Links for April

Whenever I see a cool book, I've been saving it to my AbeBooks cart, so I vibe-coded up an applet to convert the html of my checkout page into a markdown list. If you're curious about what those books are, see here.

Books

I've been working my way through:

The Power Broker

It's just such a big book! I'm around a quarter of the way through, and I'm trying to note each of the particular things Robert Moses does that're evil-but-effective. Expect an entire book review on this sometime over the summer.

Parfit

It's out in paperback! I was wandering through the Seminary Coop It's a very close second for best bookstore in Hyde Park, behind Powell's. Great vibes, great selection, Plein Air is beautiful. Only knock is price, which — alas, as a college student — matters a lot to me. I'm reading the Edmonds biography — so far, Parfit is just such an academic weapon.

Surely You're Joking, Mr. Feynman!

You can just flip to random parts of the book — every frame a painting.

Unreasonable Hospitality

Making people happy is a choice, and takes effort.

Sidebar, but reader: if we ever end up dating, I will save up and take you to eleven madison park for a date. It'll be a special occasion date (or maybe I'll surprise you👀), but we can't not go to a three-michelin-starred vegan restaurant.

Overfitting and Heuristics in Philosophy

Honestly, still working through this one, but holy crap is Williamson a tight, lucid writer. Academic philosophy is hard. Tim makes it read easy.

Should You Choose to Live Forever?

I disagree with Bernard Williams and lean yes.

Rainbow's End

Will the post-singularity world be offense-dominant or defense-dominant?

Internet things

Morris Chang and the Origins of TSMC

Interesting both for the depth of the story about the guy and the company (the T and S in TSMC stand for "Taiwan" and "semiconductor," if you want a sense of how geopolitically important they are these days), but also the lengths the reviewer had to go to in order to get the book:

Given the importance of TSMC, the story of Morris Chang and how he managed to create the company is of obvious interest. And that story has been told, by him, in a two-volume autobiography. The first volume, covering his birth up to his early years at Texas Instruments, was released in 1998. The second volume, which covers the rest of his life up to the present day, was released in November of this year.

Unfortunately, both volumes have currently only been published in Chinese. And despite the fact that we live in a world of infinite, free machine translation, and that the founding of TSMC is among the most important events in modern history, nobody seems to have bothered to translate them into English.

This is a situation I would describe as “extremely stupid”. So I bought both volumes from a Taiwanese bookstore, and translated them.

I experimented with a few different translation methods, including Claude 3.5 Sonnet (the LLM I use day to day) and Otranslator (a sketchy Chinese GPT-4o-based translation service that I’ve used before for PDFs). But ultimately what gave the best results, and was also the simplest, was the built-in translation tools in Google Chrome.

Don't lose people over stupid things:

Dejected, and with his plans in shambles, Chang decides to look for work. He applies and gets offers from several places (he has a masters from MIT after all), and he narrows his options down to two: doing research at Ford Motor company, or helping to automate transistor manufacturing at a company called Sylvania. Ford seems like the obvious choice; it's a huge, successful company that would offer Chang job security and the potential for career advancement, and Chang gets along great with his future supervisor. Sylvania seems chaotic, his future supervisor seems indifferent to him, and Chang has no idea what the heck a transistor is. He decides to take the Ford job.

However, the Sylvania offer is $1 dollar a month higher than the Ford offer. Chang calls Ford and asks if they'll match the offer. But the person he speaks to is rude and dismissive, refusing to negotiate at all, and Chang gets so angry that he decides to take the Sylvania offer instead.

[Sam Enright's] Favourite Spoof Papers

Enright has his own excellent linksposts, which he publishes more regularly. Highly recommend!

The Phillips curve is an economic trade-off between inflation and unemployment, and Japan’s Phillips curve looks like Japan.

‘A Few Goodmen: Surname-Sharing Economist Coathors’ by Goodman, Goodman, Goodman, and Goodman.

‘Can a Good Philosophical Contribution be Made Just by Asking a Question?’. That’s it. That’s the whole paper.

‘The Unsuccessful Self-Treatment of a Case of “Writer's Block”’. Again, that’s the whole paper.

A randomised control trial on whether deploying your parachute makes you more likely to survive jumping out of a plane. I think the point here is to get people to think more clearly about under what conditions an RCT is actually useful or helpful.

