I wrote about what I learned after buying a house a while back, but I’ve also been talking with some friends about what to consider beforehand. Here are my takes on two questions: whether to buy or rent, and how much to budget if you’re buying.

Buying vs renting

First, you may not have a choice. If you’re young and living in a high-cost area, you probably can’t afford a down payment on a house; if you’re looking for a single-family home in the suburbs, you won’t find one for rent in many areas (which is changing somewhat, unless some legislators have their way). For the reasons below, you definitely shouldn’t live somewhere you don’t want to just for the sake of buying property. So only ask this question if the home you want is available either way (say, a two-bedroom apartment or condo near a major city) or if you’re flexible between types (say, you know you’ll move to the suburbs eventually but could do it now or later).

Once you’re weighing those options, don’t fall into the trap of thinking that “renting is just throwing money away.” If you’re renting, you don’t have to make a down payment, which means you can instead invest that money in stocks (or whatever else, but really it should be stocks). And if you’re buying, a lot of what you pay goes toward mortgage interest, property taxes, repairs, and other costs that don’t directly build wealth.

In a free market, homes should be priced in a way that makes it equally attractive to invest in a house or to rent and invest your money elsewhere.1 And in fact, a study of the “rate of return on everything” over more than a century found that, after accounting for all the costs, stocks and residential housing have each returned about 7% per year. I’m sure you know someone who bought a house in like January 2020 and got wealthy immediately, but that’s not the average outcome. If you know your house will go up in value soon, of course you should buy it. But by the same token, if you know Build-A-Bear stock will go up soon, you should put the down payment there instead and rent for a while. If you’re reading this newsletter for advice, you don’t know either of those things.

So in general, neither renting nor buying is clearly better financially. Instead, the decision comes down to secondary factors—financial or otherwise—that matter more or less to you.

In favor of buying:

• Your expenses are predictable. In the US, standard mortgages have a fixed monthly payment for 30 years;2 if you rent somewhere (without rent control), you might have to pay more when you renew your lease. Similarly, buying guarantees that you’ll be able to live in the same place indefinitely, which might not be possible at any price in some rentals.

• You can customize a house you own as much as you want; rentals may have restrictions.

• To stay in your house, you have to keep paying your mortgage and building equity; this “forced savings” is helpful for some people who would otherwise be tempted to spend all their money each month instead of investing some.

• A mortgage gives you leverage on your investment. If you take out a $400k mortgage on a $500k home, and the next day its value rises to $600k, the house price has gone up 20%—but your home equity has risen from $100k to $200k, a 100% increase.

  • This cuts the other way too: if it went down by the same amount, your equity has fallen 100%. This sounds bad, but there’s a decent economic argument that young adults are better off using leverage—you can compound your gains over time, and if you get wiped out, you have a long career to make it up. It’s easier to get a mortgage than to get leverage on stocks.

  • And housing specifically is an even better investment to leverage: if you lose 100% on a purely financial asset, you’re left with nothing. But if your home equity falls by 100%, you still get to live in the house! Price changes only matter when you sell, and you always have the option of just keeping the house instead.

In favor of renting:

• The process of buying a house is a lot more expensive than renting. You generally face costs for agents, a mortgage lender, title work, an attorney, inspection, and other setup costs like changing the locks; those cost 3-6% of the home price plus a lot of your time. If you want to move later, you’ll pay a similar amount again to sell.

• If you’re renting, you don’t have to arrange for repairs or pay for them out of pocket. This is a mixed bag—you might be stuck waiting for your landlord to fix something urgent, but it means less work for you, and you don’t need to finance any large repairs yourself.

• A mortgage is a bet on the local labor market. If you work at General Motors in Detroit and you lose your job, that probably means American automakers are struggling, which means nobody wants to move to Detroit, which means your home value will also go down. If you’re renting, you can invest your assets in a diversified portfolio of stocks, giving you more security if your own industry is doing poorly. (This may not matter so much in larger or more diversified areas. If you live in NYC, anything that damages your local economy will also probably hurt all American markets, so you’re screwed whether your savings are in home equity or stocks.)

One thing to notice: many of the pros of buying become more valuable over time. You’re probably not going to get kicked out of your rental or face a huge price increase in one year, but over time there are more opportunities for something bad to happen. Customizing your house takes some upfront work, but you get to enjoy it more if you live there longer. In contrast, the biggest advantage of renting is avoiding the transaction costs, which are a one-time thing. So the relative value of buying a home goes up if you expect to stay there longer.

My three points of advice:

1. As a baseline, rent unless you’re expecting to stay for at least five years. This is a rule of thumb I heard a while ago, and it’s also about what I find when putting comparable rental and purchase prices into the NYT rent-vs-buy calculator. This reflects the large but temporary transaction costs of buying a home against the ongoing long-term benefits.

2. Understand your non-financial preferences and adjust accordingly—just be aware that’s what you’re doing. I suspect that the “renting is just throwing money away” belief is sometimes just a rationalization for other reasons that you want to buy. Maybe you really want to customize your place, or you’re terrified about being forced to move—then you should buy earlier. Maybe you really don’t want to take on debt or coordinate maintenance—then you should rent for longer. That’s great; just don’t confuse it for a financial decision.

3. The best financial investment is the house you most want to live in. Unless you’re truly a real estate expert, you won’t be able to predict future prices better than the market, so trying to pick a house that’s going to go up in value is foolish. But you do have a put option of sorts: your house is never less valuable to you than the value you get from living in it, because you can always stay there instead of selling.3 The more you enjoy living in your house, the higher that floor is, and the more valuable your investment is to you. (Caveat: this only applies to houses you can responsibly afford; paying a high interest rate for an unusual mortgage on a mansion isn’t a good investment even if you really want to live there.)

How much to spend?

If you’ve decided that you want to buy a home, it’s pretty easy to figure out how much you can afford. Use the 30% rule, the Freddie Mac calculator, or just apply for mortgage pre-approval at a high number and the bank will turn you down if they don’t think you can afford it—all of these will roughly agree in most cases.

But that doesn’t mean you should spend that much:

• You’ll probably need to spend more money on furnishing and maintaining your house, which is much easier if you have comfortable savings left over. (In the first few years we had to replace a furnace, boiler, washer/dryer, and have some trees cut down, and we also chose to install leaf filters and renovate the basement, each of which cost four or five figures.)

• Spending more on your house concentrates your savings in one property, which is riskier than a basket of stocks and makes it harder to save for education or retirement (which you may need to pay for before selling your home).

• Your income might go down in the future, and you’ll want to still be able to afford your house if that happens.

Here’s how we approached it. Our goal was to find a monthly housing cost (mortgage + taxes + insurance) where we’d still break even in a reasonable downside scenario for our income. This involves four steps:

• 1/ Estimate your long-run monthly income in a “downside” scenario. (”Long-run” meaning over a period of years, not just being between jobs for a little while.) This is necessarily speculative, but you can ground it in questions like: what kinds of work could you find if you lost your current job and how much would it pay; how much would you lose if you or your partner stopped working or went part-time to care for family; how much do you think your income will go up normally if nothing bad happens (which may increase the baseline for your downside scenario).

• 2/ Calculate your long-run monthly non-housing expenses. Start with your current budget and adjust for big things that might change: are you expecting to buy a new car, or to have kids in the future?4 Most other categories probably won’t change enough to move the needle unless you’re moving to a radically different cost-of-living area.

• 3/ Subtract #2 from #1. This is the amount you can spend on housing to get to breakeven.

• 4/ Estimate the house price that would get you to the amount in #3. Use an online calculator or your favorite large language model to estimate this for your region and any unusual factors. You don’t have enough information to calculate it exactly—your property taxes, mortgage rate, and insurance payment will depend on what home you buy—but you can get close enough.

Looking back, I’m very happy with the framework and the confidence it gave us in setting a budget. I think our specific downside scenario in #1 was a bit too conservative, so we probably could have budgeted more. (We also had some errors in forecasting expenses, but they roughly canceled out in each direction.) That said, we got really lucky with the house we found, and I can’t imagine being happier anywhere else even with a little more to spend.

1 | AI makes our jobs easier in obvious ways, but in the long run it’ll make work harder. That broke into the mainstream with a recent HBR article, but the study behind it focused on obvious incremental effects: basically, people who adopt AI tend to work harder (because they’re more productive, can work on a wider scope of projects, and can multitask more). That’s true, but more fundamental changes to the nature of work are coming:

• Work will be cognitively harder: As AI takes over more routine work (implementing code, updating spreadsheets, writing emails), everything that’s left for human workers will be non-routine—designing large systems, making hard decisions, deciding what to automate and how. That’s rewarding work, but it’s exhausting to never be on autopilot. Steve Yegge, one of the most prolific AI-assisted-coders, says he can only work 3-4 hours per day now before getting exhausted, roughly matching what Cal Newport suggests in Deep Work. (Yes, there’s an HBR study for this too.)

• Work will be emotionally harder. AI may eventually do even the cognitively hard things, but people will remain in the loop for accountability. Being held responsible when you’re not doing the work—acting as “shit umbrellas” for AI agents—is going to cause a lot of anxiety. And the agents will make all the easy decisions, meaning whenever you have to step in and act, there’s a real chance you’ll make the wrong choice and be held responsible for it (Obama writes about this).

2 | A video game developer explains pricing psychology: “What does it mean to spend five bucks? Well, that’s five bucks. But six bucks? Well, that’s still five bucks. Four bucks is also kind of five bucks. Three bucks is two bucks. And two bucks is basically free. So we’ve got these tiers: You know, twelve bucks… that’s ten bucks. But thirteen bucks is fifteen bucks. And we found that eight bucks is still five bucks. It doesn’t become ten bucks. Seven ninety nine, that’s five bucks, right? So, eight bucks going to five bucks is the biggest differential we could find in pricing, so we found it very optimal.”

3 | Everyone complains that Congress delegates too much to the executive branch, but Matt Glassman points out a systematic reason that I never thought about: the veto makes delegation a one-way ratchet. A bill to give the President authority can pass with only a majority of both houses, but a bill to rescind authority needs two-thirds (assuming presidents will veto anything that gives them less power). So more power is delegated than rescinded over time.

You could get around this by writing a bill that says, Congress gives the executive branch this authority, but it can be revoked with only a majority vote (not a new bill that could be vetoed). However, this “legislative veto” was struck down by the Supreme Court in the 1980s—though they left all the laws with a legislative veto in place and just removed the ability to revoke authority. One of those delegations was IEEPA, the authority at the heart of the Trump tariffs that the Court struck down last month. Congress passed a bill this year to revoke the President’s IEEPA authority, but it was naturally vetoed; had the legislative veto remained in place, the authority might have been rescinded without a Court case.1

The kicker, he writes, is that the Court’s opinions (in this case and others) often blame Congress for not controlling executive power without mentioning their own role. In the IEEPA case, Neil Gorsuch was the proverbial man-in-the-hot-dog-costume, pointing out the ratchet-up issue but not the judicial history behind it.

4 | This NYT piece finally changed my mind on why backing into a parking space is better than pulling in headfirst. I’d heard it was safer, but it never made sense to me until reading this: the most dangerous accidents in a parking lot involve running into a pedestrian, and people walk in the aisle, not parking spots. I still pull in sometimes (mainly if I need to get a stroller from the trunk), but I’m trying to switch when there’s no good reason not to.

(Although I still think most people are probably motivated by selfishness: when you back into the space, other cars have to wait for you, whereas when you pull in and back out, you have to wait for other cars.)

5 | Mac Barnett and Jon Klassen’s appreciation of The Very Hungry Caterpillar is delightful and filled with fun trivia. (Eric Carle originally called it A Week with Willi Worm, and the cover was just a leaf. And bug experts wrote in to complain that caterpillars build a chrysalis, not a “cocoon” as the book says.) Barnett contrasts Caterpillar‘s simple way of playing with the page structure to the flashy but fragile flaps and pop-ups in today’s kids’ books: “You get the feeling that the people making [modern books] are having more fun than the people reading it.” I admire this take but I don’t agree with it, mainly because nothing delights my seven-month-old son more than the Where’s the [blank] flap book series.

6 | After major OpenAI and Anthropic releases near the end of last year, software development almost immediately went from 80% manual, 20% AI-assisted to the reverse. A few random side projects I’ve ripped off:

• The old version of my personal website was impossible to update due to dependency issues from using an old framework. Codex recreated it from scratch with a simpler framework in about an evening, and then I spent another evening rearranging and fine-tuning it.

• I used an RSS reader to keep up with some work- and blog-relevant reading, but I didn’t check it that often, and there was too much content to read it all, so I made a little app with Claude to rank the top 5-10 links from that firehose and email me every morning, including feedback buttons on what I liked and didn’t like. The difficulty here has been filtering for “good links”—I’m not trying to read the top news about a single topic (which is easy to do with ChatGPT or plenty of other places) but the best blogs of the day across a wide range of topics. So far it’s working a little better than my previous system but I haven’t fully cracked it yet. I’ve iterated on a bunch of different versions of this but none of them took more than an hour of part-time focus.

• I was disappointed that I couldn’t find a BracketMatrix equivalent for women’s basketball, until I realized I could make my own by scraping and aggregating bracketology data from several sites. This took a bit longer than expected (a couple of the sites were in weird format, and nailing down the output took some iteration) but still was finished in about three evenings with Codex.

7 | The downside of AI programming is that today’s organizational processes aren’t really capable of managing that much code. It turns out we’ve been in a similar situation before: A tweet I can’t find taught me about the “software crisis” of the 1960s and 70s, and the “humble programmer” speech that laid out solutions. Until that point, most programming was done by individuals writing simple scripts in low-level detail. But computers were becoming powerful enough to do bigger things, which meant software was becoming too complex to manage; that spurred the development of what became known as good programming practices, such as continuous testing (not just when everything was finished), abstraction layers, and functional programming with clearly defined inputs and outputs.

Another fun tidbit from the speech: “Nowadays one often encounters the opinion that in the sixties programming has been an overpaid profession, and that in the coming years programmer salaries may be expected to go down.” There was an economic logic for this: software and hardware costs had been roughly equal until that point, and hardware costs were projected to fall by a factor of 10, which seemingly meant software costs would have to fall too. That did happen, but only because software became more than 10 times as efficient—and also the amount of software demanded went up even more than that—so instead of falling salaries, we got rising salaries and rising employment.

8 | Regulation is hard and has unintended consequences: Australia banned “social media” for people under 16, but that includes Substack, which essentially fills the role that newspapers and magazines used to fill (although the platform also has social features). I say this is an “unintended consequence” not because the bill was drafted poorly—there’s apparently an explicit set of exemptions and Substack isn’t included—but because the people pushing for social media regulation probably didn’t have this outcome in mind. A lot of the writers I read at 15 would be on Substack today, and I probably wouldn’t be writing now without that experience.

9 | I liked this post framing a lot of what we do for toddlers (like giving them an ice pack they don’t really need) as a “totem”: something that helps them move on from a problem not by actually solving it, but by giving them a tangible symbol to recognize they’ve been helped and they can move on.

10 | After initially refusing to review Moderna’s totally normal mRNA flu vaccine, the FDA eventually reversed course, but Derek Lowe says the damage has been done: new drugs or vaccines require tons of spending before getting to the approval stage, and it’ll be hard to justify investing in anything that might be blocked on a whim. (Moderna has already said it’s cutting late-stage vaccine trials.) From a purely consequentialist perspective, this might be the worst thing the Trump administration has done; the PEPFAR cuts prevented lots of lives from being saved, but we’ll never know the impact of the medical advances that never made it to market.

Related: it’s hard for liberals to do anything about this trend, because even though skepticism has risen, vaccines are still way more popular than the Democratic party—so anything that makes it more of a partisan issue will probably backfire. The only real solution is to win elections and install better leaders.

11 | Pedestrian Observations points out a specific problem with “free bus” policies: they distort the demand for buses vs subways, making it hard to make both modes work together. In a well-designed system, buses generally go places that subways don’t (for example, NYC’s bus lines generally go crosstown where subways go up and downtown, or they cover subway deserts in the outer boroughs). But if buses are free, people will demand they take popular routes that are already covered by subways—as happened in DC in the 90s—making the system less cost-effective overall. If you want to subsidize transit, it’s more efficient to reduce fares across the board.

12 | As someone who, like Scott Alexander, read every Dilbert book by fifth grade (even the serious-ish “how to survive corporate America” ones), I thoroughly enjoyed his Scott Adams eulogy. In the last decade Adams became, uh, a complicated figure, and the eulogy analyzes all that in detail, but it was also nice to just flash back to funny cartoons that first exposed me to management.

Bonus thought-provoking comment on that post: “Boomers are tolerant of contradictions and find them amusing whereas millennials are horrified.” This arguably explains different attitudes toward careers (Boomers could work the same job for decades while also complaining about it constantly, whereas Millennials have to earnestly do something they find fulfilling), but also probably politics?

13 | This game-winning fire show went viral, becoming the most famous college basketball pyrotechnics since the 2016 national championship’s buzzer-beating confetti ... and apparently both celebrations were set off by the same guy?!

But Patrick McGee’s work is much more ambitious than the rest. Apple in China has a novel take that contradicts conventional wisdom on what both named parties get out of their relationship. It’s supported by strong reporting, which pulls no punches but is reasonably upfront about whose side it’s telling. It’s remarkably clear, laying out the thesis in five clear prologue pages. And it doesn’t hurt to be a little lucky, such as when America’s president laid massive tariffs on China just a couple months after publication.

