Links for April

Railways, Han Kang, Bach

Here are the most interesting things I read, listened to, and watched so far this month. But first, some announcements:

1. If you are aged between 18 and 22, apply to Invisible College, a weeklong residential seminar in Cambridge hosted by Works in Progress between August 17th and 22nd. I will be speaking at it. I haven’t decided what I’ll talk about yet, so if you have any strong feelings on that, let me know.

2. Apply to the progress conference, from October 8th to 11th, in Berkeley, California.

3. As an alumnus, I also recommend applying to the Roots of Progress blog-building fellowship before June 1st.

4. The Fitzwilliam AI Circle has now launched. If you’re technical and can make it to regular meetups in Dublin, please email sam [at] thefitzwilliam [dot] com with a short bio to join.

5. Those of you in Berlin on the 19th of May should attend the Works in Progress meetup, on the topic of European transport and housing policy. In planning your journey, recall that German trains are now less punctual than British ones.

Blogs and short links

We have a new Director of Communications at Progress Ireland, who will help expand our work to new levels, and her name is literally Scales. Nominative determinism remains undefeated!

I spent last weekend in Tennessee, which I enjoyed a great deal, and to which I will provisionally assign ‘underrated’ status. Here is Tyler Cowen on his favourite things Tennessee. I drank some lovely whiskey from Shelbyville, which I learned is a real place, and not just a fictional town from The Simpsons.

Mark your calendars for July 7th for a conversation with Henry Oliver at Samuel Johnson’s House in London. Henry also hosted his terrific book launch there, where I got a signed copy to gift to Julian Gough, my favourite late bloomer.

Rest in peace to Robert Skidelsky. Alas, I only read the abridged version of his three volumes on Keynes, which are difficult to find in paperback. Someone ought to do a reprint. Here’s a passage from Skidelsky’s Money and Government (2018):

Ronald Coase recalled the English economist Ely Devons saying to him, ‘If economists wished to study the horse, they wouldn’t go and look at horses. They’d sit in their studies and say to themselves, “What should I do if I were a horse” And they would soon discover that they would maximise their utilities.’ This joke gives a profound insight into the economic method.

I found that quote when reading about what Rebecca Lowe has been reading. The first piece in that roundup is this essay by T.S. Eliot about literary criticism. She read it for a mini-conference on Eliot in St. Louis, Missouri, hosted by Sebastian Garren and inspired (I’m proud to say) by the Fitzwilliam Milton Friedman conference. I’m not a frequent user of emoji, but: 🥹

Shruti Rajagopalan on Delhi’s recent AI Summit. Shruti’s Substack and podcast are indispensable.

What are the Irish good for?

What Sam Altman believes. I’ve noticed security getting noticeably stricter at tech-adjacent events I’ve been to in the last six months or so, which makes me terrified that we’re entering an era in which the taboo on violence against public figures is gone. Does anybody really understand the terrorism decline of the 1990s, and how durable we should expect it to be?

A disused alley in San Francisco was accidentally bought by a hapless couple who thought they were bidding on the adjacent house, and thus overbid by $25,000. Riley Walz, the Mozart of internet pranksterism, and his pals have now bought that street. They can’t legally do almost anything with it other than paint on it. In order to decide what would be painted, anybody on the internet could submit a design, and public voting determined what would be included in the resulting “quilt”. Voting has now concluded, and they will start painting.

Nobody seems to be pointing out the primary lesson of this tale, which is that this whole mess would have been avoided with the proper application of auction theory. Sometimes I doubt whether Californian politicians have even read Hal Varian’s Intermediate Microeconomics.

We are getting more Leibniz.

Rest in peace to Asimov Press. The editors kindly sent me a copy of their book written in DNA, but I still haven’t read it, because I don’t own a DNA sequencer.¹ If anyone can help with this, please email me!

David Friedman responds to my essay about his dad. Every source I looked at (biographies, Wikipedia, The New Yorker) said that Milton Friedman was 5 feet nothing. David says his dad was 5’3”, and I am in no place to argue. What’s going on here? There are so many photos of Milton next to other people of known height, and three inches is not small. This seems like a big error to me. Also, for people who said I was being too mean in my original post, I only mentioned Milton Friedman’s height because it genuinely seems like he got a kick out of being short. I respect all short kings.