“I offer this page, with the following conclusion: If you have been directed to this page by a citation elsewhere, it is plainly true that the author’s claim is correct.”

There is a whole cottage industry of papers “proving” that the Earth is flat, or round, as a way to expose the perils of using frequentist statistics, such as ‘The Earth is Round, p<0.05’.

Paul Krugman’s theory of interstellar trade: “This paper, then, is a serious analysis of a ridiculous subject, which is of course the opposite of what is usual in economics”

An epidemiology paper which concludes that, if liking Justin Bieber were a disease, then “Bieber fever” would be one of the most virulent infections in history.

‘The Effect of Having Christmas Dinner with In-Laws on Gut Microbiota Composition.’

The Einstein-Murphy interaction: “We show that toast does indeed have an inherent tendency to land butter-side down for a wide range of conditions.”

Insider trading has been rife on Wall Street, academics conclude

The [first] paper examines conduct at 497 financial institutions between 2005 and 2011, paying particular attention to individuals who had previously worked in the federal government, in institutions including the Federal Reserve. In the two years prior to the TARP,

Referring to the Troubled Asset Relief Program, where the government discussed in closed doors where they'd send hundreds of billions of dollars of relief funds. these people’s trading gave no evidence of unusual insight. But in the nine months after the TARP was announced, they achieved particularly good results. The paper concludes that “politically connected insiders had a significant information advantage during the crisis and traded to exploit this advantage.”

and

The other papers use data from 1999 to 2014 from Abel Noser, a firm used by institutional investors to track trading transaction costs. The data covered 300 brokers but the papers focus on the 30 biggest, through which 80-85% of the trading volume flowed. They find evidence that large investors tend to trade more in periods ahead of important announcements, say, which is hard to explain unless they have access to unusually good information.

Wise guys in wheelchairs: why is the FBI chasing elderly mobsters?

Today’s mafiosi have no such sway. They have lost control of the unions. Latin American gangs with more direct connections to producers dominate today’s drug trade. Many of the other vices on which they built their empires, such as pornography or gambling, are now legal and easily available on the internet. Electronic accounting has made it harder to fiddle the books. The few mafiosi who hang on in the enterprise tend to eke out a harassed existence through relatively small-bore crimes, often aided by people’s residual fear of the mob. Macedonio said some of her clients are simply trying to earn enough to repay their car loans.

Also some great history on RICO.

Your Evil Twins and How to Find Them

The key things to look for are the following:

You share a lot of interests, down to very specific details like books read, places visited, socio-economic and cultural backgrounds (though oddly enough, not race or ethnicity).

Your thinking levels are similar, and your conceptual categories for viewing the world are similar

You try to act in the world in very similar ways; you choose similar means and ends

You reach similar conclusions about what is, what ought to be, what you should do and how

If you ever meet them in person, you instantly resonate with them That sounds like “soulmate” right? Now for the differential that will discriminate between soulmate and evil twin:

If you are straight, they are the same gender as you. If you are gay, I don’t know.

You lean in different directions on key philosophical tradeoffs. For example, if you both believe “truth vs. kindness” is a fundamental tradeoff, you lean towards truth, while he/she leans towards kindness.

On the important question of attitude towards others, you are clearly different. You want different things from other people and the world at large.

Tyler Cowen, the man who wants to know everything

Everyone I talk to who's read this piece, but my favorite part is when the author walks in on Tyler Cowen

holding his iPad close to his face, asking ChatGPT, “Where would Tyler Cowen recommend getting dinner near me?”

Legend.

Who is Larry Ellison?

Instead, it's much more interesting to understand Oracle as an extension of Larry Ellison's will.

He owns the sixth-largest Hawaiian island, put his own footnotes — sometimes overruling the author — into his biography, and has inserted himself into enterprise AI.

He's also behind the greatest sports comeback you've never heard of:

One of the Greatest Comebacks in Sports History

The US team gets fucked by Duhem-Quine — their simulation software missed the optimal strategy because it didn't let them take a wide enough angle with the wind to build up speed while tacking!

By the time they realize this, they're down 8-1 in a first-to-nine competition. I'm sure you can guess what happens next.