The central idea: “Apple wouldn’t be Apple today without China ... and China wouldn’t be China today without Apple.”

• On Apple’s side: it’s well known that most Apple products are made in China, but conventional wisdom says that was just about lowering costs. In fact, Chinese manufacturers are uniquely able to adapt to new product requirements, scale up to meet massive demand, and respond to ruthless feedback on quality—allowing Apple to continually serve leading-edge products around the globe.

• On China’s side: Apple obviously provides jobs for Chinese workers and products that its relatively affluent citizens enjoy. But the greatest benefit, McGee argues, is that Chinese manufacturers learned a lot by working with Apple that they wouldn’t have learned from other companies. This has made China the unquestioned leader in making tech products, allowing it to dominate other companies’ supply chains and increasingly develop its own consumer brands.

My only complaint about Apple in China is its subtitle: The capture of the world’s greatest company. This makes it sound like China was forward-thinking and cunning, setting a trap that Apple myopically stumbled into. The book’s reporting shows there was no such master plan: Apple itself practically had to say, “Hey China, you have us captured and it’s really good for you, please don’t force us to leave.”

And although Apple’s dependency is a problem worth highlighting at book length, it’s not obvious that it—or even America—would have been better off doing things differently.

How Apple got here

Like many leading companies in complex industries like hardware, Apple doesn’t actually make its final products—it designs and sells them, but it works with third parties to do the manufacturing. So it needs to find reliable partners, ensure they have the right resources, teach them what to do, and test the final product quality. This is called supply chain management, and Apple is considered one of the world’s best at it.

That’s in part because Apple has mastered the generic MBA syllabus, such as: have multiple suppliers for everything, so you can play them against each other;2 make products only when you know you’re able to sell them, so you’re never stuck holding too much inventory; and use your power as a big and popular company to negotiate incredibly strict terms with vendors, making sure nobody can undercut you on price. (Apple once made a small company sign a contract without reading it, saying there wasn’t enough time.)

But what really makes Apple unique is how closely it works with suppliers to get exactly what it wants. All outsourcers send some oversight to the plants making their products (Tim Cook already worked with Chinese manufacturers while working at other computer companies before he was hired by Apple), but they generally expect partners to work independently to make a specified design. Because it sells high-end products, Apple can afford to invest lots more in helping its suppliers make better quality goods:

• Apple sends a lot more employees overseas, who work a lot more closely with factory workers. For instance, when it realized that only one company (Lens Technology) could correctly shape glass for iPhones, it sent people to that companies’ rivals and trained them on the process, giving Apple more options and reducing Lens’ leverage.

• When some manufacturers couldn’t afford enough equipment to meet Apple’s growing demands, it bought machinery itself and installed it at their plants. (This also helped Apple negotiate with them harder, because if the supplier didn’t keep its Apple contract then it would lose the equipment.)

• Apple was unusually fanatical about testing its products—more than half its assembly lines were dedicated to testing, which was unheard of at the time, and it often tests every single product instead of just a representative sample.

In addition to quality, Apple’s supply chain is uniquely flexible, because it chooses suppliers who are willing to work collaboratively and pivot rapidly. This dates to the iMac days: Apple made the computer in five different colors and produced them on-demand without having to guess which colors would be more popular in advance (this way it wasn’t on the hook for losses when Tangerine flopped). And it’s become even more prominent for iPhones: Steve Jobs demanded that the screen be switched from plastic to glass only six weeks before the first version launched, and Apple’s suppliers delivered.

To find partners willing to give that much flexibility, Apple basically did everything in China—about 90% of its products are manufactured there—because:

• Politically favored businesses could scale up quickly. Foxconn built a new iPod Nano factory from scratch in only nine months; later it used ample government infrastructure funding to double iPhone capacity on its own before Apple was willing to pay for it, becoming first in line when the demand came.

• The business landscape was ripe for shaping. Since China was rapidly industrializing in the early 21st century, Apple could to set up clusters of suppliers in the same city, making it easier to change specifications quickly.

• And with few labor rights protections, workers had no choice but to adapt to new demands. Per Haunted Empire, when Apple switched from plastic to glass screens, workers immediately started a 12-hour shift in the middle of the night “after being given just a cup of tea and a biscuit.”

Focusing exclusively on China wasn’t a deliberate strategy at first; it happened more gradually out of convenience. (Apple didn’t have a VP-level executive in China until 2017, and other senior leaders arrived only a few years earlier.) But it’s now stuck: its business model requires the scale and flexibility that only Chinese companies can provide today, and although it’s trying to move some production to other countries, it can’t anger China by doing so too quickly. At the same time, it has to manage tariff threats from a US government that’s concerned, for better or worse, about a massive trade deficit with China.3

Apple’s plight is an example of a phenomenon called Highly Optimized Tolerance: when systems are optimized against everyday risks, they may become more fragile to large, unplanned risks. Apple relentlessly mitigates the risks that suppliers will fail, input prices will rise, products will be faulty, and even that customer needs will change. But now it’s exposed to the big, weird risk of geopolitics.

How China got here

Attracting all of Apple’s production was a coup for China, the book argues, but not for the reasons you may think. It created a lot of jobs, which is a plus, but those are hard jobs that workers often quit. It didn’t make Chinese factories wildly profitable, because Apple negotiates so aggressively that nothing is left over for suppliers (even though Foxconn’s revenue doubled in five years after it started making iPhones, its profits were basically flat).

Instead, what China gained was tacit knowledge—what experienced engineers know about how to create technologically advanced products well at scale. It’s hard to point to examples of tacit knowledge, but you can see its effects: the same companies that served Apple started winning contracts for other suppliers like Samsung and Nokia. Eventually, homegrown Chinese companies like Huawei took advantage of their local expertise to create their own products, removing foreign partners from the mix entirely (not Apple, but lower-end designers).

Believe it or not, an American tech reporter found it easier to get information from former Apple employees than from CCP members, so McGee doesn’t have nearly as much reporting on China’s perspective. But what he has demonstrates that the arrangement was hardly a master plan on China’s part. For one thing, the most effective operational model—rotating workers frequently from Apple projects (to get the training) to other companies (to put that to more profitable use elsewhere)—was pioneered by Foxconn, which produces a lot in China but is headquartered in Taiwan.

For another, Chinese policymakers didn’t seem to really understand what they were getting out of the relationship. Their mental model was that technical knowledge was acquired through “joint ventures”, in which an experienced foreign company works with a local partner to sell products in China. (In theory, the foreign company had contracts in place to protect its intellectual property, but in practice, you know...) Apple wasn’t in an industry where it had to follow this model, and the fiercely independent company wouldn’t have abided it anyway, preferring supplier relationships that could be cut off anytime.

By the mid-2010s, Apple’s full presence in China was being felt—not just as a producer, but as a seller of iPhones to the country’s rising middle class—and the government fired a few warning shots: it started enforcing arcane length-of-stay laws against Apple workers, nitpicked its lack of local partnerships, and claimed it owed taxes for underreporting sales. It was Apple that had to go to government officials and show them its impact: arguing it was “investing” more than $50 billion per year in the country, because that’s what it was spending on Chinese suppliers, which involved teaching skills that were helping them dominate global high-tech manufacturing. (Of course, Apple kept this story private, not wanting to broadcast to the US how much it was doing for a geopolitical rival.)

With some other signs of good faith—an unusually large investment in Chinese ridesharing company Didi, establishing R&D centers in the country that may or may not have been useful—Apple eventually convinced China’s leadership that its form of technology transfer was even better than the old joint venture model. China loosened its JV requirement for automakers in 2019, allowing Tesla to build its own factory there; the country was already a leader in electric vehicle production, but it’s since zoomed way beyond the rest of the world.

Was it worth it?

For Apple: The book doesn’t say this explicitly, but the subtext and cover clearly point to no: Apple stumbled into a situation where it has to do whatever China says and can’t really move production out of the country, which means trade barriers or geopolitical tensions could ruin its business at any moment.

But that minimizes just how much Apple has benefitted from the arrangement. Having practically its entire supply chain in China has allowed the company to build complex products at unfathomable scale: per The One Device, each iPhone took 24 hours of labor to produce (as of 2012), and Apple sells more than 200 million of them every year. It’s hard to know the counterfactual, but it’s possible that without going all-in on China, it’s not just that iPhones would be more expensive—we might not have them at all, and certainly they wouldn’t be as good or upgraded as frequently.

So the benefit was massive. Was the cost? This comes down to two questions:

• 1) How much leverage does Apple have over the governments that could give it trouble?

  • In China: Apple is vulnerable because all of its production is there, but it’s also well aware of the benefits it brings the country—not only the skill transfer, but the iPhones and other products that hundreds of millions of Chinese citizens love. China’s government has extracted some small concessions from Apple (giving a local partner access to data on Chinese users, removing VPN and certain news and messaging apps), but nothing that seriously hurt its business. Apple’s influence is by no means absolute—it wasn’t able to prevent controversial Covid-Zero lockdowns from disrupting iPhone production in late 2022—but so far it’s been able to prevent any serious repercussions.

  • In the U.S.: The biggest risk has been trade barriers—the Trump administration put a 25% tariff on imports from China in 2019, and then threatened a 145% levy on some goods in 2025. But Apple ultimately got an exemption each time. That’s in part because it’s become savvy at politics (McGee writes that Cook called Trump every few weeks during his first term to stay in good graces), but also because America needs Apple to succeed—everyone loves iPhones, and Apple is a major contributor to the health of the stock market and Silicon Valley.

• 2) To the extent that some existential risk remains, is it okay for Apple to bet the company on it?

  • Companies fail all the time. Usually it’s the small stuff that kills them: they’re not profitable enough, or another competitor beats them to the next big thing. Apple’s model has minimized those risks, and generated tons of value, at the cost of exposing itself to a very big tail risk. Is that a bad thing? Traditional corporate governance theory would say no; if anything, executives generally take too few risks. But that’s easy to say about a normal-sized public company; it feels harder to accept when the risk threatens a product as ubiquitous as the iPhone.

For China: By being a good place for Apple to do business, China is now a chokepoint of massive and strategically important manufacturing industries. That’s obviously good, right?

To Western readers, it sure seems so. But there are complications that are harder for us to see. For one thing, it’s been less obviously good for Chinese citizens: the government’s crackdowns on labor protections led to poor working conditions and lower wages, and its relentless investment in business infrastructure has meant more inequality and a weak social safety net. For another, serving Apple has meant making concessions toward becoming more Western and capitalist: however much influence you think the Chinese government has over Apple, it’s nowhere near the control it has over local producers or JV partners. That doesn’t seem like a cost to us, but to Chinese leaders it probably is; I’d love to read a version of this book from their perspective.

Finally, it’s just hard to predict what will shape geopolitical power. In the mid-2010s—when the important-in-retrospect Apple investment was taking place—all of the financial press was instead writing about China’s seemingly transformative Belt and Road program, which the country has now basically given up on.

For America: Even if you think the bargain continues to be good for Apple and China, there’s another interested party: the U.S. as a whole. McGee doesn’t actually talk about this angle, but it’s the real reason his book resonated so widely.

Apple helping improve Chinese electronics manufacturing is on its face a good thing for America; we get better partners to make us better and cheaper stuff. And though it helped China become a chokepoint for consumer electronics, it’s not an existential crisis if a rift interrupted our supply of iPhones. However, if you believe that “everything is computer“ now, China might be able to leverage dominance of consumer electronics into dominance of manufacturing basically everything. We’re already seeing this happen to some extent in cars—as more vehicles get electrified, China is gaining share through its expertise in battery manufacturing. Could energy infrastructure be next (via solar panels)? Industrial machinery (autonomous robots)? Military hardware (drones)? If China becomes a chokepoint in those industries, America’s power could be truly threatened.

That isn’t an immediate cost for Apple: as Lenin famously (and perhaps apocryphally) said, “The capitalists will sell us the rope with which we will hang them.” But it would be bad for America, and the effects would be bad for Apple as well.

That’s a worst-case scenario; China might not dominate all those industries, and even if so it might not use that status to threaten American power. It’s become unfashionable to say that economic integration causes peace, and it’s certainly not that simple, but it’s not wrong either. There’s a future in which both countries double down on their comparative advantage—China on manufacturing, America on innovation and design—and we’re all better off for it. (And in the meantime, having awesome and relatively cheap iPhones is pretty great too.)

• Now Is Not the Time to Panic: The main timeline tells a story from when the narrator was a teenager, with flash-forwards to her being confronted about it by a journalist as an adult.

• The Brief Wondrous Life of Oscar Wao: Stories by or about Oscar and other family members that move back and forth across generations.

• The Night Circus: Each chapter opens with a specific date, helping the reader track a beginning, middle, and end of the story as they’re told out of order.

• Americanah: Two stories of immigrants settling in America and the UK interleaved with substantial chapters flashing back to their time together in Nigeria.

• A Visit from the Goon Squad: A book of short stories that jump around different times in the same characters’ lives.

• The Overstory: Several characters’ stories are told in alternating chapters until they all come together. (Admittedly this one is more about different strands in space than in time, but it has a similar effect.)

When I first wrote about multi-track stories, I hypothesized that they’re popular because all the linear stories have already been told: playing with time increases possibilities exponentially, allowing an author to create a new shape. But that relied on my anecdotal observation that multi-track stories are a recent invention. Is that true?

I can now answer that question without reading hundreds of books myself—just ask large language models to research a representative sample of novels and quantify which ones jump around in time.

I gave ChatGPT, Claude, and Gemini two lists: the New York Times’ 100 Best Books of the 21st Century (removing 30 non-fiction books) and Penguin’s 100 Must-Read Classic Books. I gave each the same prompt explaining what I meant by multi-track novels, asked them to classify each book from both lists, and followed up with clarifications to my criteria if they were wrong about any of the books I’d read.1

According to their analysis, multi-track novels are more common today, but not dramatically so:

• ChatGPT says 14% of the modern list fits, and 7% of the 20th century list

• Claude (24% vs 19%) and Gemini (26% vs 15%) are looser on the classification overall, but the relative change is in the same ballpark

The full book-by-book results are here (tell me if you see anything that looks wrong). I expected the multi-track novels to be much rarer historically, and perhaps a bit more common in modern times.

Previously: 2024, 2023, 2022, 2021, 2020, 2019

1 | About two-thirds of gendered animals in children’s stories are male. Frogs, wolves, and foxes are almost always male; birds, ducks and cats are the only common animals that are more often female.1 [The Pudding]

2 | Netflix commonly tells writers to “have this character announce what they’re doing” for people who have the show on in the background and aren’t really following. [Will Talvin via Derek Thompson]

3 | ChatGPT would need to answer at least 10 million questions to emit as much carbon as one cross-country flight. (If you’re concerned about water depletion instead: 200 responses use as much water as growing one almond.) [@AaronEstel, The Weird Turn Pro, more in this follow-up]

4 | Medieval Englanders stored coins in containers made from a clay called pygg. The vessels were called “pygg jars,” which became the “piggy banks” we know today. [Liberty Street Economics]

5 | Chicago pizza had been around for decades before it became known as “deep dish,” at which point it wasn’t that much different from other regional pizzas. But once it had the name, Chicago pizzerias leaned into it, making it “deeper and dishier” over time. [United States of Pizza]

6 | Fly balls travel a foot farther in Miami than in Seattle because gravity is stronger in the north. (Earth is not a perfect sphere; it bulges out a little at the equator.) [Effectively Wild]

7 | “Food comas” seem to be real: people who had eaten in the past hour scored 8% lower on a surprise test than people who hadn’t, with a larger effect on harder tests. (This held no matter when the test was given, so it’s not just a circadian-rhythm effect.) [Econstor via Gastropod]

8 | 17-year-old Birgit Felden won the 1984 Nathan’s Hot Dog Eating Contest for eating 9.5 dogs in 10 minutes. (For comparison, Joey Chestnut won this year’s with 70.5.) She was visiting New York on a nine-day trip from Germany with her judo club and had never eaten a hot dog before. [The Athletic via Shutdown Fullcast]

9 | Ikea serves food to 700 million customers per year; some say it’s the world’s sixth most popular chain restaurant.2 [Acquired podcast]

10 | Prediction markets for elections have shown a ‘momentum’ effect: if the price went up today, it’s more likely to go up over the next couple days, which also better predicts the final outcome. [No Dumb Ideas]

11 | Loud restaurants aren’t just annoying, they’re more dangerous, because shouting while eating spreads airborne diseases. [Complex Systems]

12 | More than 50% of Americans approve of using spanking to discipline children, even in the 18-29 and 30-44 age groups.3 [Family Stuff, see underlying GSS data here with some more details/crosstabs in this thread]

13 | The 1970 Vietnam draft lottery wasn’t random—people born in later months were disproportionately likely to come up earlier, because the balls weren’t mixed well before the drawing. [Misha Teplitskiy via Kent Hendricks]

14 | Rabbits rarely eat carrots; they prefer leafy greens. The association of rabbits with carrots was popularized by a 1940 Bugs Bunny cartoon. [@alexschmidtstagram]

15 | New York City’s hiring rules are absurdly bureaucratic: you have to choose an existing job description (or go through a long process to get a new one approved), find people who already took a test to make them eligible for that role without knowing which department might hire them, and then choose one of the three highest scorers on that test. (In practice, managers often look for weird titles that haven’t been tested so they can do their own interviews.) [Eating Policy]

16 | The “Taylor Rule”—a highly influential formula that observers often use as a guide for what interest rates should be (though the Federal Reserve doesn’t actually follow it)—comes from a paper that used only six years of data (1987-92).4 [Odd Lots]