George Stigler, Milton Friedman, and John Kenneth Galbraith. I’ve seen this picture a few times, and I’m bemused that I’ve never seen anyone suggest that it’s a hoax. Why does Galbraith look like he’s superimposed? Why does his foot look so weird when the people next to him are standing still?

What the Advanced Research and Invention Agency has been up to.

Gavin Leech on the state of Chinese AI. He concluded that they are a year behind the West, but the post was in November 2025, and he says it’s outdated by now. Still worth reading for (among things) discussion of how China ‘distils’ knowledge about frontier models from the open internet.

Tragedarianism, the political philosophy.

Most climate communication is innumerate, edition #4723.

Nicholas Decker makes the case that Steven Berry, Jerry Hausman, and Ariel Pakes should have won the Nobel Prize in economics. Read also for some intuition for the Hausman test between fixed and random effects estimation in panel data, which went over my head as an undergraduate.

The most disturbing thing I read this month was what Peter McLaughlin read and thought about in February. In certain communities in Nigeria, one in six men and one in eight women support infanticide for children born after a set of twins.² That makes it by far the most “popular” form of blood sacrifice of children in Sub-Saharan Africa. It’s also much more popular among women than among men, which is the opposite of the norm. Infanticide of twins is a well-known practice about which there is an enormous literature. But neither he, I, nor any LLM, can find a single reference to baby-after-twin infanticide, other than an extremely racist anthropology paper from 1928 about a different ethnic group than the one surveyed. You can definitely see why, under certain forms of folk religion, twins might be thought of as “cursed”, but what is the origin of believing that being born next after twins would also be unlucky? Do mothers who have had twins know that, by having another child, they are painting a target on their back? What the fuck???

If you can find anything more about this, please email me.

On a lighter note, Peter also writes in the post about:

• The great curling controversy of 2026.

• Why all Egyptian art looked the same for thousands of years, then started looking like weird aliens for a few years under the pharaoh Akhenaten, and then went back to looking the same again for thousands of years.

• Why ‘poverty’ is not a natural category, but subsistence farming is.

• Why belief in witches is nearly a human cultural universal.

• How African megafauna managed to avoid extinction by evolving a fear of human speech.

• The necessity of repetition for appreciating poetry.

• An argument that quantum field theory is the phlogiston theory of the modern day.³

Stephen Malina’s links. My favourite entry here is Afra Wang’s on the story of a Beijing vibe coder. The fact that this quote comes from someone named Liu Xiaopai is a point in favour of Dan Wang’s “no two peoples are more alike” thesis about Americans and Chinese:

My past two decades of building software have been profoundly inspired by Paul Graham’s Hackers & Painters. The core thesis—PG argues that the software business is the world’s best business because its marginal cost approaches zero. And software creators, like painters, are engaged in creation, not construction labor. These concepts have deeply influenced me. Additionally, I enjoy Tools of Titans and Tribe of Mentors by Tim Ferriss.

From Justin Skycak, a vastly underrated predictor of extreme subsequent success is a willingness to be low status.

Ruxandra Teslo, Saloni Dattani, Witold Więcek, and others have started a new group blog to promote improvements in clinical trials. Here is Adam Kroetsch on the recent guidelines on the use of Bayesian statistics released by the FDA.

The FDA goes Bayesian.

An archive of all the books mentioned on the Conversations with Tyler podcast.

I am “retired” from the world of competitive trivia, but if you want to get better at it for some reason, you can start by grinding Protobowl.

Riley Waltz on the smallness of the world.

I commend the efforts of Celine Nguyen to expand the market for culture. Here she is on everything she read in January. This is a year in which much of my friend group is also Chinesemaxxxing.⁴

March update from the Centre for British Progress.

Bookbear Express update.

Pre-order the new textbook on Bayesian workflows from legendary statistician Andrew Gelman.

Via Henry Oliver, Gelman is also a remarkably sophisticated literary critic. On Andrew’s blog Statistical Modeling, Causal Inference, and Social Science, guest blogger Witold Więcek also asks about what major works of literature have been written at advanced ages. (I’m told these gentlemen sometimes talk about statistics.)

Update from the world of neuroscience, especially connectome mapping. Read for some background about why inhibitory neurons are more diverse in their structure than excitatory ones.

From Andy Coravos: In the world of the machines, the human is the conductor. Read for lots of examples about how AI agents can save you money, especially in overcoming healthcare bureaucracy.