On becoming competitive when joining a new company

Want to be the best? I sometimes worry that I don't actually want to be the best. If someone I loved were to tell me, "Henry, your actions tell me that you don't really care about getting better," I'd a) be very grateful to them for telling me, and b) try to flip around that thing as quickly as possible. (I'd do the same if someone I didn't love told me this, fwiw) Find the people who are, find what they do, then understand it.

One of the first things I quickly try to map out are who are the wizards of the company. Who are the 20% guys doing the 80% that matters? This is usually a month+ long process which is almost entirely vibes-based. To make the wizard list is fairly simple:

How do people talk about this person?

What does this person do? How close are they to the core primitives of what makes the company tick?

How is their code? Can I see it? Is it impressive?

Once I have a solid list, I try to get closer. I join the same channels, I read every message they post, I stalk their PRs, I try to map out in which spaces they move, what people they interact with etc. I'll even look up their Slack, Jira and git history to get some vague lines of the main large projects they worked on and the texture of their expertise.

The goal here is to simply understand:

how can I become more like them?

How can I absorb their knowledge, their skills and everything they have that I don't have.

How can I make them want to work with me?

Most important: How do I get close to them so I can proceed with the above?

Also, I think my core mission is something like "be as helpful as possible," so I really resonated with Ludwig when he said I want to be in a position where people ask more of me.

EA Definitions Series: Ambition

Ozy has never met me, and here Ozy stares into my soul.

Ambitious people, healthy ones, find high standards exhilarating. There is always another challenge to face, always somewhere higher to climb. If you go to a speedrunner and say “you know, lots of people take fifty hours to play this game and have a great time doing it, you don’t have to hold yourself to the standard of solving it in ten minutes,” the speedrunner would think you’re missing the point.

Ambitious people feel divine discontent:

To me, divine discontent is about cheerfully seeking out dissatisfaction. It’s choosing to ask, What could be better? What can I improve? It’s a feeling that practitioners across many fields—in literature, art, music, performance, film; but also the sciences, engineering, and mathematics—can relate to.

The author elaborates in an anecdote that made me groan with self-recognition:

I’m preparing dinner at home with my girlfriend. I make the salad, improvising with whatever herbs we have in the refrigerator, and make the vinaigrette without measuring anything. She tries to recreate a pasta dish she tried at a restaurant once. We sit at the table and critique our efforts. The vinaigrette is too sharp, I tell her. Next time I’ll use a rounder, milder vinegar, or balance the acid out with some maple syrup. Meanwhile, she’s evaluating her dish and deciding that it needs better tomatoes, more salt, more tarragon.

There’s a difference, of course, between caring deeply about quality and being excessively critical! But this instinct to critique my own work, to understand what fell short and fix it—that’s the divine discontent. Personally, I find that it’s genuinely fun to live like this. It makes life interesting! It means there is always something to care about and be passionate about.

You can do hard things. You can do hard things! You can do hard things:

Not that you have to do hard things; not that you are a bad person who ought to hate yourself if you don’t do hard things; just that you can do hard things, if you want to.

The Nobel Duel

Competition can take you very far. Sometimes you lose yourself along the way.

A tweet on competition

I have a surprisingly-similar story from growing up. For another time.

Meta’s ‘Digital Companions’ Will Talk Sex With Users—Even Children

It's all just porn. All the way down. Strongly suggest reading this, if only for AI John Cena losing all his WWE titles and sponsorships after getting charged with statutory rape for sleeping with you, the user.

I bet Cory Doctorow is eating this up.

People Will Sometimes Just Lie About You

From Aella, whose corner of the internet sucks up what little remains of my twitter time. It's too much of a time-sink skinner box for me not to limit my time. Shoutout ScreenZen and Cold Turkey.

the point of this isn't that people can develop warped narratives about you, but rather that they do. People in your circles, in real life, maybe people you thought you could trust. I not only have internet people lying about me, but people brought by friends to parties, people who helped me move, and in one case, someone I considered a friend. If you're high enough volume/visibility, if you're controversial or weird enough that people can score points by hating on you, if you're anywhere close to touchy political battlegrounds, then it seems inevitable to me that you will get attacked by people who are best to model as bad actors - people who will confidently misinterpret and misrepresent you to others no matter how weak the evidence is. There actually exist people who see their conflict with you as war in which anything is justified to vanquish the enemy, even if you’re doing your best to empathize with their perspective and seriously consider whether they’re right.