17 | According to one study, attractive waitresses earn significantly more in tips, mainly because they get bigger tips from other women. [@MackenMurphy]

18 | Eric Adams was sworn in as NYC’s 110th mayor, but his successor Zohran Mamdani will be #112—a historian found that the city’s count missed Matthias Nicolls’ second (nonconsecutive) term in 1674, which the Department of Records has accepted. [Original finding reported by Gothamist, follow-up in archives.nyc with some more complications]

19 | In 1970, about 50% of US manufacturing was in the Rust Belt and 25% in the South; today those shares have flipped. [WaPo via Marginal Revolution]

20 | The reported total fertility rate has fallen in the US since 2000, but that’s just because women are having children later; 44-year-olds today have as many children as 44-year-olds in 2000. (That doesn’t mean everything is fine; even that corrected fertility rate is a little below replacement level, and the problem is worse/accelerating in many other countries.) [Mike Conczal]

21 | Mars (the company that makes M&Ms, Snickers, Skittles, etc.) owns more than 5% of US veterinarian practices and makes more revenue from pet care than from candy. [Acquired podcast]

22 | When one gram of ice melts, it releases 100 billion Terabytes of entropy. (Entropy in information theory represents how complex a message is—for example, adding extra characters to a password increases entropy—which is measured by how many bits it takes to encode the message, as in computer memory. Entropy in the physical world can also be framed the same way: you know exactly how molecules are arranged in solid form, but you don’t know where they are as liquid because they move around, so describing the matter requires more information.) [Simon DeDeo and relevant SMBC]

23 | According to the founder of Trader Joe’s, Thanksgiving is one of the best days to sell wine. Christmas is a bad day for wine but a good day for hard liquor. [Becoming Trader Joe]

24 | Mark Zuckerberg wanted to change his company’s name as early as 2018 (because Instagram and its other properties were growing while the “Facebook” app was declining), but he didn’t want to be seen as running away from bad PR at the time, so he was searching for a defensible reason. This probably explains the somewhat ridiculous change to ‘Meta’ a couple years later. [Facebook internal emails from FTC trial, they’ve since been taken down but the idea is mentioned here]

25 | Sidewalk chalk is no longer made of what chemists call chalk (calcium carbonate); it’s made of calcium sulfate instead. Blackboard chalk is usually chalk, but the chalk that gymnasts and LeBron use to improve their grip isn’t; that’s magnesium carbonate. [Wikipedia via xkcd]

26 | In 2005 Snapple tried to set a record with a 25-foot, 17-ton popsicle in NYC’s Union Square—but the frozen treat melted in the truck that was carrying it, spreading strawberry-kiwi goo across several streets. (”What was unsettling was that the fluid just kept coming,” a Guinness Book of World Records representative said.) [NBC News via ChatGPT]

27 | Lyndon Johnson and his wife owned a TV station in Austin while he was president and benefited—probably disproportionately—from advertiser payments. (He got a CBS affiliation after he was elected to Congress; it didn’t make business sense on the merits, but a senior network executive decided the political connection made it worthwhile. After Johnson became president, he frequently called CBS leadership to complain about their coverage.) [Charlie Brown’s Christmas Miracle]

28 | “Kermit”, “Grover” and “Elmo” were top-300 American boys’ names in the first half of the 20th century. (All fell out of the top 1000 by the 1970s or earlier; Sesame Street began in 1969.) [SSA via @whovian_delights]

29 | An often-cited stat claims that the median first-time homebuyer’s age has risen from 31 to 40 in the last decade. But that’s based on a single industry survey (with a low response rate) that’s probably wrong; government surveys and financial data show a younger median and little change in recent years. [Connor O’Brien, Lance Lambert]

30 | 2,000 calories of food (a standard daily recommendation) provide the same amount of energy used in driving a traditional car two miles, or an electric car 20 miles. [Energy Cheat Sheet]

31 | Intel became famous for powering Windows PCs in the 1980s. Apple probably would have used Intel chips in its first PC even earlier, but Steve Jobs lied to cofounder Steve Wozniak about how much they would get paid per computer (to keep a larger share for himself), so Wozniak chose a cheaper competitor to stay within budget. [The Intel Trinity]

32 | It’s generally illegal to use native American bird species in movies. Film animals are considered working pets, and early-20th-century avian protection laws banned domestication of native birds. [Slate]

33 | Placebo effects can affect social science studies, not just medicine: people who were told they were assigned to the control group of an experiment had worse outcomes than people who were in the same group but weren’t told about it. [Marginal Revolution]

34 | Although there are fewer living veterans today than in 2001, the number of disabilities compensated by the VA has grown from six million to 41 million. [@evanhill]

35 | Finland has more than one sauna for every two people. [NYT]

36 | Babies are gaining some gross motor skills like rolling and crawling later, because they’re developed on your tummy, and parents are now instructed to have babies sleep on their backs to prevent SIDS. [What to Expect the First Year]

37 | “Black raspberry” flavor comes from a plant that’s distinct from red raspberries or blackberries; it’s used in ice cream because red raspberries decay and lose their color when mixed with dairy. (”Blue raspberry” is an artificial flavor representing regular raspberries; food companies used blue dye for it because cherry and strawberry were already red.) [several sources, see this ChatGPT thread]

38 | Post-tax incomes rose at about the same rate for all economic classes in the US over the last 50 years. (The top 1-5% had higher pre-tax growth, but policy has become more redistributive.) [SSRN]

39 | “Big spoon” and “little spoon” are nonsensical terms for cuddling, because spoons only fit next to each other neatly if they’re the same size. [xkcd]

40 | The number of Americans in prison has declined by 25% since 2009 and is expected to fall another 50% over the next decade. This is because crime rates have gone down since the 90s, but many people convicted in that era are still incarcerated on sentences that will end soon. [The Atlantic]

41 | Denver International Airport is larger than the entire city of San Francisco or Boston. [@mehran__jalali]

42 | According to one estimate, a single pitch being a ball or strike is worth about $80k to a baseball team on average.5 [Effectively Wild]

43 | When the USPS first accepted large packages, at least eight families mailed their children to a relative or neighbor (including one all the way from Florida to Virginia). [History TV]

44 | Private equity firms often seek out small local companies that don’t have a website, make a website so more customers can find them, and benefit from the easy growth. So some business owners are now taking down their websites to fool lazy PE buyers into buying them. [Don’t Worry About the Vase]

45 | Sans-serif fonts have taken over because they render better on small phone screens. [Sharp Tech]

46 | Chromosomes only form the classic “X” shape when they’re dividing; they’re usually more tangled. [How Life Works]

47 | How do camera crews keep their pictures clear in rain or snow? They put a fast-spinning disk above the lens, which isn’t noticeable but shakes off precipitation. [vt.physics]

48 | When home improvement companies (or other big-ticket purchase sellers) offer something like “12 months to pay back without penalty and then a high rate”, it’s not really a trick to try to get you to pay more—they’re just trying to get you to lock in a sale as soon as possible, even if you need time to complete your own financing later. [Patrick McKenzie]

49 | In 2018 a golfer played the entire country of Mongolia as one hole, taking 20,093 shots to cover 2.2 million yards. His caddy carried 400 balls for the 80-day trip. (A random dog joined them on day two and stayed with them the entire rest of the way.) [The Longest Hole via Shutdown Fullcast]

50 | Public opinion swings against the views of the president’s party even during successful terms. (For example, Americans became much more liberal throughout the Reagan era.) [James Stimson via @SpecialPuppy1]

51 | “Hot chocolate” and “hot cocoa” are technically different drinks—the former is made with real chocolate while the latter is made with cocoa powder (which is sweeter but less rich since it doesn’t have most of the fats). [@halfbatchbaking]

52 | The length of written and spoken sentences has been falling for centuries. There isn’t an obvious explanation, but one theory is that shorter sentences are clearer and people have just become better at communicating. [Arjun Panickssery]

Roughly ranked within each section.

Business and economics

1. Chokepoints: American Power in the Age of Economic Warfare (2025): An outstanding nonfiction book. It’s a really important topic: how the US (and to a lesser extent other countries) use control of financial systems and technology to enact sanctions on perceived bad actors. There’s lots of detail, but it’s actually interesting, and it’s focused on the topic at hand.

2. The Nvidia Way (2024): A solid “management book” about one of the most successful and uniquely run companies. Keep in mind that it’s very much NVIDIA’s (and really CEO Jensen Huang’s) perspective—controversial issues like crypto mining, international sales, and how to choose customers aren’t mentioned at all. See a full review here.

3. The Intel Trinity (2014): I was planning to skim this quickly to look for any tidbits to use for the NVIDIA book review, but it was compelling enough that I ended up reading the whole thing in earnest.

4. Trade Wars Are Class Wars (2020): This had been on my list for a while but became more relevant this year. The argument is mainly a) our trade deficit is mainly driven by other countries wanting to invest in the US and sending us stuff to get dollars, not greedy Americans (this is actually pretty mainstream in economics) and b) those financial demands are driven by inequality in those other countries, especially China and Germany (this is a heterodox take but not a zany one, and they argue it clearly).

5. The Box: How the Shipping Container Made the World Smaller and the World Economy Bigger (2006): I was slightly disappointed relative to how much modern tech/finance types talk about this book—it’s great on the invention and adoption of containers, but less than I expected on how they actually changed the economy.

6. Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist (2025): Annoyingly ideological but also packed with useful facts about Spotify’s success. The most important theme is its evolution from being like iTunes (a place to find music you already want to hear) to a service for filling your life with background music—explaining algorithmic playlists, the rise of bland coffee-shop music, and white-label artists. (A Daniel Ek quote: “Apple Music and Amazon aren’t our competitors. Our only competitor is silence.”)

Other nonfiction

1. The Boys in the Boat (2013): I resisted reading this for a while because it sounded cliche, and it was, but the writing lived up to the hype: amazing detail about every race 80 years ago and some neat insight about rowing. It was interesting to learn the “underdog story” everyone talks about was only sort of true—some of the individual rowers were unlikely heroes, but not the Washington rowing program as a whole (it had won a national title already and was a close finisher in others, plus its West Coast compatriot Cal had won the prior two Olympic golds).

2. Rickey: The Life and Legend of an American Original (2022): Rickey Henderson is a worthy subject of a biography—perhaps the most underrated of the top 25ish players of all time?—and I appreciated the vast range of perspectives in here. The most interesting takeaway is that Rickey sat out for injuries when he probably didn’t have to, which was looked down upon at the time, but is more in line with how teams treat players today and may have contributed to his longevity and success when he did play.

3. Excellence Without a Soul (2005): I found this from college while cleaning my office and it had much more relevant commentary than I expected on grade inflation, distribution requirements, and athletics. It’s very “academic” in the sense that it’s heavy on facts and light on clear positions, though. (Except for how much Lewis hated then-Harvard president Larry Summers: “Lawrence Summers’s principal failing was not that he was too strong or too uncongenial, but that the wisdom, knowledge, and judgment he lent to faculty affairs were too feeble.”)

4. The Last Great Walk: The True Story of a 1909 Walk from New York to San Francisco… (2014): A cute light read, though I could have done with fewer/shorter digressions into human nature and more focus on the walk itself. The subtitle doesn’t give away the astounding fact that Weston was 70 years old when he attempted his walk (nor the explanation that he was a professional distance walker for most of his life, which makes it a bit less astounding).

5. How Life Works: A User’s Guide to the New Biology (2023): I don’t know much about biology so this was exhausting; I only made it through about six chapters and that was with a heavy amount of ChatGPT help. It seemed important and the first few topics related to things I kind of knew in interesting ways (how genetics are more complicated than “this gene creates this protein / causes this trait”; why protein folding matters) but then it became a blur.

6. The Scaling Era: An Oral History of AI (2025): Recommending because I liked the first section on the theory and history of scaling. Otherwise there are lots of insights throughout but I don’t know who this book is actually for: you have to be smart enough about the basics of AI to follow its lack of structure and have enough interest in the topic to want to dive into the details, but not so into the topic that you’re already listening to Dwarkesh’s podcast or reading other things from the people interviewed here.

7. Twitter and Tear Gas (2017): A thesis that seems important and still relevant: the Internet made it easy for protest movements to arise spontaneously, but that meant they didn’t have (formal or informal) leadership structures, so they don’t last very long and/or can’t articulate clear demands. It’s very set-in-the-early-2010s so it’s a little hard to read now, and I think it partially but didn’t completely predict how future digitally native protests like MeToo and BLM played out.

Fiction

1. Cutting for Stone (2009): I enjoyed all the medical details, which taught me about a field I don’t really know very well, and as a pure story it was also beautiful.

2. The Night Circus (2011): This was the first (only?) audiobook I read that I actually looked forward to, and that didn’t just feel like a worse version of reading a regular book. The story was really compelling and the voice acting helped it work really well in audio too.

3. The Secret History (1992): I devoured this, in part because I was sick and had nothing better to do. The first 100 pages were a little cringey, but I still wanted to see what would happen, and then the rest was actually enjoyable.

4. City of Thieves (2008): Really engaging, chapters flew by, and the ending was disappointingly abrupt.

5. Circe (2018): I enjoyed it a lot despite (because of?) not knowing much about Greek mythology—the characters were interesting and most of the plot wasn’t too predictable nor forced. The voiceover was nice too. (audiobook)

6. His Dark Materials trilogy (2000): Not my usual genre but I really liked the pace of the first volume and wanted to keep reading. It dragged a little more through the second and the first half of the third.

7. A Visit from the Goon Squad (2010): I don’t always like this type of book but I highly enjoyed this one—favorite stories were #4, #7, #12.

8. The Curious Incident of the Dog in the Night-Time (2003): A quick read with a character whose head it felt worth getting inside. I switched from audiobook to hard copy midway through and the latter was much better.

9. Bleeding Edge (2013): Only recommended if you’re into tech and/or finance, but I found the setting and the gratuitous dot-com-era references fun. I also wish I’d read the print book for this one because the narrator’s voice took a long time to get used to. (audiobook)

10. The Song of Achilles (2011): I liked the first two-thirds a lot but then it got too “unrealistic,” which sounds silly to say for a mythological fantasy, but the climax was hard to believe even in the context of the rest of the story. (audiobook)

Parenting

1. Siblings Without Rivalry (1972): The same advice from the How to Talk to Kids... series but focused on sibling situations with more examples. The most novel parts are the ones about a) don’t compare your two kids or put them in roles (”you’re the responsible one”) and b) although you of course shouldn’t favor one sibling generally, you also shouldn’t try to make everything “equal” (because circumstances will be different for each child).

2. Oh Crap! Potty Training… (2015): We didn’t really try to stick with this plan nor did we find what it said happened exactly in practice, but this book did one really important thing, which was lower my expectations for how quickly potty training would work.

Children’s books (2-3 years old)

1. If Animals Went to School

2. The Science Girls

3. I Hereby Crown You Big Sister

4. Paddington Plays On

5. Elephant & Piggie: Should I Share My Ice Cream?

6. Richard Scarry’s Cars and Trucks from A to Z

Previously: 2024, 2023, more books throughout the year on Goodreads

1 | Charles Darwin’s 10-year-old son doodled the above pictures on the back of Darwin’s original manuscript of On The Origin of Species.

2 | Adults should correct other people’s kids (and welcome the same for their own). The idea is that children should learn to respect others in society—you should expect negative feedback if you aren’t behaving well, and you should be receptive even if it’s not coming from mom or dad. “Modern parents agonize over how to raise community-minded humans and then go out of their way to model for their kids that the opinions and concerns of others don’t really matter.”

3 | Decoder Ring tracked down several creators of Charlie’s Angels to ask a burning question: How did a scene show one bird species, name it another, play the sound of a third, and hinge on a fact about it that’s totally untrue? You can more or less guess, but one surprise is that it’s illegal to use native birds in American movies.

4 | I wrote last time about a study showing that AI hasn’t increased software developers’ productivity much. Shortly after, a much more rigorously controlled experiment came out, showing … AI tools slowed programmers down by about 20%! This is despite the fact that the subjects expected it to make them 20% faster, even looking back after the experiment was over.

Why? The best explanation seems to be:

• Most of the subjects weren’t very familiar with Cursor, the AI tool in question, although both this fact and how much it matters are in doubt.1

• Working in large, complicated codebases is hard for AI relative to creating new projects or changing isolated elements like user interfaces.

• The subjects were very experienced developers, especially with the specific open-source projects used in the experiment. (Studies of less experienced software engineers tend to show a significant benefit from AI tools.)

But those explanations don’t really undermine what this study tells us about AI’s impact on our world. Yes, it’s genuinely transformative that large language models let non-experts write more or better code. But it’s still the case that most of the global value of software is created by experienced engineers working in established codebases. I think you can generalize that to other sectors, which is why I’m a little pessimistic on AI automating tons of jobs soon.

A second angle from this study is how the developers spent their time:

Of the 20% time increase, about half of that was spent either “waiting on AI” or “idle”. (The rest was spent reviewing AI output or prompting AI, which together took more time than was saved on coding or researching.) This happens because AI agents take time to edit and test your code. You could spend that time working on something else, but you might not have another exactly-three-minute task available, or it might not be worth context-switching.

From an economic or business perspective, that’s wasted time. But to you it may not be: scrolling social media or getting a coffee is fun! (Likewise, prompting AI is probably less effortful than active coding, which you might also prefer.) In this light, we can perhaps understand why the subjects were wrong about what happened: they spent less effort with AI, so they felt like they were getting things done faster, even though they weren’t.