Rationalist Garfield, a comic strip procedurally generated from LessWrong.

For those of you whose issue with these posts is that monthly isn’t frequent enough, you can check out this Substack of daily links.

Congratulations to the new cohort of Emergent Ventures winners.

After Japan’s surrender in World War II, General Douglas MacArthur was given essentially dictator-level power in directing its reconstruction. However, he didn’t speak Japanese. There were apparently only 200 American civilians at the time who knew how to speak Japanese,⁵ which meant that people who coincidentally spoke both English and Japanese were suddenly thrust into positions of remarkable power. For example, a 22-year-old child of Russian Jews who happened to speak Japanese is the reason why marriage law is so egalitarian in the Japanese constitution.⁶

It’s a shame that the Tommy Lee Jones film about MacArthur is so poorly reviewed, because he’s the perfect casting. There could have been an incredible Tarantino-style alternate history movie about his life in which he later became president and nuked China.

Lauren Gilbert’s links. Read for some context on why the Anglosphere is so much better at assimilating immigrants than anywhere else.

Katja Grace on AI as a Trojan horse race.

Klara Feenstra’s best novels of February. Relevant to my discussion two months ago about Tom Stoppard and reading versus watching plays:

I recently edited a screenplay and found it truly excruciating. No matter how good a script, I never enjoy reading it because it’s an act of translation but without the joys of a second language. My brain gets exhausted. I can’t appreciate Arcadia because I’m traumatised by doing years of drama school where, if a piece of dialogue is set beside a stage direction, it immediately takes on a stilted quality. That same dialogue in direct speech within a novel wouldn’t bother me at all.

Podcasts

Tyler Cowen and Alex Tabarrok on the history of options pricing. I am struck by the history of polymath (usually French) economists whose work was so advanced that (usually Anglo) economists didn’t have the mathematical background to understand it for decades or more. Augustin Cournot was applying the concept of Nash equilibrium in the 1830s. Louis Bachelier derived the equations for Brownian motion five years before Einstein independently rediscovered them, and had worked out the core conceptual framework of Black-Scholes theory over seventy years earlier.⁷

Andrej Karpathy on the case that AGI is still a decade away. I enjoyed this most for the unusually thoughtful discussion about what education is going to look like in a few decades. As Tyler wrote on Marginal Revolution, like it or not, this is where the serious and important philosophy is being done today.

Henry Oliver on Shakespeare’s Measure for Measure, and other literary matters.

Joe Studwell on African economic development. I still haven’t read How Africa Works, but I look forward to doing so; perhaps we’ll do an African economic development edition of the reading group.

James Philips, one of Dominic Cummings’s “weirdos and misfits”, on the British government’s COVID response and the creation of ARIA. Every time I hear an Oxbridge-educated Brit who appears incomparably more culturally sophisticated than anyone in Irish politics, I can’t help but think about this quote:

People in the EU are super wise. You have a meal with some sort of French person who works in Brussels—it’s very impressive. They’re cultured, they have wonderful taste, they understand all these different countries, they know something about Chinese porcelain. And if you lived in a world ruled by them, the growth rate would be negative 1%.

99% Invisible on the “Ford to City: Drop Dead” fiasco. This is a sort of coda to their great series breaking down The Power Broker. Like everyone else, I’ve been reading The Power Broker on and off for years, although I almost finished it while I was in New York last week.

Rebecca Lowe and Tyler Cowen on the definition of freedom. This podcast has a fun format: the host and guest discuss a philosophically laden word, for the purposes of reaching a ‘working definition’ of it. If I were ever on such a podcast, ‘alignment’ or ‘causality’ would be fun topics.

Allan Dafoe on Google DeepMind’s governance work and philosophy. Dafoe’s team recently closed a job listing which looked cool.

Music

Johann Sebastian Bach, St Matthew’s Passion. A Passion, if you’re not aware, is a choral retelling of the story of Christ’s arrest, trial, crucifixion, and burial. I heard this one live at a recent performance in St Patrick’s Cathedral. I’m told the John Eliot Gardiner recording with the Monteverdi Choir is canonical, which is what I listened to in preparation. I was most moved by the Judas character and his significance. It’s also interesting in itself that the singing is in German; Catholics would be less likely to use the vernacular. I’ve long struggled to get into baroque music, and, logistically, I’m not sure how best to approach live choral music. I do ultimately have a monkey brain that will get distracted if I have the lyrics up on my phone, but then, should I print them off in a booklet instead? For Bach, I’ve mostly been learning from the Year of Bach podcast and Substack. Next up will be St John’s Passion.