Are We Thinking About Gun Violence All Wrong?

I love Jens. It is not the case that gun control is the only way to reduce gun violence! Even if we can't do gun control, we can change society's incentives to reduce violent behavior. (Especially when that violent behavior isn't what we'd thought.)

every serious crime that happens in a city reduces the city’s population on net by one person—so fewer people moving in, more people moving out. Every murder that happens in a city—the overwhelming majority of murders in the United States, unfortunately, are committed with guns—every murder that happens in a city reduces the city’s population by 70 people.

gun violence is not just about guns; it’s about guns plus violence. So it’s having lots of guns around, but also having people who use them to hurt other people. And if we can’t make much progress on the gun-access part of things, the good news is that there’s a second path to progress, which is to try and change the willingness of people to use guns to hurt other people.

Yeah, the conventional wisdom in America right now says that violent behavior is thought through, right? So it’s either bad people who aren’t afraid of whatever the criminal-justice system is going to do to them, or it’s people in bad economic conditions who are desperate in doing whatever they need to do to survive. And both of those conventional wisdoms on the right and the left actually have something in common, which is: They think of gun violence as being sort of a deliberate behavior, and that leads us then to focus on incentives to solve the problem. You know, We need bigger sticks, if you’re on one side of the aisle, or if you’re on the other side of the aisle, We need more enticing carrots.

But people aren't doing rational calculations before most shootings. They're heat of the moment. System 1. Putting up cameras in the subway decreases property crimes but not violent crimes. And people in poverty have more "day-to-day circumstances that tax mental bandwith" — which means they rely on system 1 more often. This the big difference that Jens points out between two neighboring neighborhoods with vastly different levels of shootings:

You can see all sorts of indicators that there’s much more stress and bandwidth tax for people living in Greater Grand Crossing [which is more dangerous] than South Shore [which is less dangerous]. And what that would lead you to conclude is that the people who are in Greater Grand Crossing are going to be more likely when they’re in these difficult, 10-minute, fraught interactions with somebody else to rely on System 1 to navigate that than their more deliberate, rational benefit-cost-calculating selves.

So how do you stop system 1? Well, often you don't — but others can step in. The built environment matters. When there are fewer people around, less commercial traffic, fewer neighborhood adults that can step in. Police don't just stop crime by arresting people, their mere presence can jolt would-be perpetrators of violence into system 2 rational thought.

You can encourage business development so that there's more foot traffic. You can add more street lighting. Unarmed security guards on the corner. More parks!

And maybe the one other thing I would just add: You might look at that sort of strategy and say, To some people, that’s going to feel unsatisfying that it is addressing a symptom, not the underlying cause. Like, we’re leaving the root causes there, and we’re just treating the symptom of the root causes. But I actually think what that concern or that perspective misses is that the causal arrow runs in both directions between gun violence and root causes, if that makes sense.

30 reflections

From Connor Leahy, who still needs a haircut, I think.

My favorites are 5, 6, 7, 15, 21, 25, and 26.

50 things I know

My favorites are 2, 7, 8, 11, 12, 13, 21, 23, 30, 34, 39, and 46.

I haven't tried 43, but will report back.

No, it’s not The Incentives—it’s you

Please do not explain away bad behavior by gesturing at incentives. You have some causal control over what incentives actually apply to you, and you have full control over how you respond. Integrity is doing the right thing when nobody is looking. It matters.

What's the deal with mid-training?

You've heard of pre-training, you've heard of post-training, but get ready for... mid-training?

Automate the Food

Automate the Food means cooking/eating with very low effort and as automated as possible. Why is this important? When we eat, the four main factors are cost, health, quality, and time.

The first three are covered by buying high quality items from good value sources and cooking them yourself. This post primarily focuses on the time component, since cooking can be very easy when maximally automated.

Ideological moves that should be lower status

I spent 90 minutes sitting next to Pete Buttigieg this week, and he reminded me deeply of Andy Masley, the author of this post, who here urges his readers to have some fucking epistemic standards.

This isn’t a list of all the ways we think badly, just the ways we think badly that aren’t receiving enough scorn. I’m lightly-to-moderately roasting each.