Once AI actually does make us significantly faster at accomplishing things, will we be able to, and choose to, use the saved time productively, creating more growth and potential for shared prosperity? Or will we just slack off more, enjoying the extra leisure time? (And is it obvious which outcome we should want?)

5 | Relatedly—here are three facts: AI has exploded in recent years; AI is especially good at doing what entry-level white-collar workers do; and new college graduates are increasingly unemployed. It doesn’t take a billion-parameter neural network to deduce that AI is displacing the young workers. And yet, several recent takes (1, 2, 3) argue that’s probably not true. The main arguments:

• Hiring is down across the economy for other reasons (policy uncertainty, higher interest rates), and new workers are the most sensitive to such fluctuations.

• The gap between college and non-college grads is at a new low, but it’s been shrinking consistently for 15 years, not a new trend.

• The trend by sector isn’t really correlated with AI exposure outside of computer science (where hiring has been cool since 2022-23 in response to rate increases and a glut of workers who came in right before then—not necessarily AI—and hiring rates are starting to pick up again).

• Productivity, though hard to measure, doesn’t seem that much higher in the aggregate (as you’d expect if companies could just get by with fewer workers).

These still don’t fully contradict the AI story, but they make it less obviously true.

6 | A widely shared ESPN.com article several years ago included the cool fact that top chess players burn 6000 calories per day during competitions. (It made my 52-things list that year.) It turns out that was bogus; a professor got that number by reading a study about breathing rates, confusing the maximum reading for an average, and then assuming it translated linearly to caloric expenditure (it doesn’t).

7 | Margaret Wise Brown (of Goodnight Moon) was really particular about line breaks, timing them so children can chime in during the pause.

8 | What it looks like inside various musical instruments:

9 | Everything is tax policy: An office building in the UK put a bunch of snails on their vacant floors to claim an “agricultural facilities” tax credit, saving hundreds of thousands of dollars over the last few years.

10 | One of the few novel arguments I’ve seen on AI existential risk (the long original is here, though this follow-up is more tractable): Large language models don’t seem to have an intrinsic personality; they take on a role based on what they learn from training data. They’re trained on approximately all human writing, which now includes commentary about AI itself—of which a lot focuses on how to “align” models to do what we want instead of, say, going rogue and killing people to achieve some other goal. Human writing also includes a lot of literary fiction about heroes who overcome people trying to get them to stop in order to achieve their goals. Put those together and it sounds like we’re implicitly leading AI to take on a role that could turn out pretty badly!

11 | Prediction markets sound great: use the wisdom of crowds to get the best information. So people have tried to take them beyond big obvious things (“who will be the next President”) to more specific ones (“which of these candidates would make the best CEO for our company?”). One trick for this is called conditional prediction markets—people bid on contracts for something like, what will our stock price be next year if Alice/Bob/etc is CEO, and the one that settles at the highest price is the most promising.

Niche prediction markets have generally failed for reasons that are kind of boring (see some examples here). But this post brought up a fundamental flaw that I hadn’t considered before: conditional prediction markets only tell you about correlations, not causation. An example: you might run a conditional prediction market on “what’s the chance of nuclear war if Trump does or does not declare a no-fly-zone over Ukraine” to assess how risky it is. But that market will set a much higher price in the “does” scenario—not just because of the impact of the no-fly zone itself, but because in worlds where Trump does that, he’s also more likely to take other aggressive stances that make war more likely.

12 | Tweet of the month:

It might not be true! They define productivity as commits per user, which might understate the impact here—open-source development often isn't a full-time job, so maybe people are producing that output more quickly and working fewer hours. But it passes the smell test to me.

In the canonical software management book The Mythical Man-Month, Frederick Brooks gives this breakdown for how programmers actually spend their time. (Things have changed since he published that in 1975, but not in ways that really affect this analysis, so I'll relegate them to a footnote.1)

• 1/3 planning what to do

• 1/6 writing functional code

• 1/4 component testing - does your code work as you want in isolation

• 1/4 system testing - does your code work as you want in the context of your larger system

Component testing today includes writing tests in your code that will tell you if things go wrong, so let's count half of that as "writing code" too. That makes the total time spent coding 1/6 + 1/8 = about 30%. The AI tools available at the time of this study were almost purely focused on writing code, so they couldn't help with things that take 2/3 of a programmer's time.

Still, if 30% of time is spent writing code, and AI wrote 30% of that code instantaneously and perfectly, you'd get a 30%*30% = 9% increase in productivity. But it's definitely somewhat less than that:

• Code "written by AI" generally includes autocompletes (where you start writing something and it fills in the rest); that's far from a 100% speedup, since it only saves you the time actually typing out the commands, not the time thinking of what to start writing in the first place.

• Even code fully written by AI (through a chat interface or an autonomous agent) requires time to give it context, review the output, and iterate when it gets something wrong—which is especially common when you're working in an existing codebase, as most development involves, rather than starting from scratch.

• And "percent of functions written" (the paper's definition) isn't the same as "percent of work done"; in my experience AI is more likely to write new helper functions than to reuse or modify existing ones, which would make it appear to be producing more than it actually is. This is likely a small effect at most.

The paper's 2.4% improvement, versus the theoretical maximum of 9%, implies that even the AI-written code takes humans about 75% of the time they would have normally spent. That's perhaps a little bit lower than I'd expect, but not far out of the ballpark, especially for a time when AI coding assistants are brand-new and might take a learning curve to use.

But there's a big caveat: data from December 2024 is already practically obsolete, because AI coding models in particular have taken massive leaps forward in the first few months of 2025. The new reasoning models are now helpful in the "planning stage", helping translate rough ideas into instructions that a coding model can use. The core coding models are getting better, and agent systems are able to deliver larger chunks of their work without human intervention. I still wouldn't bet that any of these have radically transformed productivity yet, but they're pushing the envelope in more dimensions, and as tools keep improving and people become better at using them, the impact will accelerate.

So although I think productivity growth in the low single digits was correct at the end of 2024, and is still probably roughly correct today, I don't think it'll stay that way for too long.

1. Be Playful:

• Make it a game. “Can we get all the cars into the box before the timer beeps? Ready … set … go!”

• Make inanimate objects talk. “I’m an empty sock. I need a foot in me!”

• Use silly voices and accents. “I … am … your … robot … Must … buckle … seat … belt … now.”

• Pretend! “We need to climb this slippery mountain into the car seat.”

• Play the incompetent fool. “Oh dear, where does this sleeve go? Over your head? No? On the arm? This is so confusing! Thank you for helping me!”

2. Offer a Choice: “Do you want to hop to the tub like a bunny, or crawl to the tub like a crab?”

3. Put the Child in Charge: “Johnny, would you set the timer and let us know when it’s time to leave?”

4. Give Information: “Tissues go in the trash.”

5. Say It with a Word (or a Gesture): “Trash!”

6. Describe What You See: “I see most of the blocks put away in the toy box. There are only a few blocks left to go.”

7. Describe How You Feel: “I don’t like food thrown on the floor.”

8. Write a Note: “Put me on your head before riding. Love, your bike helmet.”

9. Take Action Without Insult: “I’m putting the paint away for now. I can’t let you splatter the other kids.”

Simple, right? That's nine different techniques—13 if you count the sub-bullets of #1—to have ready to choose from on a moment's notice, all while your child is probably shouting and you’re running late for school. I couldn’t imagine being able to do all that.

But when I re-read it a year later, I realized almost everything on the list had become second nature.2 That's not to say that they're always successful; just that I use them instinctively, and they often help. From what I've seen, most other parents have developed their own tricks that generally follow this advice, even if the specifics differ. (In fact, that’s an advantage—if my attempt at using a stuffed animal to help us put shoes on isn’t working, Meredith can come to the rescue with her own tactic of making the shoe talk with a funny voice instead.)

Another way to put it: parents develop expertise at communicating with young children. Expertise is largely a know-it-when-you-see-it thing, but here's a definition I find useful: 1) you have a lot of knowledge about how something works, and 2) you can act on that knowledge to achieve your goals much more effectively than a non-expert, which requires 3) a sophisticated mental model, usually trained through experience, that compresses your knowledge into something you can act on instinctively.3

• An expert jazz pianist knows the fundamentals of jazz music (what each chord is comprised of, what progressions they usually follow, which melodies sound interesting on top of them), can play complex sounds on the piano, and chooses what sound they want to play as they're performing.

• An expert basketball player knows the best way to attack different defensive schemes, can dribble and shoot really well, and makes decisions for when to dribble and shoot based on split-second cues from the defense.

• An expert at corralling a toddler knows different ways to help them focus, can communicate those strategies clearly, and chooses which one to use in the moment based on the mood, time pressure, which stuffed animals are nearby, etc. (Don’t think you’re not an expert just because it doesn’t always work—even the best basketball players miss half their shots too.)

People typically study expertise in fields like music, sports, and dance—where you practice a lot, get coaching on weak areas, and measure yourself against benchmarks so you know when you're getting better. But most things we do aren't like that, so I'm really interested in how expertise manifests in these other areas. Parenting is an example: there's no real way to practice beforehand (the closest I got was swaddling a stuffed pelican) and no time to practice once you’re doing it; you can only learn "on the job."

And yet most parents get much better over time at engaging with children, especially their own children. How does that happen? According to a framework that I love, there are four conditions that help us learn well:

• Repetition: practicing something over and over again gives you opportunities to improve.

  • A child has to get dressed, get out the door on time, eat dinner, take a bath, brush their teeth, put on pajamas, go to bed ... and do it all again the next day!

• Timely feedback: seeing the outcomes right away helps you learn what works well and what doesn’t.

  • You know right away if they're cooperating, distracted, or having a tantrum.

• Task variety: facing different challenges encourages you to experiment with new techniques and forces you to learn generalizable skills, not just specific situations.

  • You corral a child with different goals, in different locations, and while experiencing different moods.

• Progressive difficulty: as you improve, you need to face harder challenges to keep learning.

  • Older children learn more ways to get what they want even when they shouldn’t, and new considerations like potty training enter the mix over time.

You might notice that many aspects of parenting are not like this, and those can be really stressful. Is your preferred method of discipline helping your child grow up to be an independent decision-maker, or will they resent you for the rest of their lives? Don’t expect timely feedback; you’ll have to wait a decade or two to find out, and also it’ll be confounded by everything else they’ve experienced in their life, sorry!

But even for the everyday communication, it hasn’t really felt like leveling up: there are no exercises to master, no teacher giving out stickers, and (thanks to the progressive difficulty) I’m still failing a lot of the time. Remembering how I felt a year ago—rereading something from the past, or seeing parents of younger children—has helped me realize what I’ve learned and gain confidence for future challenges. This probably applies to any other unstructured domains too.

This is my best attempt to 1) explain the benefits of trade; 2) steelman the case for why to limit it anyway; and 3) show how the recent tariff plans wouldn't achieve that goal.

You might have heard that the tariffs were “paused,” but that’s not completely true: we still have the highest effective tariffs in nearly a century, with lasting effects on businesses and financial markets. And there’s always a chance that the original policy could return or something totally new could come next.

How America benefits from trade

In your Econ 101 class, you learned that trade happens because of comparative advantage. A stylized example: America can make airplanes or clothes; Indonesia can make clothes but not airplanes. So America makes airplanes, Indonesia makes clothes, and they trade: Indonesia gets airplanes it can't make itself, and Americans get cheaper clothes than they could make on their own. This is a major reason why stuff has gotten cheaper in the US, relative to expenses that can't be traded like housing and medical care.1

In the textbook example, the countries exchange goods for each other, but in reality there’s usually an imbalance: one side gets more stuff and the other gets money. Maybe Indonesia doesn't really need airplanes, but it needs oil—so the US instead sends its airplanes to Saudi Arabia, and Saudi Arabia sends oil to Indonesia. Maybe value is being exchanged outside of physical products: a lot more foreigners pay tuition to study in the US than vice versa, which is a US "export." Or maybe goods are crossing multiple borders: the full cost of an iPhone counts as a US import from China, even though 75% of its value comes from supplies made in other countries.

For all those reasons, the trade balance in goods between two countries can be misleading. But the hardest case to explain is when overall trade is still imbalanced: the US imports about $1 trillion more from the rest of the world than it exports. How does that work?

• Americans pay out more money for imported products than we receive for exports, which means foreign entities accumulate excess dollars they aren't immediately spending. They invest those dollars into US businesses or government bonds, or they deposit them into US banks who do the same thing. The US is borrowing from the rest of the world, and our debt rises.

  • People often frame this as “we borrow because we want to buy more stuff”, but you can also reverse the causality: if foreigners want to invest in the US more than Americans want to invest abroad, then they have to get that money from somewhere, which means selling us more stuff. In fact, economists tend to use this investor lens to explain trade imbalances more than the spending lens, although you can’t really isolate either—maybe it’s more desirable to invest in the US because we buy more stuff than other countries.2

• How do we eventually pay back the rest of the world for this borrowing? We can stop importing so much some day (or export more) to reverse the trade balance. Or, since we owe them dollars, we can just print more dollars. This would have a lot of bad effects (inflation, less future investment) but it would at least be within our control.

• But nothing says the deficit has to be paid back on any particular timeframe. As long as our economy grows faster than the interest we owe, we can keep borrowing more to pay back old debt sustainably. And as long as foreign investors want American assets more than we want foreign assets, the trade deficit will continue.

So Americans benefit from the pre-tariff landscape: we can afford more stuff at cheaper prices. We have to borrow more, but because everyone wants to invest in the US, we can do so very cheaply; there’s no indication that we’ll have to pay it back in the near term, and if that changes we have several options for doing so. (There are also “soft power” benefits: we can do things like enact sanctions for political purposes because US dollars are the most commonly used currency in the global financial system. The trade deficit doesn’t cause that entirely, but it plays a role.)

As Matt Levine writes: “We give the rest of the world entries in [financial] databases, they give us back food and clothing: That is a magical deal for us!”

Not everyone wins from our trade deficit. If you’re a US thingamajig-maker, and foreign companies can make thingamajigs more efficiently, then you might be better off selling to Americans without trade than having to compete in a global market. (Of course, this is only true if you can get all the materials and equipment for making thingamajigs in the US.) But US thingamajig buyers benefit when foreign competition drives down prices, and in general that outweighs the cost to thingamajig companies and workers: even China’s admission to the WTO, the most controversial trade policy of recent decades, helped American consumers more than it hurt producers.

Why to consider changing things anyway

If we’re getting more stuff as a result of the trade deficit, doing something to reduce it would make us poorer. To argue in favor of tariffs, therefore, you need to say that they’d have important second-order benefits.3 There are plenty of arguments for why that’s the case, but I think most of them don’t hold up:4

• "Negotiating leverage": It’s sound to argue that some tariffs would hurt other countries more than they’d hurt us, so we could bargain for concessions. But this just begs the question—why do we want to change the terms of trade if they benefit us already? You’d need to explain that with one of the other goals from this list (or something non-economic, which is a different discussion).5

• Reducing debt: Tariffs can reduce the national debt, but they’re not a very effective way to do so:

  • If we keep importing goods anyway, tariffs will generate revenue—but most of that is ultimately charged to US customers via higher prices, which makes them effectively a tax on Americans. Other forms of tax would probably be better: tariffs hit lower-income households hardest, and they have an outsized impact on capital equipment, which is important for sustaining growth.

  • If we stop importing goods, tariffs don’t raise revenue—but the trade deficit will fall, likely slowing down debt growth. But this means either we get less stuff, or we pay more for it by producing it domestically; in either case it’s also an indirect tax. This reduces economic growth, which is bad for the national debt. With lower growth, we can afford a smaller level of debt (because the same debt payments become a larger share of GDP). And debt rises faster when growth slows (because people pay less in taxes and get more in benefits like unemployment insurance).

• Restoring financial stability: Some finance wonks argue that America’s large trade deficit is bad, not because it’s unaffordable or makes us poorer, but just because having a lot of borrowing in the system makes it more fragile. This is reasonable, and it may have even played a role in causing the 2008 financial crisis. But regulation has made the financial system more stable since then, and there wasn’t much reason to be concerned about imbalances entering 2025. (The tariffs themselves destabilized markets, of course, but we still haven’t seen any fundamental concerns like bankruptcies or financial crises.)

• Creating jobs: This is the most popular argument, but I think it involves several misconceptions:

  • People think that the job market is bad, but unemployment is near a historical low, and middle-class wage growth is as strong as ever. (From that same report, there are also fewer people than before juggling multiple jobs or stuck in part-time work.) Jobs have become a little harder to find over the past 18 months, but the situation is still pretty good.

  • People think that manufacturing jobs are “good jobs,” but they’re not so much anymore: the average non-manufacturing job actually pays more than the average manufacturing job today, and workers voluntarily quit factory jobs more often than others. And a lot of foreign factory jobs are definitely not appealing: they involve tedious, low-paying work like stitching sneakers or assembling iPhones, or dangerous activities like tanning leather in acid baths. (I think this also undercuts the related case that reshoring manufacturing will bring back “manly” jobs.) I’m sure some outsourced manufacturing jobs would be better than some people’s current jobs, but I doubt it’s a large fraction of our imports.

  • People think that trade is mainly to blame for the loss of manufacturing jobs, but that’s more about technology replacing labor. US manufacturing employment was declining long before our largest trade deals, and the same pattern happened in other major countries. More careful analysis finds that trade explains at most 20% of the lost jobs.