Talking Heads, Speaking in Tongues. Moon Rocks is my favourite track here, while This Must Be the Place is less interesting to me. Interestingly, Brian Eno was not the producer on this album after having been it for the previous three. I gather there was some drama because he was leaving an increasingly large footprint.

The Clash, Combat Rock. There is wisdom to the critical consensus that this album is weaker than either London Calling or The Clash. Rock the Casbah is mighty, and I suppose now is as fitting a time as any for satires of decrees by Ayatollah Khomeini. I also learned that Straight to Hell is the Vietnam protest song sampled on Paper Planes, very cool.

Mingus Big Band, The Charles Mingus Centennial Session. I love Charles Mingus; the Mingus Big Band is an ensemble started by his widow⁸ over a decade after he died. The opening performance of Work Song (Break the Chains) is very strong, which, confusingly, is completely separate from the Nat Adderley jazz standard. You may be interested to know that the trombonist on this album is internet public intellectual Coleman Hughes.⁹ What’s the correct lesson to infer from the fact that successful people were so frequently highly successful in some unrelated area?

Papers

Alec Radford et al., Language Models are Unsupervised Multitask Learners. This is the paper that introduced GPT-2 all the way back in 2019, which to some of my friends means I might as well be reading the Epic of Gilgamesh. This showed that language models trained on sufficiently diverse corpora would implicitly learn language skills without supervised fine-tuning on each individual task. It also showed a monotonic relationship between loss and number of parameters, a precursor to the scaling laws. Like seemingly everything else in ML, I don’t think there was all that great a reason for anyone to expect this was going to work. But it did!

Some of the papers from this era about what’s going on inside transformer components are easier to read today than they must have been at the time. More useful abstractions have been developed for thinking about information flows through neural networks, e.g. Anthropic’s conceptual framing of the ‘residual stream’ as being the primary channel of information flow through a transformer.

I’ve been making spaced repetition flashcards like a madman for important ML concepts, like the learning rate, the Adam algorithm, and perplexity. What an incredible time to be learning new things. :)

Tom Brown et al., Language Models are Few-Shot Learners. This is the paper that introduced GPT-3 in 2020. When I was reading this, I was thinking about the poverty of the stimulus, the combinatorial argument from Noam Chomsky about how children aren’t exposed to enough instances of words to acquire the features of their language unless they are hard-wired for language-specific biases. There’s some good stuff in Appendix A about the clever ways that ChatGPT improves upon the rubbish data quality that it scrapes from the internet. And Appendix E has details of a quiz they ran in which people were not reliably able to tell apart AI-generated and human-generated news articles. It’s interesting that this was covered in the original paper, and I remain confused as to why misinformation and deepfakes haven’t been more of an issue so far.

I chuckled at the use of a Student’s t-test on page 26.¹⁰ The Student’s t-test assumes that the variances of the two samples are equal, which is almost certainly false in this case. A “baby’s first hypothesis test” like this wouldn’t fly in other fields, but the authors know this; the effect size is so gigantic that it doesn’t matter. Another way of looking at it is that computer scientists don’t need fancy statistical methods, because the effects they are studying are actually real.

Jonathan Roth et al., What’s Trending in Difference-in-Differences? A Synthesis of the Recent Econometrics Literature. I read these papers a few months ago in preparation for Jeffrey Wooldridge and Pedro Sant’anna’s summer school on causal inference in Greece in June. It slipped my mind to include them in the links, even though they passed the quality bar. I am still writing up my notes from that event, which are too technical to be covered here to any meaningful extent. This paper is a helpful overview from 2023 of what’s happened in the literature on differences-in-differences since it fell into crisis around 2020. This post from Beatriz Gietner does a good job explaining the background about that.

Clément de Chaisemartin, Xavier d’Haultfoeuill, Two-Way Fixed Effects and Differences-in-Differences with Heterogeneous Treatment Effects: A Survey. If you’re going to read one review paper about developments in the literature on differences-in-differences econometrics since 2020, it should be the Roth paper above. But if you’re going to read two papers about that, this should be the second one.