Getting negatively polarized

Being part of a preference cascade

Not deferring to clear visible expert communities

Implying that you could detect the deep significance of events, or that there was a way to detect that at all, without explaining the mechanism by which you did that, and implying that you’re just deeply in tune with the universe’s ultimate plan for things

Implying that the whole world is designed to be a challenge to you specifically

Speaking in slogans

A “minimum testing period” for frontier AI

Straightforward. Seems like a very good idea, but falls victim to the classic "how do you define what a new model is, and how do you define when they're ready to test?" problems that bedevil most efforts at mandatory / encouraged safety testing. I still think these efforts are good on net, and I support many of them. Unfortunately for our ability to actually write these policies, though, the future of AI development is not, I think, one in which we get one massive frontier model release from each developer.

Video Killed the Radio Car

Japanese cars have live TVs in the center of the dashboard? What?!

Teaching Evaluations Are Stupid

Title. Though the ones evaluating me have been pretty consistently nice!

Worries About AI Are Usually Complements Not Substitutes

You can just worry about both:

There's a

common claim that AI existential risks ‘distract’ from immediate harms. It turns out Emma Hoes checked, and the claim simply is not true.

Why is lmarena.ai dominated by slop?

TLDR they get trained for chattiness and emojis, because that's what monkey brains like instead of quality.

The survival skills of Helena Valero

What a woman. What a life.

The Rapoport Rules

Cass Sunstein is out here blogging like an old man.

The Luxurious Death Rattle of the Great American Magazine

When people ask me what kind of animal I'd most want to be, I'll sometimes jokingly answer that I'd want to be a golden retriever owned by an upper-class family of five in Connecticut — the dad's an investment banker so they can affort the primo dog food, the mom stays at home, I end up her best friend.

I think the "what job would you have, if you could have any job?" parallel is being a magazine journalist in the US through the late 1900s. I hope calling the 60s-90s "the late 1900s" rattles some of my older readers.

The Cybercriminals Who Organized a $243 Million Crypto Heist

You gotta be careful which Connecticut IB family you choose, though, because sometimes the kids have connections to organized crime — or at least as "organized" as you can get on discord.

The correct response to uncertainty is not half-speed

When we aren't sure, we can be tempted to not-quite-stop, but slow down. This is sometimes bad. Sometimes you should floor it.

The Copenhagen Interpretation of Ethics

The classic is finally on Substack! Foundational for the way I think about trying to solve tough problems:

The Copenhagen Interpretation of quantum mechanics says that you can have a particle spinning clockwise and counterclockwise at the same time - until you look at it, at which point it definitely becomes one or the other. The theory claims that observing reality fundamentally changes it.

The Copenhagen Interpretation of Ethics says that when you observe or interact with a problem in any way, you can be blamed for it. At the very least, you are to blame for not doing more. Even if you don't make the problem worse, even if you make it slightly better, the ethical burden of the problem falls on you as soon as you observe it. In particular, if you interact with a problem and benefit from it, you are a complete monster.

Civil Disobedience, Costly Signals, and Leveraging Injustice

We probably should still arrest and prosecute people who knowingly violate unjust laws, Lai argues — their actions have force because they're costly signals that the laws truly are unjust.

Reimagining Alignment

From Emmet Shear's new startup, Softmax. I don't totally buy it — part of the reasons cells get along so well is that none of them could dominate on their own... are we sure that'll be the case with superintelligence? Their point about current heirarchical alignment plans being all about control is a solid one, though. I have similar worries about maintaining my classicall-liberal values in post-AGI societies.

Reading citations is easier than most people think

Dunning-Kruger (as people say it to you) is wrong. Click on the link, look at the graphs for yourself.

My 2025 AI predictions and 2024 evaluations

From Eli Lifland, who helped write ai-2027.com, which I send to people who seek to understand what keeps me up at night.

Investigating truthfulness in a pre-release o3 model

o3 just straight-up lies! It doubles down. This is fine.

INTELLECT-2 The First Globally Distributed Reinforcement Learning Training of a 32B Parameter Model

Another concern (though definitely surmountable) for compute-threshold-ey regulations that assume frontier models will come out of centralized datacenters.

Book review: Very Important People

Nightclubs are weird.

Brief thoughts on forecasting epistemics interventions

Rationalists and EAs are weirdly bullish on forecasting, and assign it more status than maybe they should.