  • People think that manufacturing employment is still falling, but it’s actually been pretty steady since 2010. There’s a good case that we should have done more to prevent manufacturing job losses in earlier eras, even at the cost of being poorer overall—unemployment has negative effects beyond the economic ones, and in many cases it was especially hard to find a job in the same area once a factory closed down—but that’s not happening much now.

In my opinion, the best case for worrying about the trade deficit is that our economy is more sustainable if we make more things ourselves.

The most concrete reason is that if we import critical items from other countries, and our relationship with them sours, we might find it harder to get those items. That’s usually framed as “we can’t win a war with China if we rely on it for lots of stuff” (or its corollary, “we can’t import semiconductors from Taiwan if China invades it”). That’s true, but you don’t have to imagine armed conflict for this to matter.6 If a single large country is a chokepoint for making lots of goods, they could exert non-violent power we don’t like—just as the rest of the world probably feels about American dominance of the financial system. Outsourcing manufacturing also reduces resilience to certain types of disruptions, like Chinese companies’ shutdowns during years of “zero-Covid” policies or the Ever Given disrupting global shipping.7

A second, more speculative reason: making things yourself probably helps you come up with new ideas. (Think of this as “it’s good to get in the weeds” at national scale.) There’s some evidence supporting this—manufacturing industries tend to spend more in R&D, which helps drive future growth and innovation. This might not matter in all cases, but it seems important at least in high-tech products or their inputs.

How the actual policy falls short

But if you want to change the balance of trade in order to make more stuff in the US, the current policies don’t achieve that goal effectively for several reasons.

All goods are taxed equally. This has two downsides. First, it’s unnecessarily costly, because we can’t make everything here. That obviously applies to things like coffee that physically can’t be grown in most of the US. But it also applies broadly—we just don’t have nearly enough workers to make what we’d need to in order to offset the entire trade balance.8 (Unless you dramatically increased immigration, lol.) Even in the most optimistic scenarios, we’re still going to import a lot from abroad, especially things that involve low-paying work. Tariffs on those products are a pure cost increase; they don’t move us closer to our goal.

Second, about half of what we import helps us make other stuff, such as machines for factories or components of finished products. Those “intermediate” goods are also taxed today, which undercuts the goal entirely: if a carmaker is considering moving a plant into the US to avoid the 10% tariff, but some of the parts and equipment has to be imported at a 10% tariff, it may not be worth it. (And any cars they sell outside the US are now more expensive without any benefit.) This helps explain why many US manufacturers have stopped production since the tariffs were announced. This is particularly bad because it might not be reversible—2002 tariffs on steel led to closures of US steel-consuming factories, which didn’t come back even when the tariffs were rescinded.

(The original tariffs did exempt some raw commodities such as oil and metals. More exemptions have been added, but those include finished products like iPhones, which doesn’t address the intermediate-goods problem.)

Everything is subject to change. Last week’s pause fixed some of the worst elements of the original proposal, but it also made clear that the tariffs could change at any time. From the standpoint of encouraging manufacturing in the US, that’s terrible: maybe it was better to move a factory onshore last week, but it’s not this week, and who knows what will happen in 90 days. Factories take years to build and decades to pay back an investment, so it’s not realistic for companies to make new plans now.

“Policy uncertainty” has become something of a bogeyman in economics: everyone was talking about it in 2017, yet businesses still invested like usual. But that was largely because, despite lots of talk, not much of substance happened on trade or regulation. This time is different—large cost increases are already here.

Investors are spooked. When markets nosedived after the tariff announcement, some supporters said, who cares what investors say, we’re focused on what happens in the real economy. That was dumb on the surface, because investors were reacting to a change in the real economy (higher costs). But it was also dumb on a fundamental level: if the purpose of the tariffs is to move production to the US, we’ll need to build a lot more factories, which requires investment. The stock market is down and interest rates are up, so it’s more expensive for companies (or the government) to raise money for big projects.

And although the narrative flipped to “markets are back to normal” when stocks rallied after the tariff pause, that’s not really the case: borrowing costs are still higher, a sign that there’s less demand to invest in America. Recall from the first section that foreign countries invest a lot in the US. If they stop wanting to do that—because they’re pessimistic about our economy, worried about uncertainty, or just pissed off because of punitive tariffs—that might improve our trade balance on the surface, but it would hurt the fundamental goal of producing more here.

Exchange rates might offset tariffs anyway. There’s a fairly mainstream view in economics that tariffs don’t really affect net imports at all: if we apply a tariff, then we’ll be less willing to exchange dollars for imports (because costs are higher), so there’s less supply of dollars in currency markets, so they’ll become more valuable relative to other currencies, so foreign exporters can lower prices (because they care about what they earn in their own local currency). Depending on other assumptions, that at least partially offsets the effect of tariffs on domestic production, and it might cancel it out entirely.

I’m putting this one last because it depends on a lot of other factors being held constant, which isn’t the case; to wit, the dollar is actually down this month (when that model predicts it should be way up), perhaps because foreign investors are less willing to hold US assets.

(A fifth weakness of the original proposal was that it charged each country a different tax based on our existing ratio of goods import to exports. Focusing on bilateral deficits is wrong in any case, as we saw in the first section, but it’s especially wrong from the standpoint of making our production more sustainable. That would argue for something like a flat tax, one that varies based on the absolute size of our imports, or one that favors close allies who we know we can count on in any geopolitical circumstances. You can definitely argue with the new structure and the motivations behind it, but it’s much more logical than the first version.)

***

So what would be a good plan for making US manufacturing more sustainable? I think it would start with making direct investments in American manufacturing and infrastructure, particularly for products that you think we can least afford to give up. This would cost money, but so do tariffs (through higher prices or foregone purchases); the upsides are that it would be more likely to work and less likely to poison foreign relations.9 I’m not sure if there would be a role for tariffs at all, but if so they’d be much more targeted: they’d focus on specific technologies that we want to produce here, minimize the effect on intermediate goods (at least at first), and exempt close allies. They’d be rolled out with months or years of notice so businesses can plan for them, and they’d be enacted as a law through Congress so they can be considered permanent.

(You might notice this is basically what the previous administration was doing, through the CHIPS act and to some extent the Inflation Reduction Act. There are valid criticisms of how those were implemented, but they were basically working: after years of stagnation, factory investment more than doubled from 2022-24.)

Trade has come under scrutiny from both the left and right in recent years, for reasons that are sometimes valid but more often not. If the tariffs persist and turn out as I expect, a silver lining is that maybe attitudes will change.

They claim that the book is a classic because going to sleep is itself strange, especially for children. Goodnight Moon validates that feeling.

2 | Restoration Hardware, a home goods company that imports a lot of stuff from Asia, happened to hold its earnings call at the same time as the tariff announcement. Its CEO’s live reaction might be the best five-word poem ever: "oh really, oh shit, okay."

3 | The “hot hand” is an eternal question in basketball, but this fun paper looks at it in Jeopardy: are contestants more likely to get the next question right if they’re on a streak of correct answers? It finds that they are, but contestants overestimate the effect (betting too much on Double Jeopardy when they’re hot), which is also where the original basketball debate landed. Just as in hoops, it’s hard to control for all factors, but I love the idea.1

4 | Most of this history of pizza in the US is great, because pizza is great. Especially interesting: Chicago pizza had been around for decades before it became known as “deep dish,” at which point it wasn’t that much different than what you’d find everywhere else. But once it had the name, Chicago pizzerias leaned into it, making it “deeper and dishier” over time.

5 | I think I’m fairly tech-aware, but I had no idea that a drone delivery company, Manna, is already sending hundreds of items per day to houses in one Irish neighborhood. (I heard about it from a subscription-only podcast, but this article covers the highlights.) Surprisingly, they focus on not packages, but small orders of prepared food (cups of coffee, burgers, ice cream) where speed is important and human delivery is cost-prohibitive in the suburbs. I'm a bit of a Luddite when it comes to food delivery—I winced a little when he said "Suburban living will have all the advantages of urban living: get delivery in 10 minutes to your house from anywhere in the whole city"—but I’m curious to see how it goes.

6 | Here’s a funny story of a guy whose car wouldn’t start when he got vanilla ice cream, but worked fine when he got other flavors.

7 | When I interned at Sports Illustrated in college, a veteran photography editor held a lunch-and-learn session where he told us, “winners are boring; the best photos are of the losers.” That’s stuck with me forever, but I never really thought about why it’s true. It clicked when I read Sam Miller on celebrations: players or fans of the winning team all just put their hands in the air, while losers do all sorts of different things.

8 | Matt Clancy shares a paper showing that 10% more startups are formed in a town when the first Starbucks arrives, because it’s a new place for people to collaborate. That effect size seems crazy to me, and all social science findings have to be taken with a grain of salt, but the study’s controls seem reasonable: it only holds when Starbucks is the first coffee shop in the area; there’s a similar effect for other sit-down chains but not for primarily to-go chains like Dunkin; and Starbucks openings cause the arrival of other coffee shops, creating more areas for collaboration.

9 | My mind is always blown by these behind-the-scenes with directors of live sports broadcast dictating which of dozens of cameras is going live every second.

10 | Every NYC restaurant (or SF/LA) rated by the attractiveness of people who leave Google Reviews.

11 | “Why did our six-year-old tell our four-year-old that mom was foaming at the mouth when giving birth?”

This isn’t comprehensive, and it’s inherently subjective, especially on the x-axis. (I’ve tried to ground the y-axis in data, but there’s some wiggle room in how to frame those too.) If you have arguments or other nominees, I’d love to hear them.

Deep-diving on a few interesting ones:

Work from home (seemed permanent, actually permanent): I think today’s return-to-office narrative is oversold. It’s hard to get a clear picture here because different measures vary, but government surveys show working from home increasing since 2022 (both as a share of workers who have the option, and as a share of total hours worked). And although new job postings have shifted back toward in-person, that’s mostly just because the WFH-friendly tech sector is hiring less; when controlling for the industry mix, remote postings have barely fallen.

Part of what’s going on here is that new companies are much less likely to be in-person they were before the pandemic, and that trend may even be accelerating: according to one survey, 57% of startups founded in 2023 were remote or hybrid, up from 35% in 2021. So although you’ll see headlines about older and bigger companies bringing workers back to the office, that’s offset by the small but growing cohort of new businesses.

Business travel (seemed moderately permanent, actually moderately permanent): This is one of the few I could find concrete data for on both dimensions. Industry groups claim US business travel is back to record highs, but those numbers aren’t inflation-adjusted; even taking their estimates at face value, they’ve grown 10% below inflation since 2019. Furthermore, business travel spending was growing at 1-2% in real terms before the pandemic, which means it’s more like 20% below the projected trend. And for all the rhetoric about how all meetings would happen on Zoom forever, businesses actaully projected in a mid-2020 survey that their long-run travel spending would fall by about 30%—pretty close to what actually happened.

E-commerce acceleration (seemed permanent, actually temporary): In the first months of the pandemic, people stopped going to stores and started ordering things online; prompting a lot of talk about how these new habits would stick and the future of e-commerce was permanently brighter. Five years on, people are definitely buying more online, but almost exactly the level you would have predicted by just extrapolating the pre-pandemic trend:

In contrast, the curve for online groceries seems to have shifted more permanently; I couldn’t find good US data, but in the UK and worldwide it’s still 1-2pp above the trendline.

Gig economy (seemed ambiguous, actually temporary): As a lot of laid-off workers turned to non-traditional platforms to earn money in 2020, and others got bored while working from home alone, some wondered if we were seeing a permanent shift away from full-time employment. No matter how you define it, this hasn’t really happened:

• If you look at all freelancers and contractors, they represent the same share of all workers as they did in 2019 (about 10%).

• If you look at specifically who earns on “gig platforms” (ridesharing, delivery, social commerce), it’s up from 3% to 4%, but flat in the last three years.

  • Despite what you’d expect from scrolling Reels or Tiktok, the share of “content creators” is a rounding error at 0.2% and has been declining since 2021.

• And even among those gig workers, less than a quarter actually work on gig platforms year-round.

• If you like Google Trends, searches for “gig economy” are below pre-pandemic levels, which matches my sense of the zeitgeist.

Startup boom (seemed temporary, actually permanent): There were a lot of weird economic indicators in the early months of the pandemic. Most of these were easy to explain, but a slightly puzzling one was that new business applications went through the roof. (These aren’t necessarily what you think of as “startups”—most of these are small businesses, not venture-capital-backed tech companies.)

You could tell a lot of temporary stories about this: maybe people put their stimulus checks into getting a business off the ground; maybe they were laid off and had nothing better to do; maybe they were committing fraud for PPP loans. But five months on, new company formation has remained way above pre-pandemic levels:

Usually a permanent break in a trend like that isn’t real; it means the data provider changed how they measure it, or people found a way to game the system, or something like that. But I really couldn’t find any theories for why this data is skewed—which means people are actually starting a lot more new businesses. Maybe that’s to take advantage of all the other trends on the chart, and/or maybe it’s being propelled along by the recent AI boom.

One caveat is that these new companies tend to employ somewhat fewer people and may be more likely to go out of business sooner. But as long as there’s more of them, that’s not necessarily a bad thing.

==

One thing I didn’t put on the chart, because I don’t think we know yet0 how it’ll play out, is the technological advances in pandemic prevention. It’s safe to say that we aren’t investing as much toward fighting the next outbreak (in dollars, regulations, or mindshare) as we expected from the vantage point of mid-2020. But if COVID-19 was just the tip of the iceberg for what new vaccine technology can do, or if innovations like germicidal UV light dramatically reduce transmission, we might still be much better prepared for future threats.

I don’t really think the evidence behind the other items is worth reading, but if you’re curious or want to argue about it, see this long footnote.1

NVIDIA makes GPUs, a type of computer chip optimized for processing graphics. (Technically, NVIDIA only designs them; it outsources production to specialized manufacturers like TSMC, the same way Apple doesn't actually "make" your iPhone. But the designers have the most control over the final product and earn most of the profits.) Historically, the most valuable chip companies like Intel made general-purpose CPUs; graphics were a smaller niche primarily used for video games. But it turns out that other applications have the same attributes as graphics processing—especially deep learning, the algorithm used by modern AI.2 So NVIDIA went from leading in gaming chips to leading in AI chips, which is worth a whole lot more.

You might be tempted to dismiss that story as a lucky break, and that's sort of true: NVIDIA was founded as just a gaming company, and others did most of the initial work in adapting its chips for AI. But every trillion-dollar company gets lucky somewhere; it's just not usually so obvious.3 To reach its current stratosphere, NVIDIA had to lead the gaming chip market to begin with, be open to new applications, and maintain its advantage in a more lucrative market.

Tae Kim's The NVIDIA Way is the first mainstream take on how NVIDIA took advantage of its luck. Really, it's better thought of as CEO Jensen Huang's take; Kim got lots of access to top management, which is strongly shaped by Jensen's4 worldview, and he doesn't address some controversial topics. That isn't really a criticism—any business book that isn't a hit piece will play up management's brilliance, because "learn one weird trick to make your own business better" is a better marketing pitch than "that company sure was in the right place at the right time"—but it should be read through that lens.

Kim still provides compelling evidence that a) NVIDIA did some things differently than competitors would have, and b) those decisions generally helped it succeed. But it's not so clear how much of that is replicable by other companies.

Stage 1: Moving fast

Two engineers from early-90s computing giant Sun Microsystems had the initial idea for NVIDIA: they already knew how to make graphics cards for expensive business computers that were used for architecture, video editing and animation; could they adapt that technology for cheaper PCs to bring good graphics to the masses? They brought in a former business partner, Jensen Huang, and founded NVIDIA together at a Denny's; Jensen became the first CEO and remains in that role today.

NVIDIA had an inauspicious start: its first chip made great graphics, but in order to get most of the benefits, game developers had to significantly rewrite their programs (rendering objects with quadrilaterals rather than triangles and using an integrated audio card). They might have done this if the alternative had been much worse, but since other chips were becoming cheaper and more powerful, they could get away with the status quo. So nobody used NVIDIA’s chips to their full potential, Sony canceled a major contract, and the new company was heading toward failure. According to Jensen, a more successful competitor could have bought them out easily if they hadn’t waited too long. (Jensen is constantly saying things like this, because the semiconductor industry is famous for fast rises and fast falls; one of the most famous quotes in all of business history is longtime Intel CEO Andy Grove's "only the paranoid survive." Sometimes Jensen's paranoia is hard to believe, but it was warranted here.)

Out of that crisis came the company's first major innovation: dramatically speeding up its production cycle. At first this was out of desperation; chips usually took multiple years to develop and sell, but the company only had enough money to last nine months. Instead of stretching its runway as far as it could go, NVIDIA spent a third of its remaining cash on a chip emulator. This shrank the timeline to six months, but it sped up the process enough to hit that goal—engineers could test software as if it was running on the chip, find bugs and fix them ahead of time instead of waiting to test on a complete product. NVIDIA also gave up on quadrilaterals and licensed some technology from a competitor to move back toward industry standards, finishing a prototype just in time to make enough sales at a major industry conference.

When new competitive pressures came from Intel, Jensen's response again was "go faster." The industry standard was to release a new generation of chips every 18 months, but new computers came out twice a year, and having the newest chip was a selling point, so different chip companies would go in and out of fashion. NVIDIA decided to deliver a new chip every six months so it was always top of the line.

How was this possible? Organizational design played a role. NVIDIA restructured its engineering department into three parallel teams: one that worked on major architectural refreshes on a slower cadence, and two that worked on smaller releases in parallel.

