. It finds that a big primary schooling expansion in Indonesia in the 1970s caused boys to go to school more — by 0.25–0.40 years on average over their childhoods — and boosted their wages as young adults, by 6.8–10.6% per extra year of schooling.

I reproduced the original findings, introduced some technical changes, ran fresh tests, and thought hard about what is generating the patterns in the data. I wound up skeptical that the paper made its case. I think building primary schools probably led more kids to finish primary school (which is not a given in poor regions of a poor country). I’m less sure that it lifted pay in adulthood.

Key points behind this conclusion:

The study’s margins of error” — the indications of uncertainty — are too narrow. The reasons are several and technical. I hold this view mostly because, in the 21 years since the study was published, economists including Duflo have improved collective understanding of how to estimate uncertainty in these kinds of studies.

The reported impact on wages does not clearly persist through life, at least according to a method I constructed to look for a statistical fingerprint of the school-building campaign.

Under the study’s methods, normal patterns in Indonesian pay scales and the allocation of school funding can generate the appearance of an impact even if there was none.

Switching to a modern method which filters out that mirage also erases the statistical results of the study.

My full report is here. Data and code (to the extent shareable) are here.

Background

The Indonesia study started out as the first chapter of Esther Duflo’s Ph.D. thesis in 1999. It appeared in final form in the prestigious American Economic Review in 2001, which marked Duflo as a rising star. Within economics, the paper was emblematic of an ascendant emphasis on exploiting natural experiments in order to identify cause and effect (think Freakonomics).

Here, the natural experiment was a sudden campaign to build tens of thousands of three-room schoolhouses across Indonesia. The country’s dictator, Suharto, launched the big push with a Presidential Instruction (Instruksi Presiden, or Inpres) in late 1973, soon after the first global oil shock sent revenue pouring into the nation’s treasury. I suspect that Suharto wanted not only to improve the lot of the poor, but also to consolidate the control of his government — which had come to power through a bloody coup in 1967 — over the ethnically fractious population of the far-flung and colonially constructed nation.

I live near the Library of Congress, so I biked over there to peruse a copy of that 1973 presidential instruction. It reminded me of James Scott’s Seeing Like a State, which is about how public bureaucracies impose homogenizing paradigms on the polities they strive to control. After the legal text come neat tables decreeing how many schools are to be built in each regency. (Regencies are the second-level administrative unit in Indonesia, below provinces.) After the tables come pages of architectural plans, like the one at the top of this post.

T...