A senior member of the economics profession who shall remain nameless once told me that “everyone in econometrics is friendly, except for the French”. The way a lot of these things go is that someone identifies a problem, like recovering an average treatment effect for a policy that was rolled out in stages, and then a new “heterogeneity-robust” estimator is developed to deal with that. The authors of this review paper created one such estimator in 2020. Claude Opus 4.7 was quite snarky about the ways in which the authors present themselves as the protagonists in their own literature review.

I recently heard from another economist who admitted to using the Callaway-Sant’anna estimator in his research instead of the Chaisemartin-d’Haultfoeuil one “because it’s easier to pronounce”. I still can’t really follow what makes the Chaisemartin-d’Haultfoeuil estimator different from Callaway-Sant’anna and Sun-Abraham, or the imputation methods I don’t know anything about. This (gated) explainer by Scott Cunningham would be a good place to go next.

Guido Imbens, Jeffrey Wooldridge, Recent Developments in the Econometrics of Program Evaluation. This is a review paper from 2009, so it’s from too long ago to cover the spicy differences-in-differences controversies. This is quite a practical guide that covers econometric issues that come up when doing empirical work, but it also has no GitHub, code, or implementation. This is in contrast with the papers above that have full implementations. So I’m not sure if there is still a reason for papers like this to exist, which is not a criticism of the authors, so much as it is a comment about the skeumorphic quality of consuming research in this format. Beatriz calls Jeffrey “Wooldridge the Great”, and he is particularly thoughtful. The other author is Guido Imbens, who won the Nobel Prize in 2021 for work related to what’s in this paper.¹¹

Bryan Graham, et al., Inverse Probability Tilting for Moment Condition Models with Missing Data. Inverse probability weighting (IPW) is a classic way in statistics of responding to the problem of data being missing at random, in which you model each observation’s probability of being observed (propensity score), and then weight the observed cases by those scores. The term ‘missing at random’ (MAR) is incredibly confusing, and means almost the opposite of what you might expect it to from everyday speech. Missing at random means that an observation is missing with a probability that depends on the observed variables in an unknown way you have to estimate. What you probably think of when you hear the term is there being a fixed probability p of data not being observed in a particular case, which is called ‘missing completely at random’ (MCAR). The classic IPW setup has some pretty technical issues about the bound on variance of the estimator as the sample size goes to infinity, and then also it doesn’t have this property called double robustness. The authors suggest an alternative called inverse probability tilting, where you reweight the complete cases in a sample to be perfectly representative of the population as a whole.

They test their model on a simple, uncontroversial, dataset: explaining Black-White gaps in wages using IQ tests.¹² I’m not kidding. The only empirical part of the paper involves forming a much more precise estimate of ethnic differences in cognitive ability than what previously existed. The context is that they are responding to one of the most famous and contested papers in labour economics, which is about how much of the racial difference in earnings is explainable by characteristics acquired before entering the labour market. But still, a striking choice!

I am not sure that anybody cares about this other than professional econometricians. At the time, reading the other causal inference papers felt equally niche, but they actually subsequently came up in my life quite a lot. This one still feels niche to me.

Jared Kaplan et al., Scaling Laws for Neural Language Models. This is the famous ‘scaling laws’ paper from 2020, showing that test loss scales as power laws of parameters, dataset size, and compute.¹³ Maybe it’s just that I’m a hammer to whom everything looks like a nail, but this really reads like an economics paper. The authors take some guesses at the functional form of a relationship, and then calculate certain optimal ratios assuming perfect rationality, which have now been widely adopted by profit-maximising firms. It’s just like macroeconomics!

And, like macro, we have no idea what’s actually going on. There is some speculation in this paper about whether the scaling laws are explained by a mysterious “data manifold”, but as Dario Amodei points out, nobody knows why scaling works. I’ll quote Peter McLaughlin’s review¹⁴ of The Scaling Era by Dwarkesh Patel:

My main takeaway is how little any of these people know what the fuck they’re doing. I don’t blame Patel or [Gavin] Leech for this; indeed, I’m really glad that Patel knows enough about the field, and is clever enough, to ask the questions that reveal the real limits. He opens the very first chapter with the most obvious thing that any intelligent person would want to know about the field: given that we are in the ‘scaling era’ of artificial intelligence, why does ‘scaling’ lead to better artificial intelligence? The best answer comes from one interviewee who posits some kind of totally unknowable ‘data manifold’ made up of who-knows-what components that might answer the question if it even exists; everyone else is reduced to mystic burbling. If ‘knowing what the fuck you’re doing’ were a prerequisite to ‘changing the world’, we could just write the whole field off; alas, it is not, and we cannot.