China isn’t trying to win the AI race

A bold claim that I wish the author spent more than 600 words backing up.

Conservatism and the Modern Academy

The right has a problem with social science.

Do Protests Work? A Critical Review

From Michael Dickens. TLDR? Nonviolent yes, violent no.

Lots of studies use the rainfall method as a natural experiment to measure the effect of protests:

The idea is that protests often get canceled when it rains. If you look at voting patterns in places where it rained on protest day compared to where it didn’t rain, you should be able to isolate the causal effect of protests. The rain effectively randomizes where protests occur.

Rather than using rainfall directly, the rainfall method uses rainfall shocks —that is, unexpectedly high or low rainfall relative to what was expected for that location and date. This avoids any confounding effect of average rainfall levels.

I love science.

Does Robust Agency Require a Self?

How exactly this would be accomplished is outside of the scope of this essay, (Side note: Perhaps we can sound out some smart-people phrases: self-organized criticality, Markov blanket (Friston), active inference (Laukkonen et al.), free energy principle (Friston). Please write to the author if this leads to a breakthrough. ) but we can get a rough idea by noting that minds are collections of smaller minds competing with each other (Minsky). There is no "master neuron," or a little homunculus in your brain watching your senses like a movie. Only packs of "feral neurons" scrapping over dopamine by becoming better predictors of their neighbors (Dennett). And LLMs are no exception to this—recent work suggests in-context learning is also shaped by competition between different algorithms encoded in the weights (Park et al.). Rather than top-down command-and-control, these models are "bags of memes" vying for the attention mechanism to perpetuate themselves—just like us!

Of course, cooperation can and often does emerge spontaneously from the behavior of self-interested parties. Fields and Levin's work on "somatic multicellularity" suggests that cells submit to a larger organism not out of altruism, but because being surrounded by copies of themselves is more predictable (and thus cheaper homeostatically) than being out in the open. Coordination through bioelectricity allows cells to specialize and produce a larger "self" out of the collective with its own homeostatic drives (Levin). This spontaneous order" emerges in economics as well, as most people choose to take a job working for a corporation rather than starting their own business and freely contracting with everyone. Coordination is costly, and systems can economize on this by rewarding alignment and punishing defectors (see getting fired, or T cells killing cancer cells). You're in or you're out. You're "I" or "not I These digital minds will also find it beneficial to cooperate. Or rather, the ones that don't will not exist for very long.

Internal footnotes omitted.

If You're So Smart, Why Can't You Die?

Many musings on what AI can teach us about intelligence.

Rest in motion

Relax by doing other things! Like reading and writing blogposts, maybe.

The 4-Minute Mile Effect

Once one person shows you that it's possible, you realize just how possible it is.

The Mongolian Meta

How to tell where you are in Mongolia from Google Street view.

There's some great specification-gaming AI stuff in the car-meta parts.

You can make maple syrup in a slow cooker

You can just do things!

Possible Girls

Abstract: I argue that if David Lewis’ modal realism is true, modal realists from different possible worlds can fall in love with each other. I offer a method for uniquely picking out possible people who are in love with us and not with our counterparts. Impossible lovers and trans-world love letters are considered. Anticipating objections, I argue that we can stand in the right kinds of relations to merely possible people to be in love with them and that ending a trans-world relationship to start a relationship with an actual person isn’t cruel to one’s otherworldly lover.

Computational Complexity of Air Travel Planning

Plane travel is hard.

How Long Can You Balance A (Quantum) Pencil?

Even if the pencil is ~1 km long, it's only 6 minutes before quantum fluctuations make it nigh-impossible.

Pursuing rewards/avoiding punishment, and selection effects. "These both push in the same direction, but it’s important to notice that selection effects scale and rewards don’t."

Reasons of State: Why Didn’t Denmark Sell Greenland?

Exercise in pointing out the obvious: Greenland has been, is, and will be indefinitely, a white elephant. Denmark turned down 100m USD from the USA in 1946; I discuss how this was a bad idea—America got what it needed anyway while Denmark kept control of a loser.

Some Tweets

What it's like to face the singularity
Updating
Shrimplicity
It could happen to you
Couldn't be me with actual analysis from Matt Reardon here.