Technological innovation also helped. NVIDIA's chips had a software virtualization layer that could run instructions as if they were on different hardware. This allowed the company to be more ambitious with timelines: if it didn't have a new feature working in silicon, it could build that feature into the software instead. This would run less efficiently, but that would be a much better price to pay than pushing back a release date. It also made sure its chips were backward compatible so developers didn't have to update their programs on the faster schedule.

But a big part of NVIDIA's speed, it seems, is that they simply worked harder. "We thought there was going to be a secret sauce [for delivering a new chip every six months]; it turns out it's just really hard work and intense execution on schedules," says an engineer acqui-hired from a competitor, one of at least five different employees quoted on some variation of that theme. The anecdotes aren't quite at the SpaceX level of nearly killing yourself to get things done, but they're still pretty intense: parents brought their kids to work so they could spend a few minutes with them without leaving the office; employees were asked about their work objectives for an upcoming weekend; when manufacturing defects became a problem early on, the engineering team spent a few weeks of tedious evenings nights testing every part of every chip instead of sampling them as usual.

This just pushes the question up one level: why did people work harder at NVIDIA? I can understand why people work so hard at SpaceX—you know what you’re getting into at an Elon Musk company; his companies have a track record of doing great things; the mission is literally going to Mars; and they hire mostly 20-somethings. Similarly, I can see why it would be easy to work hard at NVIDIA today; you know you’ll get rewarded when your company's stock price has risen by 30% per year for a quarter-century.

But what set apart the NVIDIA of three decades ago, when it was a little-known hardware startup? Kim says that Jensen is just really effective at getting people to work hard. Part of that is the paranoia—constantly saying you might go out of business in 30 days is very motivating if it’s credible. And another part is that Jensen has no qualms about just telling people to push harder—sometimes, like a fitness coach, that’s all it takes to get more out of people. Finally, Jensen works like crazy himself; his typical hours early on were 9am to midnight, and he says he likes taking family vacations because "this is when I can get a lot of work done."

What can we learn here? There is no speed limit—don’t assume you have to move at the same cadence as everyone else. If you try to go faster and work backward from there, you’ll often find new ways to get around bottlenecks. You might also help push your people to work harder, but there’s some alchemy here that I don’t fully understand yet; raising your ambition is necessary but probably not sufficient for that.

Stage 2: Building a platform

After stage 1, NVIDIA was the market leader in graphics chips for gaming; that's cool, but it's nowhere near a trillion-dollar company. (NVIDIA was worth $9 billion at the end of 2011.) Its next leap came when graphics chips exploded beyond gaming into AI. Such shifts never happen overnight, but this one kind of did: at an annual machine learning competition in 2012, one team from the University of Toronto blew everybody else away with a new deep-learning model called AlexNet. Deep learning had been tried before and failed; AlexNet’s main innovation was making the model much bigger, which was possible because they trained it with NVIDIA's graphics chips instead of CPUs. Over the following years, it became clear that deep learning models are great for lots of things, and that NVIDIA's GPUs were the best way to train them.

In one sense, this gold mine fell into NVIDIA’s lap: it wasn't directly involved in AlexNet at all and only had a couple people working on deep learning at the time. But it was made possible by the company's decade-long strategy of creating chips that others could build on in new, unexpected ways.

NVIDIA's first chips in the 90s only had specific built-in methods for rendering graphics (for example, adding shadows to a 3-D image), as was typical at the time. With its 2001 GeForce chip, NVIDIA made its graphics programmable—developers could write their own techniques for applying visual effects, which might work better for a given situation in a given game.

Soon the company noticed that some researchers were using the same pipelines to do stuff that wasn't graphics rendering at all; they were doing scientific computations like simulating fluid dynamics. This worked well on GPUs because it had the same structure as graphics processing (doing lots of calculations across a surface in parallel), but pulling it off involved a ton of tedious work. So NVIDIA redesigned its chips to support a new software layer called CUDA, which had built-in support for doing non-graphics computing.

The company invested a lot in CUDA: it went through a four-year development process (in contrast to its typical fast, incremental iteration); it made the new feature available on all of its chips to attract the most developers to write programs for it (not only an initial high-end model); and it gave away chips to researchers to encourage them to experiment further. And it did so without an immediate payoff: there were a few early use cases (Adobe used it to make Photoshop more efficient; protein-simulation researchers used it to replace supercomputers with cheaper PCs), but nothing that changed the company's fortunes. Between higher spending and the Great Recession, NVIDIA's stock price fell by 80% in the late 2000s. Then came AlexNet and the deep learning revolution, and the company’s growth took off.

Looking back, NVIDIA's path to AI success is easy to explain: the company followed a classic platform strategy, creating a product that others could develop their own new programs for. Usually, when people talk about why platforms are valuable, they focus on "network effects"—if all the customers use one platform, everyone will build products on top of it, which means all the products are built for that platform, so all the customers have to use it, etc. That's a big advantage for NVIDIA today: everyone writes AI software for CUDA, which means everyone using AI software uses CUDA, which means any new programs are written for CUDA, etc.

But NVIDIA's leap illustrated another benefit of platforms: they externalize innovation. When anybody can build a product that uses your technology, they might come up with something that you didn't think of, or invest in something more quickly than you can. That means you can explore a lot more ideas more quickly, making it more likely that someone will find a killer idea. For example, when Covid first hit, Etsy was one of the few places you could buy face masks—not because the company had amazing foresight, but because its e-commerce platform allowed sellers to list whatever they wanted. Developers have built all sorts of applications on NVIDIA chips with the help of CUDA—the keynote at their annual conference a few years ago featured "Robots and digital twins and games and machine learning accelerators and data-center-scale computing and cybersecurity and self-driving cars and computational biology and quantum computing and metaverse-building-tools and trillion-parameter AI models"—and deep learning has been the clear winner.

(Another example, curiously not mentioned in the book: during the crypto boom of 2017-18, customers started buying lots of NVIDIA GPUs and adapting them for blockchain mining. NVIDIA didn't really endorse this; it allegedly underreported crypto sales and later tried to disable mining-friendly features of its chips. You can spin that into another positive management story if you want: NVIDIA realized that crypto demand would be temporary—miners eventually moved on to more specialized chips—and didn't want it to crowd out longer-term deep learning users. But the blockchain boom helped prop up prices and sales at a time when NVIDIA needed profits to reinvest in future AI dominance.)

What can we learn here? Software platforms encourage innovation and make your technology more useful. Unfortunately, everybody knows that now! NVIDIA was ahead of its time in investing in CUDA—other companies, such as Windows and eBay, had successful platforms by the early 2000s (rare Beanie Babies were the face masks of the dot-com era), but the concept of platform strategy wasn't codified and popularized until the early 2010s. It’s still a good idea to follow that path, but it’s now conventional wisdom.

Stage 3: Thriving at scale

2002 marked a turning point in Jensen's leadership of NVIDIA. The company was reeling from a disastrous product launch—its chip was late, customers mocked its cooling system for sounding as loud as a vacuum, and competitor ATI already had a better-performing version available. How did that happen? Key features were dropped because of poor communication between hardware, software, and design teams; the design wasn’t right because too much time was wasted negotiating a contract with Microsoft before getting product specs; and leadership's focus was fragmented because it was chasing new customers and acquiring a large team from a competitor. In other words, NVIDIA had the same problems as any growing business.

All CEOs are familiar with these challenges, and they love to talk about how they need to act like a smaller company—reduce hierarchy, get closer to the details, focus on innovating rather than playing corporate politics—but they rarely do anything meaningful about it. Jensen, however, took these ideas to the extreme:

• He radically flattened NVIDIA: whereas most CEOs manage 5-10 executives, Jensen has nearly 60 direct reports. His management isn't about mentorship or 1:1 conversations; it's about making decisions efficiently and getting information to and from whoever needs it. (An outside executive says that you usually hear different stories when you talk to different people from the same company, but NVIDIA employees never contradict each other.)

• He established a "Top 5 Things" routine: every employee is expected to send regular emails with five important things that they're working on or that they've noticed outside the company. These emails go not just to their manager and team, they also go directly to Jensen, and he really reads them and sometimes replies. This is Jensen's way of bypassing the curse of hierarchy, in which bad news or unusual tidbits are sanded down when they get communicated up the chain of management; by hearing directly from every employee, Jensen can pick up on anomalies that might be harbingers of a new trend: “It’s easy to pick up the strong signals, but I want to intercept them when they’re weak,” he says.5 For example, when a bunch of top-5 emails mentioned interesting advancements in machine learning, Jensen added resources to a project to develop ML frameworks for CUDA.

• He structured the organization to avoid internal politics. Most companies eventually break into different business units (each focused on a specific country or product line), which inevitably leads to some infighting for resources. But NVIDIA is "functionally organized"—everyone is part of a reporting line that’s responsible for the whole business. There are downsides to this approach, but it means there's less competition and people can change priorities more quickly. (Jensen also publicly criticizes other executives unusually harshly—in his words, he aims to “torture people into greatness.” His ostensible reason is so everyone else can see that there are high standards and that they can do better, but you can also read this as a president undermining his generals to make sure nobody consolidates enough power. The book doesn't have any Ray Dalio-like examples of Jensen explicitly putting himself above everyone else, but it also doesn't have any serious examples of someone calling Jensen out on something and him accepting it, which you'd think would make good fodder if they existed.)

The upshot: NVIDIA is completely built around Jensen, much more than most companies leverage their CEOs. He analogizes this to a racecar: “I built Nvidia in a way where I could drive it, and manage it, and steer it perfectly.” Workers have a lot of agency and responsibility, but it strictly falls within boundaries set by Jensen alone. ("Don't take this the wrong way, but you may not have the brainpower or wherewithal to detect something I think is pretty significant," he says to his executive team.) Even the seemingly quirky note that NVIDIA uses whiteboards in meetings instead of PowerPoints probably fits into this: Jensen can walk up to a whiteboard and take control in a way he couldn't do with a slide deck. (People vividly remember one meeting where Jensen listened to someone else present for a full hour, because it was such an anomaly.)

In this sense NVIDIA reminds me a lot of Skunk Works: Kelly Johnson eliminated bureaucracy by wielding all the authority himself, which meant anyone else could go to him for a decision and get unblocked quickly. But that was at 50-person scale; Jensen is operating NVIDIA at 30,000-person scale. (Ben Thompson draws a parallel to the Steve Jobs era at Apple—notably, another functional organization.)

There was a direct line to success from moving faster in stage 1 and building a platform in stage 2. You can't do the same with NVIDIA's organizational changes in this stage; if you're cynical, you can argue that it's just a coincidence, and they're just living off their AI platform moat. But I do think NVIDIA is uniquely positioned to see what's coming next and change priorities quickly, and that's been a big factor in their dominance:

• After AlexNet, NVIDIA redirected a lot of resources to AI quickly. They built an AI-optimized library for CUDA and supported less precise data types that would make training neural networks more efficient; most importantly, they added a new core processor for matrix multiplication to their next chip design even though it was only a few months away from production (by which point designs are usually locked in except for bug fixes).

• NVIDIA was prepared for the large language model wave. A few months after Google released the foundational paper on transformers, NVIDIA started building support for those models, which it released a month before ChatGPT. NVIDIA worked with its manufacturing partners to build out new capacity and speed up their cycle times; it still wasn't enough to keep up with the massive influx of demand, but it helped.6

• In 2019 NVIDIA bought Mellanox, a networking company to help link multiple GPUs together for training larger models; Mellanox's business has septupled since the deal, and Kim quotes one executive as saying it's “going to go down in history as one of the best acquisitions ever.” This belongs in the lucky-break category to some extent—NVIDIA wasn't shopping for this technology, but Mellanox wanted to sell after getting pressure from activist investors; they first got an offer from a different company, then reached out to NVIDIA—but Jensen at least capitalized on it fully.

What can we learn here? This is the most tantalizing part of NVIDIA’s story: it’s a trillion-dollar company that acts on opportunities with the speed of a much smaller organization. There are a few little things that anyone can and probably should copy, like having a Top Five Things-like mechanism for sharing specific ideas directly to leadership. But the real magic seems to be Jensen’s “racecar”—minimizing politics and maximizing the speed of decision-making by centralizing authority in one driver, with everyone else fulfilling a specific role around them. Many start-ups are run this way, but Jensen shows that it can work at much larger scale (and the coming age of AI agents might increase the possible scale further).

How many people can drive such a powerful car, though? Not only do you need to see the big picture well enough to make high-level decisions, you need to be an expert at all of the important technical details for everyone to listen to you. (Kim writes about one meeting in which Jensen basically invents a new technique on the fly for using deep learning to enhance graphics; it’s reminiscent of Skunk Works meetings where Kelly Johnson estimates complicated values off the top of his head that turn out to be almost exactly correct.)

And even if you have all that, what happens when you leave? Unless there’s someone else who’s just as brilliant and charismatic as you are, and has lots of experience with the company—and was willing to defer to you for all that time for some reason instead of striking out on their own—nobody can replace you, and the organization will have to transition to more traditional management. Jensen himself pretty much admits that he has no idea what will happen to NVIDIA after he leaves, but he doesn’t plan on retiring soon, so he doesn’t really care.

That means the racecar model is only possible for founder-led companies with a technical and business savant at the helm. But for the small set of businesses in that position, NVIDIA shows that the ceiling is nearly limitless.

Previously: 2023, 2022, 2021, 2020, 2019

1 | Mice don't particularly like cheese. They're known for eating it because they could easily find it aging uncovered in cellars, but exterminators say their favorite food is actually peanut butter. [Livescience via Kent Hendricks]

2 | If you pick a random ancestor of yours from at least 12 generations ago, you probably don't share any genes. [gcbias via ACX]

3 | Parents of newborns today who follow recommended practices sleep an hour less per night than previous generations. About half of that is from putting babies to sleep on their back instead of facing down (which is important for preventing infant deaths, but causes them to wake up more often); the other half is due to breastfeeding instead of formula feeding (which is said to have health benefits, but takes longer and doesn't fill them up as well1). [@ruthgracewong]

4 | NASA tried to put Big Bird onto the 1986 Challenger space shuttle launch to get kids interested in science, but there wasn't enough room. (That shuttle exploded on live TV, with no survivors.) [NBC via Shutdown Fullcast]

5 | Even with bigger homes, bigger vehicles, and countless new gadgets, the average American uses less energy today than in the 1960s because everything is much more efficient. [Not the End of the World]

6 | Some children (though not all, and perhaps not even many) are now learning a different version of the alphabet song: ABCDEFG / HIJKLMN / OPQ / RST / UVW / XYZ. This is because the fast "LMNOP" and the ending "Y-and-Z" can make it hard to hear the letters distinctly. [@mindfulteacherrachel]

7 | NFL teams collectively serve 80,000 Uncrustables in a season (mainly at halftime of games and after practices). The Broncos alone go through 700 per week. [@samkouvaris]

8 | Donations of medical equipment to poor countries often backfire because the recipients don't have enough materials or expertise to fix things when they break. (For example, a town in Indonesia received eight incubators for premature babies in 2004, but they were all out of commission by 2008.) Aid providers now seek out equipment that's not state-of-the-art but is instead designed to break in easily repairable ways. [Where Good Ideas Come From]

9 | Pokémon Go made more revenue in 2023 than Duolingo. [@juliey4]

10 | All 24 eligible players who have made an All-WNBA team were ranked in the top 40 of their high school recruiting class; 70% were in the top 10. That's a much stronger correlation with high school reputation than NBA players have, probably because women tend to finish growing at a younger age than men. [The Next]

11 | Teenage girls are now 20% more likely to have tried marijuana than boys. (Ten years ago it was the opposite.) [MDPI]

12 | By one estimate, Prohibition increased the share of American GDP spent on alcohol, because it increased prices more than it reduced consumption. [The Rise and Fall of American Growth]

13 | There are more golf courses than McDonald's in the U.S. [NYT]

14 | Muppets creator Jim Henson was supposed to play Big Bird on Sesame Street, but the show's costume designer "did not think that Henson was walking like a bird is supposed to walk," so Caroll Spinney got the job (and held it for nearly 50 years). [Wikipedia]

15 | People can feel imperfections as small as 10 nanometers (.00001 millimeters) on a smooth surface. [@AskYatharth]

16 | Paris Olympics organizers ordered more than two million bananas for the dining halls. For 15,000 competitors over two weeks, that comes out to 10 bananas per person per day. (I'm not sure if coaches or other non-athletes also dined there, but they'd eat less anyway, and the article makes it seem like they're planning around athlete needs in particular.) [Eater]

17 | In the 1980s, New Jersey and Chicago banned teens from using pagers (fearing they were used for drug deals). [Pessimists Archive]

18 | None of the sentences in the original "Ship of Theseus" Wikipedia page remain there today. [@depthsofwiki]

19 | In 2004, about one gigawatt of new solar power was installed worldwide; in 2024, nearly twice that was installed per day. [@CardiffGarcia]

20 | If you tell ChatGPT that you're a Dallas Cowboys fan, it'll think you're more conservative, and it'll avoid left-leaning responses; it does the opposite if you say you're a Detroit Lions fan. (For some reason, it tends to avoid answering sensitive questions at all if you say you support the Chargers.) [ArXiV via Ethan Mollick]

21 | There's no evidence that any immigrants' names were changed at Ellis Island. [Documents to the People via Marginal Revolution]

22 | Cookie Monster's cookies are made of things like puffed rice, instant coffee, and black glue chips—optimized to explode amusingly and not get stuck in his fur. [NYT via Kent Hendricks]

23 | In one study, parents who were told their child had eaten a lot of sugar said they were much more hyperactive than parents who weren't, even though all the children actually got a placebo. [Otherwise]