I mentioned to a friend that I’d love to see a more literary essay about the lessons of the scaling era, inspired by J.B.S Haldane’s On Being the Right Size. Does it exist?

Jordan Hoffmann et al., Training Compute-Optimal Large Language Models. This was the third of the trio of papers I selected on scaling for the Fitzwilliam AI circle; it’s the report from 2022 introducing the Chinchilla family of models. Contra Kaplan et al., this argued that model size and training tokens should scale roughly equally.¹⁵ This meant that pre-2022 models were substantially undertrained relative to their number of parameters. The most readable explainer of this I’ve come across is by Trevor Chow, whose post about this I previously made fun of on my blog. Trevor and his enormous brain, by the way, now work with Daniel Gross on the compute team at Meta Superintelligence Labs.

This paper was also part of the transition toward people using the term “inference” in an extremely confusing way, contrary to centuries of established usage in statistics.

Dave Donaldson, Railroads of the Raj: Estimating the Impact of Transportation Infrastructure. Genuinely one of the most incredible economics papers of our time; here is Kevin Bryan on why Dave Donaldson’s John Bates Clark Medal was well deserved.¹⁶ This also made it to Nicholas Decker’s list of 11 favourite economics papers. Donaldson wrote this paper as part of his PhD, which took him eight years. The main reason it took him so long is that it involved manually (!) digitising 1.5 million (!!) data points from (largely handwritten) colonial archives (!!!). From that data, he was able to identify transportation costs in British India by looking at local variations in salt prices whose origin is known. Salt is an incredibly homogeneous good; if you know how much a particular type cost in Madras, and how much it cost in Mumbai, then you also know how much it cost to transport. From this, Donaldson is able to estimate a parameter θ, the elasticity of the volume of trade with respect to price. By combining this information with an off-the-shelf general equilibrium model of trade, he’s is able to get a precise estimate for the welfare effect of railroad expansion under the Raj.

This is extremely clever, but the issue is that railroad placement is still endogenous: it could be that railways were only built in growing areas, which would have done well anyway. What would be more convincing is if we could show that sites that were equally good candidates for railroads, that were never built for idiosyncratic reasons, did not share in the benefits. This is how Donaldson uses the Kennedy plan (1848) and the Lawrence plan (1869) for railroad expansion in India.¹⁷ This is an instance of a placebo test, a confusingly named piece of causal inference apparatus that has little to do with the term’s use in psychology. This is probably the least convincing part of the paper, because it could always be that the planned lines were cancelled for some other reason, which would also have caused those regions to have lower incomes subsequently.

Dave Donaldson, Richard Hornbeck, Railroads and American Economic Growth: A “Market Access” Approach. This paper is largely a revision of Robert Fogel’s classic 1964 study of railroads in American economic history, which argued that they were not indispensable for growth in the 19th century. He used a methodology called “social savings”, in which you take the quantity of good transported by a new technology, and then calculate what it would have cost to transport with the next-best available technology. Fogel argued there was a surprisingly high degree of substitutability between railroads and canals and other forms of transport. He concluded that the absence of railroads in 1890 would have reduced GNP by 2.7%, far below what most economists at the time would have expected.

For this paper, Donaldson and Hornbeck pioneered a new methodology called market access, where you directly model how improved transportation between cities A and B affects city C. For the computer scientists among you, this is based on a nice application of Dijkstra’s algorithm and graph theory. The market access methodology is used to combat violations of the Stable Unit Treatment Value Assumption (SUTVA). SUTVA says that there is no “spillover” from the treated units to the untreated units, which is certainly not true of the knock-on consequences of opening railroads.

Of course, railroad placement is still highly endogenous, and they have various ways of responding to this, including using proximity to waterways as an instrumental variable.

When you combine this, like the Raj paper does, with modern trade theory, you get a much better understanding of how the railroads affected economic geography. Contra Fogel, trains were a Big Deal, and the authors estimate that American agricultural land values in 1890 would have been 60% lower were it not for the railroads.