24 | writing in all lowercase letters generally saves you in data storage, because all-lowercase text can be compressed more efficiently. [endtimes.dev]

25 | The first drug test that caused someone to lose an Olympic medal was for beer. (He was competing in the shooting portion of the modern pentathlon, where alcohol arguably enhances performance by slowing your heart rate.) [SBNation]

26 | In the 1950s, the New York Times reportedly refused to run an ad for a breast cancer support group, not only because it didn't want to say "breast" in print, but also because it didn't want to say "cancer." (It suggested "diseases of the chest wall" as an alternative.) I, and many others, initially read this as a larger sign of the paper's and society's unwillingness to face hard truths about cancer, but actually the Times ran plenty of articles about breast cancer around then; the ad story must have been one editor's idiosyncratic decision. [The Emperor of All Maladies; Dan Nguyen]

27 | "Baby Shark" probably evolved from the Jaws theme song. [Decoder Ring]

28 | America has more homicides than other developed countries, but that's not new; even in the early 20th century the US homicide rate was five times higher than in any European country. [The Rise and Fall of American Growth]

29 | Tattooing was technically illegal in NYC from 1961 to 1997. [NYT]

30 | "Baker's chocolate" and "German chocolate cake" are actually named after people. (Don't even ask about San Francisco's "Main Street.") [Things Unexpectedly Named After People via ACX]

31 | The average home in every mainland US state is larger than the average home in any European country. [@StatisticUrban]

32 | Half of NYC's thousands of bodegas (and now a large share of its unlicensed weed shops) are owned by Yemenis. [Search Engine podcast, NYT]

33 | Administrative costs have been the same share of college budgets for 50 years. The long-term rise in tuition is instead driven by a) governments paying a smaller share of costs, and b) faculty wages rising (in line with other highly educated professions). [Alex Tabarrok]

34 | LEGO now heavily advertises sets to build a specific object instead of generic pieces, which is often used as a sign that we're suppressing children's imagination. But another (non-mutually-exclusive) explanation is that there's less competition: other companies could create generic blocks that are almost as good, but nobody else has a license from Disney to create Star Wars blueprints. [Byrne Hobart]

35 | Neptune is actually white-ish; the original photos in which it appears blue were not color-corrected. [The Intrinsic Perspective]

36 | Before COVID-19, people of all incomes lived 10-15 miles from work on average. Today, low-income workers live 18 miles away, and the highest-income workers live more than 40 miles away. [@crampell]

37 | In the early days of home telephones, AT&T built a "dropping machine" to test if particular models could handle being hung up casually thousands of times. [The Idea Factory]

38 | According to an independent study, Uber and Lyft reduced US vehicle deaths by about 5%, presumably by providing an alternative to drunk driving. [@albrgr]

39 | "Wacky inflatable arm guys" have been around for less than three decades. (They were invented for the 1996 Atlanta Olympics.) [Summer of Champions]

40 | Apple sorts its iPhone boxes and lids into three size groups and pairs those together so they fit ever so slightly better. [@Ryan_Gasoline]

41 | Dialysis treatment alone accounts for about 1% of the federal budget (via Medicare). [Complex Systems]

42 | Headlights on 2024 cars are nearly twice as bright as those on 2015 cars. [The Ringer]

43 | Death during childbirth is more common in the US than in other countries, which is very bad—but oft-cited statistics showing that it's getting much worse over time are misleading; that's mostly due to changes in how deaths are reported. [Noahpinion]

44 | Sending your partner funny Tweets/Tiktoks/Reels throughout the day is called “pebbling,” like how penguins leave pebbles in their partners' nests as a sign of affection. [@thisone0verhere]

45 | The official distinction between a horse and a pony is only based on size, not age. [Money Stuff podcast]

46 | Jessie J wrote the first lyrics to "Party in the USA" for her own debut album, but the label didn't think it fit her, so together they pitched it to Miley Cyrus. (Jessie J came from the UK, so original lyrics included "I hopped off the plane at LAX with my tea and my cardigan" and "jumped in the cab and the driver's on the left side.") [Reels, originally a 2014 Glamour interview]

47 | The median Gen-Z American today is more than twice as wealthy (accounting for debt) than Millennials or Gen X were at the same age. [@ernietedeschi]

48 | Old sports photos sometimes have a hazy background because so many fans were smoking. [Petapixel via Tom Whitwell]

49 | In the 1980s more than 20,000 people per year were injured by automatic garage doors (including a few dozen fatalities). In 2022 there were a total of 34 such injuries, because the industry imposed safety standards like sensors to detect if someone is in the way. [@gengelstein]

50 | Shaquille O'Neal paid for 15 friends to get a master's degree so that he had enough people for the University of Phoenix to give him an in-person course. [@DiabolicalSpuds]

51 | Driscoll's doesn't actually grow any of their berries. (It does research to identify good breeds and then contracts those to third-party growers.) [Odd Lots]

52 | NYC air conditioners look like they're stargazing with their legs out the window. [@AOWTOUDOUZAT]

Science and technology

1. Not the End of the World (2024): Loved how this puts climate change and other environmental challenges into perspective—we need to do a lot more, but we've already made some progress, so we aren't doomed—and how it backs everything up with both analysis and empathy. (See a longer review here.)

2. The Emperor of All Maladies (2010): As good as advertised (by the Pulitzer). I appreciated how it built up the history of science that led to cancer breakthroughs and mixed individual stories in. It wasn't very dense with "oh wow that's interesting" moments, but I got a lot of background knowledge on cancer, which seems useful.

3. Normal Accidents (1984): A surprisingly compelling framework for complexity given its age, even if it doesn’t completely hold up to scrutiny. (See a much longer review than you ever wanted here.)

4. Why Machines Learn (2024): Mostly review for me, but if you don't know how machine learning works and want to get into the math then this is great. (But I think it's really about "how" machines learn, not "why." Also beware it doesn’t really say much about the current wave of large language models.)

5. Complexity: A guided tour (2009): Also mostly review for me, but easy to read given the complicated topic, with compelling examples.

Management and innovation

1. Skunk Works (1994): More personal and with more war stories than I expected, which made it a surprisingly easy read for the wisdom contained. (More in this review.)

2. High Output Management (1983): Perhaps the single most influential management book. It held up in a way some classics don’t—even though most of the main ideas have become conventional wisdom, I found some new framings that I liked.

3. Dealers of Lightning (1999): For a book about a place that invented great things (Xerox’s PARC research center), I liked how it also spent as much time if not more on its limitations (messy management, few commercial successes). I would have preferred it spend less time on each individual’s background and more on the actual innovation done at PARC, however.

4. The Idea Factory (2012): The history of AT&T’s Bell Labs research department, which invented a whole lot of stuff in the 20th century. The two chapters on the inventions of the transistor and information theory were very clear, but everything else came in bits and pieces and I couldn't follow what was important, which was disappointing for such a rich topic.

5. Working Backwards (2021): Very good, unless it's the first book you read about Amazon. That's because its weakness of this book is that, amid all the details about how the company works, it doesn't really communicate what actually drove Amazon's success. (The very first chapter is a tedious description of Amazon’s hiring process, except it admits they did things totally differently for their first decade, which still worked out okay.) Once you've gotten the big picture about Amazon elsewhere, then read this book to see all the details.

6. An Elegant Puzzle (2019): A book about engineering management that’s very dense with insights, which is good (I’d like to come back to several parts again) and bad (it's dense in part because the structure doesn't make a lot of sense so everything feels random). I’m a big fan of Larson’s blog, which covers most of the same content, but I found at least one new framework that I keep coming back to: paradoxically, managing at startups is more about execution—you already have lots of ideas, the limiting step is how quickly you can test them—whereas managing at big companies is more about innovation.

7. Scaling People (2023): It devolves into an overwhelming textbook eventually but I really liked the first chapter.

8. Bury my heart at conference room B (1999): This really is a “one big idea” book—everything but a small part of the middle feels like filler—but the big idea (understand your own personal values and genuinely live them at work) was good.

History

1. More Money than God (2010): The definitive history of hedge funds, particularly impressive for Mallaby rejecting traders’ (often generic) explanations for their own success and actually doing his own analysis.

2. The Big Roads (2011): One of the most important topics that I really didn't know much about. The main point is that although Eisenhower got his name on the US interstate system by signing the financing bill in the 1950s, most of the real planning happened decades earlier. And I now understand things like the difference between "US highways" and "interstates."

Other nonfiction

1. Recoding America (2023): This is nominally about a lack of user-centricity in government technology, but the same lessons apply to government processes much more broadly. (See a longer review here.)

2. Why We Love Baseball (2023): I gorged on the “random stories about baseball history” genre 25 years ago and thought I had outgrown it, but this book had novel angles on the classics and some brand new tales.

3. Gambler: Secrets from a Life at Risk (2023): The first half was compelling enough to make me miss my train stop, even if I didn't exactly want to be friends with him. But the second half was mostly just axe-grinding.

Fiction

1. The Dogs of War (1974): Mostly about the logistics of planning an illicit war, which is more interesting than it sounds, but there were also thrills throughout and an interesting ending. I loved every part.

2. Table for Two (2024): I liked the short stories a lot (Hasta Luego the best), and I loved the long story in the second half, although I’m not sure why it wasn’t a book of its own.

3. The Rachel Incident (2023): I started and finished it in one day, which said more about the day than the book, but still said something good about the book.

4. The Rider (1978): I'm not into cycling at all but I loved the pacing, the extreme detail about tactics and opponents, the flashbacks. It's basically Levels of the Game, but about an obscure event instead of a big one, and fictional. (Though I'm a little embarrassed to say I didn't know it was fictional when I read it.)

5. Loneliness & Company (2024): This gave me a lot to think about—not even the ending so much as the subplots (how would someone with no social life codify everything about social interaction; what if talking about an issue really does make it worse; how would you live to maximize new experiences). I also enjoyed the juxtaposition of futuristic ideas with things that we already live with today.

6. Tom Lake (2023): I enjoy pretty much anything by Ann Patchett, especially how this one covers so much ground in not very many pages. A few plot points annoyed me a bit after I was done though.

7. Bel Canto (2001): See above. Not at all what I expected coming in; it was a little hard to suspend disbelief at first, but when I did it was beautiful.

8. The Glass Hotel (2020): I only sort of liked the plot but I loved the exposition.

Parenting

1. How to Talk so Little Kids Will Listen (2017): I really wanted a book for even littler kids (1-2 instead of 2-7, which is a pretty wide age range). But the advice felt promising and has been slightly useful.

2. Raising Securely Attached Kids (2024): The writing was a little annoying (too many random analogies, a lot of trying to sound scientific that didn't land), but it gave me one good point of motivation, which is all I ask out of a parenting book.

Children’s books (1-2 years old)

These are my own favorites, not necessarily my daughter’s, but I only considered books that she also liked a lot.

More fun at 1.5-2 years:

1. Kindness Makes Us Strong—A nice message and lots to look at on each page.

2. Triangle, Square—Aesthetically pleasing with just the right amount of whimsy. (Circle is fine but not at the same level.)

3. Go Dog, Go!—A quick read, and now I say “Go dogs go, it’s green ahead!” at every crosswalk. (The Netflix series is good too.)

4. Brown Bear, Brown Bear—A classic for a reason; it’s pretty and easy to engage with.

More fun at 1-1.5 years:

1. Peekaboo House—My favorite pop-up book.

2. Splash!—This became a really fun favorite before a beach trip—simple words and pictures with a solid rhyme and rhythm.

3. My First Words—Great for early speakers because of how it focuses on first-ish words but in the context of full sentences.

If you also want to see the books I didn’t like, find me on Goodreads.

Canonical stories include:

• PARC: Xerox’s computing research unit that created the first personal computer, object-oriented programming, Ethernet, and the laser printer in the 1970s.

• Bell Laboratories: AT&T's R&D group in the mid-20th century, particularly its basic research division, which made Nobel Prize-winning discoveries such as the transistor and information theory.

• ARPA (now DARPA): an innovation program run by the Department of Defense that continues to work on military technologies, but in its 1960s heyday also created the Internet, major components of GPS, and ancestors to PARC's PC.

• RAND: a government-funded think tank with leading mathematicians and social scientists that developed game theory and the foundation for ICBMs in the Cold War era.

Perhaps the most popular such book is Skunk Works, about the eponymous research team inside Lockheed that created cutting-edge military aircraft, such as the U2 spy plane and the first stealth fighter. (Not coincidentally, it's also the most entertaining read of the bunch.) "Skunk Works" has become a generic descriptor for any project to focus a small team on innovation within a larger organization, although few have come close to replicating the original's success.

What I find most interesting about Skunk Works, though, is how it breaks many of the patterns from all the other stories.

The rules of innovation

Some common themes of famous research organizations include:

Stable, long-term funding. Research on the frontier of science takes a long time and requires expensive equipment, so the best organizations have ample resources and aren’t afraid of losing them tomorrow. Universities are the hallmark of this model—professors have tenure, and they're funded by predictable enrollment funds and endowments—but companies and governments can achieve something similar:

• Bell Labs was funded by AT&T's monopoly profits; its mission was to find technologies that would help the parent company in 5-10 years. Thus it invested in time-intensive projects like launching the first communications satellite, which helped it transmit messages across the country more easily (until the government took control of space technologies).

• DARPA is less than 1% of US defense spending, so its budget isn't under much threat. Individual projects can be defunded in theory, but according to Ben Reinhardt, about 90% last for a full 5-ish-year term; this allows program managers to form long-term relationships with leading academic or industry researchers.

No need to make profits. World-changing technologies are totally different than what came before them, so anyone working on them has a high chance of failure—and even if they succeed, it might take decades to be commercialized. If you push people to focus on the bottom line, they'll do on incremental projects that have a higher chance of succeeding quickly; that might be useful, but it's not the stuff of legends.

• PARC's most famous inventions didn't benefit Xerox financially at all; its personal computer was only used for internal research (their demo inspired elements of Apple's Macintosh) and Ethernet became an open protocol. In Dealers of Lightning, Michael Hiltzer claims this reputation is oversold—PARC incubated the laser printer, which arguably paid back all of its other investments—but he doesn't deny that its leaders weren't held to any commercial pressure, allowing them to invest time and computing resources into speculative projects.

• RAND set aside 1/3 of its time to work on basic research with no immediate demand; this was funded by the rest of its work (applied-research contracts with the Air Force and other government agencies).

The best talent available. Great research is naturally done by great researchers; the best organizations find them wherever they are and entice them to join by offering more money, great coworkers, or an important mission.

• Per The Idea Factory, AT&T was among the first American companies to hire PhDs (previously, businesses looked for engineers who could get things done, not theorists who could find new things). And it bounced back from the Great Depression relatively quickly, enabling Bell Labs to outbid peer businesses and universities for top doctoral graduates such as William Shockley, a future Nobel Prize winner for co-inventing the semiconductor.

• PARC employed basically all of the top computing researchers in the early 1970s, not only because it paid more, but also because the first hire in its computing division, Bob Taylor, already knew everybody in the field and had their respect.

• RAND didn't pay crazy salaries from what I can tell, but it offered intellectual freedom and flexibility, and the mission of helping America win the Cold War attracted top talent, including more than two dozen eventual Nobelists.

Experts in lots of different disciplines. New ideas often come at the intersection of different fields, so having people with different academic backgrounds is important, as long as they bring deep expertise in their own subject.

• Bell Labs employed 13,000 people at its peak, though only about 10% worked in basic research. Their backgrounds spanned subjects (chemistry, electricity, acoustics, optics, math, psychology), but there were no "departments" like in a university; Bell's campus was even designed to break down informal silos (labs were often intentionally assigned far away from someone’s office so they have to go out and see colleagues).

• RAND initially focused on math but soon brought in lots of social scientists; it was structured by department, but those were blurry (according to Asterisk, an anthropologist joined the math department, and an MD joined aeronautics), and most projects involved multiple departments anyway. It grew to 200 researchers in less than two years.

A decentralized structure. A research organization can generate more ideas when individuals have the power to decide what they work on; if a boss told everyone what to do, they'd only test out the bosses' ideas (which would probably be the bosses' boss' ideas, etc).

• PARC had a very flat organization—all of the several dozen computing researchers reported up to one manager—and gave everyone freedom to decide what they worked on; someone developed the first computers capable of displaying colored videos and turned it into an avant-garde collaboration with artists, even though everybody else thought he was crazy.

• Unlike many government programs, DARPA lets program managers spend funding more or less however they want; each of them in turn collaborates with lots of experts in academia or business, who take on some DARPA research in addition to their other work.

Fruitful sharing of ideas. Getting challenged by peers or hearing about other ideas can inspire new directions for research.

• PARC had a weekly "dealer" session where one speaker presented their work and got grilled by everybody else; Bell sponsored external lectures and education series; DARPA programs hold workshops where all the researchers get together in one place to meet each other and discuss ideas.

A different way to change the world

Skunk Works, however, breaks a lot of those rules:

• Lockheed limited its investment in Skunk Works: its name comes from its original location, a makeshift office set up underneath a tent next to a smelly plastic factory, and its founder, Kelly Johnson, did it as a side project early on while fulfilling his full-time engineering duties at the parent company. So instead of doing open-ended research projects, Skunk Works had to be scrappy and fast. Its first project, a WWII jet fighter, was contracted with a timeline of six months (and was finished a month ahead of schedule); the first U-2 plane was completed in nine months.

• Skunk Works spent a lot of effort managing costs—not only financially, but in wasted time. It returned $2 million to the government, 10% of its U-2 contract, for coming in under budget, and another time canceled a project when it was barely off the ground but didn't seem promising. And it used existing components and familiar suppliers whenever possible, reducing the risk that one bespoke part would delay everything. By the 1980s, Ben Rich estimates that Skunk Works alone would have been a Fortune 500-size business if it had been independent.

• Johnson was a savant—he routinely estimated complex quantities outside his area of expertise almost perfectly, and he had tremendous respect from Lockheed's leadership and in the top ranks of the Air Force. But otherwise, individual talent doesn't seem to be Skunk Works' differentiator. Johnson got to pick top engineers from Lockheed, who were doubtlessly very impressive, but he didn't search far and wide for experts the way other organizations did.

• And their expertise wasn’t especially broad, in part because there weren’t very many of them. The U-2 project was launched with only 25 engineers and topped out at 80 (something like 10-25% of the industry standard). This was partly due to Lockheed's desire to minimize costs, but also due to Johnson's belief that smaller teams would be less siloed and spend more time building instead of communicating.

• Workers had a lot of responsibility to solve their own project, and in theory anyone could suggest big ideas, but for the most part everyone trusted Johnson to set the direction. (The best counter-example is the stealth fighter, which was suggested by a younger specialist who had come across the necessary science in an obscure paper published in the USSR; not coincidentally, that happened shortly after Johnson retired, and over his objection from the sidelines.) In fact, part of why Johnson liked working with a small team was so that it could be extremely centralized—anyone with a question or a problem could go to Johnson, and he could make the decision, eliminating all bureaucracy.

• Most of their projects were highly classified, so they couldn't really engage outside experts, or even others within the company. If they needed extra help, they had to give other teams generic tasks that wouldn’t reveal what the final product might be.

How can we reconcile Skunk Works' success with the lessons above? To some extent, it shows that there are multiple ways to be innovative, which should be comforting. But there are also some ways in which Skunk Works achieved the same thing with different means:

• It had freedom to choose its own direction, not because profitability didn't matter, but because Lockheed's leadership had tremendous trust in Johnson.

• It avoided bureaucracy, not by having hands-off management, but by having decision-making authority centralized in one person.

• It had excellent talent at least at the top, though it was recruited internally and not intentionally diverse.

• It tested out ideas not by spreading them widely but by putting them into practice and iterating quickly to make them better.

The Monty Hall problem is a brain teaser based on the 20th century game show Let's Make a Deal. Host Monty Hall shows you three doors; he tells you that one contains a car, and the other two each contain a goat. You pick one, and Monty says, let's build the suspense and show you what's behind one of the doors you didn't pick—which is inevitably a goat. Then he asks, do you want to switch to the other unopened door?

The answer is yes: switching doubles your chances of getting the car. (Assuming, of course, that you want the car more than the goat.)

If you've never seen this problem before, that sounds crazy; if you've seen it and understood the solution, it sounds obvious. That's true of most brain teasers, though. The Monty Hall problem is uniquely fascinating because everyone who knows the answer is absolutely terrible at explaining it to others.

I've heard dozens of people try to explain the solution, and not only has it never worked, you can tell right from the start that it's going nowhere. "You start with a 1/3 probability of being right, and that doesn't change," they usually start, which sounds like (and is) bad statistics. "This will help: imagine there are 1000 doors with 999 goats...", they continue, and that in fact never helps. They're in good company: the puzzle became famous when Marilyn vos Savant, then billed as the world's smartest person, gave the correct solution in a magazine column and got thousands of letters from irate PhDs arguing she was wrong.

(Mathematicians studying the multi-armed bandit problem during World War II supposedly claimed it was so frustrating that they should drop it on Germany so enemy scientists would waste all their time trying to solve it. The Monty Hall problem is the modern equivalent; it's computationally simple, but drop it into any group of people and they'll collapse into argument.)

I have two theories for why this particular puzzle is so hard to explain. The first is that it presents as a statistics problem, but it’s really a game theory problem. So here’s my own attempt to explain it:

• You don’t know anything about the doors, so from your perspective, there’s a 1/3 chance the car is behind each door. (I think everyone can agree on that.) Let’s say you choose door 1.

• Now Monty shows you there’s a goat behind door 2, and this is where things go off the rails.

  • If Monty flips a coin to pick which door he opens, and it happens to be a goat, then all you’ve learned is that the car isn’t behind 2. Doors 1 and 3 were equally likely before, so now they’re still equally likely, and it’s 50/50 whether you switch or stay. This is the obvious “statistics answer.”1

  • But put yourself in Monty’s shoes. He knows what’s behind each door, and he’s a game show host who wants to build suspense. Is he going to show you the car, leaving you with two goats? Of course not! He’s going keep the car hidden, so everyone still thinks you might win.2

• So there are three ways the game could go:

  • The car is behind 1. It doesn’t matter which door Monty shows you; that door is a goat, and so is the remaining door.

  • The car is behind 2. Monty won’t show you 2, because that would ruin the game, so he shows you 3 (a goat).

  • The car is behind 3. Monty won’t show you 3, because that would ruin the game, so he shows you 2 (a goat).

• Remember, there was originally a 1/3 chance the car was behind each door. So there’s now a 1/3 chance the car is behind your door (1), and a 2/3 chance the car is behind the other door that Monty didn’t show you.

So you’re better off switching, but not just because of statistics—it’s because you can predict how Monty will respond in each scenario.

Which leads to the second theory: switching feels wrong, so it’s hard to admit it’s better.3 There are well-known psychological biases toward keeping what we have instead of switching. But I’d go further and say those biases are even good.

Consider another scenario: You’re walking through a street market and see a game of three-card monte: one card is an ace, for which you’ll win $10; the other two cards are jokers, for which you’ll lose $5. You lose track of which card is which, so you choose one randomly. Then the host says, here, I’ll show you one of the cards you didn’t pick is a joker. Do you want to switch to the other card?

From a logic perspective, this is exactly the same as the Monty Hall problem. And yet the answer is the opposite: never ever switch!

That’s because you can also predict how the three-card monte host will act in each scenario, and it’s different: whereas host Monty didn’t have a stake in whether you won or lost and just wanted to make exciting television, host-monte only wants you to lose. So if you picked a losing card, he’s not going to offer to switch; he’ll take the money and run. If he’s offering you a chance to switch, it’s because you’re holding the ace.

Humans have spent a lot more time making zero-sum transactions than appearing on game shows, so it’s no surprise that our instincts are better suited to the former situation—and that the Monty Hall problem continues to flummox people.

I don't mean "on the verge of bankruptcy" the way you're used to hearing it from talking heads on CNBC, meaning that a government spent a dollar somewhere. I mean, the city didn't have funds to cover a payment due that day, talks with creditors had fallen through, and default was widely reported as a fait accompli, until a last-minute deal saved the day.

That seems like something I should have known, but I didn't until coming across Kim Phillips-Fein's Fear City: New York's Fiscal Crisis and the Rise of Austerity Politics. Though it's entirely about events in the 1970s, it's perhaps the most important book I've read for understanding the politics of the Obama era.

For one thing, as someone who came of political age in the early 2010s, I thought everything must be new and unprecedented. But bailouts, austerity, free college, disparate treatment of minority neighborhoods, private business leaders pulling strings behind the scenes, and even young protestors "occupying" public buildings were hot-button issues in New York's fiscal crisis a generation earlier.

More importantly, I can better understand the mindset of policymakers who had lived through debt crises and didn't want to repeat them. For all the similarities, the post-Great-Recession era had a few big differences that required new solutions, yet politicians still chased old ghosts of the 1970s. And now that we've moved into a new decade with new challenges, we're in danger of winning the last war again.

No money, mo' problems

How did New York get to the brink of default? The city quadrupled its budget in the 1960s, supported by grants from Lyndon Johnson's Great Society initiatives. But by the 1970s, city revenue was falling short: the Nixon administration reduced federal support; two recessions and high inflation hurt the economy, and upper-middle-income taxpayers were fleeing to the suburbs.

State rules prohibited raising taxes (and it would have accelerated suburban flight anyway), but the new benefits were popular, so NYC borrowed money to cover its budget gap. It expected banks to keep lending money out of civic obligation—as you may have heard, New York has a few banks—but that relationship frayed by the mid-70s: the poor economy made banks more profit-sensitive, and nascent globalization gave them more investment opportunities and fewer community ties.

(The first domino to fall is not historically important, but it's a fun butterfly-effect story: early in 1975, New York did a routine debt sale, but the leading bank's head of municipal bond purchasing got sick and someone from another department stepped in. He ran the deal by lawyers—standard practice on most deals but apparently not in munis—who found something they didn't like and scuttled the deal. One person's illness didn't cause New York's near-bankruptcy, but it probably accelerated it.)

The city finally cut spending in 1975: it laid off tens of thousands of workers, not only in welfare services but in sanitation, firefighting, and police. (After the latter, laid-off policemen took over the Brooklyn Bridge for hours.) It created a public-private consortium that had power to make more budget cuts directly, but they weren't enough to balance the books. And it asked Washington to lend it money, but the now-Ford administration refused because a "bailout" would set a bad precedent—leading to a famous Daily News headline: "Ford to City: Drop Dead".

Debt issuances became increasingly fraught until one day in October, when the New York teachers' union surprisingly voted down its share of a purchase that the city needed to cover a payment due at midnight. A bankruptcy press release was written and news stations reported that a default was imminent, until a late deal got the union to reconsider.

How did New York come back from that brink? According to Phillips-Fein, the conventional narrative is "all sides came together to make sacrifices"—but in reality, she says, neoliberalism won and the New Deal ideology lost. The city gave even more control of its budget to an oversight board led by business leaders, laid off 70,000 more workers, defunded CUNY's free tuition program, and gave companies tax incentives to come back from the suburbs. Regardless of which narrative you believe, the common thrust is that too much debt caused a crisis, and solving it required sacrifices.

Macroeconomics peaked when you were a teenager

Fast-forward to another crisis: the Great Recession in 2007-09. It was really bad: unemployment rose and consumer spending fell by the most in 70 years.

In a recession, the federal government can help boost the economy by spending more and/or lowering taxes, and the US did some of that early on. But if you keep doing that indefinitely, you have to issue more and more debt, which might lead to default like New York almost experienced. (You also risk inflation, which we'll come back to later.) So the political discussion in the 2010s quickly moved on to "austerity": how to balance the budget. This wasn't just Tea Party whining; the Obama administration also wanted to reduce the deficit. They just disagreed about whether spending cuts or tax increases were the way to do so.

Even after the worst of the recession, the economic recovery was notably slow compared to prior recessions, in part due to austerity. If austerity had actually prevented a debt or inflation crisis, that would have been a fine trade-off. But a crisis was nowhere in sight—not only in retrospect, but for indicators available at the time:

• People worried about runaway inflation, but actual inflation was persistently below where it's supposed to be.

• And people worried about a debt crisis, but interest rates were historically low, meaning investors wanted to lend the US government even more money than it was taking.

Policymakers thought those problems were coming, but they were wrong. And that's probably due to the shadow of the 70s.

Thanks to the Boomer Blockade, most financial and political leaders of the Great Recession era began their professional careers in the 1970s. That's a time when US economy was truly imperiled by high inflation (peaking at more than 10% per year), and when the NYC debt crisis and others like it were in the news regularly.

Survey evidence indicates that your inflation expectations and investing behavior are shaped by what happened when you entered adulthood. You might think that the world's leading economists are immune from this phenomenon, but it turns out there's evidence for it among Federal Reserve leaders too. So it's natural, if unfortunate, that policymakers around the world were too scared of debt and inflation crises to stimulate the economy appropriately.

The dawning of the age of Millennials

Today there's a new wave of economic policymakers in power, partly due to the last generation's failures and partly due to demographics. Too young to remember the 70s, today's leaders were instead shaped by the Great Recession experience: deficits are overrated, and more spending is better. You can see this kind of thinking both on the left (Green New Deal/Build Back Better-type programs come with huge price tags) and right (Trump's proposed tax cuts far outweigh his spending cuts).

This worked out great in 2020, when an even steeper economic crisis hit with COVID-19; the US government spent more on stimulus than any other major country and had the fastest recovery.

But today, and seemingly for the near future, the economic environment is different: interest rates are high, and inflation is decreasing but still higher than it should be. That's a world in which austerity would actually be helpful, and yet there seems to be little appetite for it.

To be clear, a NYC-level crisis is absolutely nowhere in sight for the federal government, and it almost certainly will never be. But I do think the culture of worshipping stimulus and ignoring deficits is likely to be entrenched going forward. If the environment has in fact changed, and different solutions are needed, policymakers will probably be behind the curve once again.

Don't believe me? Look at last week's doubles tennis event. Spain's Carlos Alcaraz (arguably the best player in the world right now) partnered with Rafael Nadal (inarguably the best clay-court player of all-time) to form one of the most talented doubles teams ever. But doubles play is a little different than singles, and after winning a couple matches, they lost in straight sets to a pair of US players who never ranked in the top 50 as singles players but reached #1 in doubles.

Likewise, dropping the best random basketball players into 3x3 competition with a couple weeks of practice would be extremely fun, but I don't think it would be the cakewalk people are expecting. (If we convinced NBA-level players to specialize in 3x3 for a while instead, that would be a different story, but that's not happening.)

**

I try not to hate on all the weird new Olympic events, because most of the old events are also weird. But it's a little hard to take the skateboarding seriously, especially the women's events, because all of the winners are 15 years old, and can 15-year-olds really be the best in the world at anything meaningful?

The best argument I've seen is that there is one thing 15-year-olds are actually the best at and helps a lot for skateboarding: not caring if you hurt yourself. Still, if that was all that it took, I'd expect 15-year-olds to dominate the trampoline and maybe a few other events.

**

The US set a world record in the 4x100 mixed medley relay, avenging a fifth-place finish in Tokyo that was caused by one of the worst math mistakes I’ve ever seen in sports.

Two factors play important roles in a mixed medley relay: 1) the best men swim about 10% faster than the best women, and 2) the freestyle is much faster than the butterfly and backstroke, which in turn are much faster than the breaststroke. This means the absolute gap is largest in the slowest event: the men’s world record in the breaststroke is 7.5 seconds faster than the women’s, compared to a gap of only five seconds in the freestyle. So if you have a reasonably deep bench of swimmers, it’s much faster to have a man swim the breaststroke and a woman swim the freestyle than the reverse. (The butterfly and backstroke are pretty close in speed, and you need two men and two women, so the choice there isn’t so clear.)

Tokyo was the first time this event was held, and the US somehow didn’t realize this, sending out its world-champion women’s breaststroker and men’s freestyler; they lost seven seconds on the field in the former and made up only five in the latter, the difference between silver and fifth place. This time, they arranged their lineup optimally—as every other team did in 2020 and 2024—and they won.

It is sort of silly to have a medley relay where math dictates what the men or women should swim; the mixed freestyle relay is much more fun, and if they want to do more of it I’d just add an all-butterfly or all-breaststroke mixed relay.

**

If you disagreed with my hot take that the freestyle is the only good swimming event, I hope you enjoyed watching the 200 IM last night, when American Alex Walsh lost a bronze medal by being disqualified for turning over a fraction of a second too early. It was a correct call, but it also sucked to watch (and probably to compete in).

**

More fun things I've seen or learned:

• I'm hardly the first to say it but Peacock's Gold Zone coverage has been terrific -- it's pretty much all I watch during the day now, giving good exposure across all sports and checking in at the most exciting times.

• If you needed another reason to hate racewalking, it started as a way for nobles to bet on which of their footservants walked the fastest.

• Every toddler’s dream: studying dinosaurs and winning gold for trampolining.

• And filming the trampoline is mesmerizing.

• The lifecycle of 1990s rappers continues to be weird.

• Don't miss the Italian parmesan gymnast or the Norwegian muffin fiend.

• "Wacky inflatable arm guys" have been around for less than three decades; they were invented for the Atlanta Olympics in 1996.

And then we turn on NBC in prime time and root for Americans in the backstroke, butterfly, or breaststroke—which are every bit as dumb.

You can technically swim however you want in the freestyle, but everyone uses the front crawl, because other strokes are much slower. In the breaststroke, for instance, the gold medalist in Tokyo swam 100 meters in 57.4 seconds. In the freestyle, that’s about the same as the US record for 10-year-olds.1

And while the backstroke is at least a natural way for humans to swim, the butterfly and breaststroke are not. These disciplines have strict rules about what makes the stroke legal, because swimming in a different (faster) way would be cheating. Olympic swimmers are good enough that you can forget about how weird these motions are, but go watch a swim meet for eight-year-olds—half of them get DQ’d in every race. We shouldn't put up with this any more than we put up with racewalking.

Of course, simply getting rid of the other three events would be a non-starter: NBC would lose that sweet prime-time swimming inventory, and America would lose medals from events we do well in. But don’t fret! Borrowing from track and field, we can invent new competitions that showcase swimming in better ways:

• Underwater hurdles: Set up buoys every 10 meters, so swimmers have to dive under them and resurface after each one.

• Inverted steeplechase: There's only one buoy in the pool, but it’s really deep, and there’s also a land bridge halfway through that swimmers have to flop over.

• Aquatic decathlon: The same people compete in several different lengths of freestyle races, diving competitions off the springboard and platform, and maybe even a game of water polo while we're at it.

All of these events would also be a little bit weird; like on the track, they wouldn’t be as popular as the pure sprints. But at least they’d reward whoever can swim through the course the fastest, not who can do it in an artificially slower way.