Ronan Lyons, Alan Fernihough, Railways, Market Access, and Development in a Small Open Economy: Evidence from 19th Century Ireland. No link for this, sorry, the paper is still under review. I will save more substantive comments for my essay on Irish railway history, whenever I get around to finishing that. This paper is essentially an application of Donaldson and Hornbeck’s market access methodology to Ireland. Here is a previous version, which, interestingly, reached a completely different conclusion. Earlier, the authors found that the construction of railroads in Ireland increased land values a huge amount because of greater market access, and then decreased them a huge amount because of accelerating emigration, in a way that almost exactly cancelled out. In any social science, having two gigantic effects that cancel out almost perfectly is very fishy, and it turned out that the working paper had a multicollinearity problem. If you have ‘perfect’ multicollinearity, one regressor is a linear function of another, and ordinary least squares no longer produces a uniquely defined answer. But if you just have a partial linear relationship between them, you get a higher sampling variance of what you’re trying to estimate.¹⁸ In the Irish case, this spuriously appeared as large effects in opposite directions. I commend the authors for catching and fixing this.

Regional trade data for the Irish economy in the 19th century doesn’t exist, which means that the authors can’t estimate the trade elasticity θ, and therefore can’t estimate the welfare effects of the Irish railroad expansion. We don’t have good historical estimates of how much Irish people benefited from improved infrastructure, which is unusually relevant to someone working at a think tank aiming to improve Irish infrastructure. You might also ask why colonial India has higher-quality economic data than Ireland. ¯\_(ツ)_/¯

I think that I’ve had enough of railroad economics for quite some time, but if I were to go further, next on the list is Banerjee, Duflo, and Qian, which uses a similar methodology to conclude that railways had almost no effect on economic growth in China. Dan Wang, telephone! You can see my discussion prompts about this trio of railroad papers here.

Books

Tyler Cowen, The Marginal Revolution: Rise and Decline, and the Pending AI Revolution. This is more of a monograph than a book, which you can read for free here. I still haven’t figured out how to use Tyler’s AI integration in a way that is more fruitful than just uploading the EPUB to Claude and asking questions.

The first part of The Marginal Revolution is an intellectual history of the ‘marginal revolution’ of 1870s economics, as led by William Stanley Jevons, Léon Walras, and Carl Menger. The Hollis Robbins review provides a good summary:

Two things happened as statistics developed in the nineteenth century: first, the rise of marginalism, the idea that economic value comes from the effect of one additional unit under fixed conditions (one more loaf of bread, one more hour of labor). Second was the rise of averageism, the idea that aggregates are best described through averages, with innovations such as Adolphe Quetelet’s “average man” and William Stanley Jevons’s calculation of index numbers and average consumption across populations.¹⁹ Tyler argues that marginalism and averageism are a kind of package deal, advancing together as mutually reinforcing methods. Averages became central to measurement, while marginal reasoning helped interpret those measurements. The marginal unit is only legible against a distribution, and the distribution is only interesting because marginal changes can be inferred from it. The coming AI revolution, as Tyler sees it, brings averageism back to the center.

Chapter three is about why, despite being quite obvious today, it took so long for marginalism to develop. He analogises economics to geology and evolutionary biology, as fields that are also conceptually simple but historically slow-moving. He also makes a sociological claim about how, since any professional economist worth their salt understood marginalism, competition was pushed into more technical domains, like the use of econometrics and more careful empirical work.

There’s also some good stuff about the precursors to marginalism, such as Jules Dupuit. Many of the ideas of marginalism were also independently discovered by the Irish polymath Dionysius Lardner in a book about railway economics (!!!) in 1850. I literally cannot believe this never came up at our Irish railway economics event. If any of you would write about this guy for The Fitzwilliam, please email me! This was also an amazing detail (page 43):

In the 1860s Jevons built a Logical Abacus, sometimes called a logical piano, a kind of early computer that could perform (some kinds of) logical operations faster than humans could. It is held in the Museum of the History of Science at Oxford University.

As is often the case with Tyler, it’s hard to tell what he’s actually trying to say in this book. Here’s what I picked up: The ‘marginal revolution’ made the economy feel much more understandable and intuitive than it actually is. Sophisticated economic theory sometimes works for a while, but then it loses its predictive value. The most influential asset pricing model of the 20th century (CAPM) has a key parameter β, the sensitivity of an asset to movements in the market portfolio. We know from a literature culminating in Fama and French (1992) that β doesn’t actually predict anything. And nowadays, the best-performing financial models are machine learning algorithms that pick up on statistical patterns and have almost no actual theory baked into them.

There are two possible reasons for this:

1. CAPM was always wrong.

2. CAPM was an accurate description of the world in the past, but it no longer is.

It’s a little from column A, a little from column B. Artificial intelligence reveals that our prior state of “understanding” was thinner than we thought. But also, contrary to the pessimists, economics is a real discipline that has aided our understanding of at least some topics. It could just be that the world has gotten genuinely less interpretable over time.

Marginalism is partly just a set of mathematical identities about first-order conditions and optimisation. Tyler calls this “tautological marginalism”. “Intuitive marginalism” is the idea that these tautologies are actually important, and that the idea of reasoning “at the margin” is useful for many applications, such as in Jevons paradox. Although humans’ intuitive marginalism might lose its force in explaining the world, Tyler says that it will survive in the form of the weights of future models (page 119):

The training of Large Language Models (and other forms of AI) will enshrine marginalism into their basic operating concepts, as those models are trained on writings that understand marginalist concepts... marginalism will not die, but we will automate it.

The Straussian reading of this monograph is that Tyler is grappling with what it’s like to be an economist in his mid-sixties today, having watched software eat the world, including his life’s work. All my economist friends now wish that they had studied maths or computer science as undergraduates. Economics has had various intellectual edifices come and go, either because the theory was wrong all along, or because it only explained a narrow range of phenomena during a certain time period. That includes schools Tyler has a soft spot for, like Chicago price theory and Mason public choice theory. His melancholy is most evident on page 106:

Sadly, price theory is fading in relevance, and it is taking marginalism down with it. It used to be that some graduate programs favored the axiomatized approach to micro and others (e.g., University of Chicago, UCLA, University of Virginia) favored the price theory approach. These days the axiomatizations have won out pretty much everywhere, except at my own George Mason University.

Finally, some fun tidbits: on whether economics is too mathematical, we have this quote from Shengwu Li;

“Can you give me an economic intuition for that result?” means “Can you explain that using math that was introduced to economics >20 years ago?”

Corollary: Topkis’ theorem counts as economic intuition. Also virtual values. And eigenvalues.

Abortions are surprisingly elastic (page 31):

We find that a hundred-mile increase in distance to the nearest [abortion] clinic is associated with 25 percent fewer abortions.

In the latest edition of “economists have predicted nine out of the last five recessions”, I was unaware that there was a sunspot theory of macroeconomics (page 59):

[William Stanley] Jevons’s obsession with the average also showed up in his macroeconomic work on the “sunspots” theory of commercial cycles. Jevons had presented evidence that business cycles occur at a typical regularity of about ten and a half years. At the same time, astronomers had told him that the solar period had a typical regularity of about ten and a half years. He concluded that the two phenomena were related, but never successfully pinned down the evidence or the case for a causal relationship.

I pronounce William Whewell (p. 103) as Victorian polymath of the month:

Alfred Lord Tennyson called him “a lion-like man.” [William] Whewell was a legitimate polymath, making contributions to mechanics, physics, geology, astronomy, mathematics, economics, and also poetry. He was an Anglican priest, he translated works by Goethe and Grotius, and he organized a pathbreaking citizen science project to study ocean tides. It was he who persuaded Darwin to become secretary of the Geological Society of London, noting […] that geology was a critical input into Darwin’s theory of evolution. On science and its philosophy, Whewell’s major work was his 1837 multi-volume History of the Inductive Sciences, from their Earliest to the Present Times, now largely forgotten outside of academic history of science, but massively influential in his time. Whewell also coined the terms scientist (in 1833), physicist, linguistics, consilience, catastrophism, uniformitarianism, and astigmatism, an impressive list. To Michael Faraday he suggested the terms anode, cathode, and ion.

Han Kang, The Vegetarian. Like many popular Korean cultural products, this novel is incredibly fucked up. I read embarrassingly little fiction. I’ve been trying to get into it more, but I find it difficult; the opportunity cost of not reading non-fiction is incredibly salient. So it was good to come across such a short novel to ease back into it. Like many people in the West, I hadn’t heard of Han Kang until she won the Nobel Prize in Literature in 2024. The Vegetarian also won the Man Booker International Prize. South Korea stay winning.

The Vegetarian is divided into three parts, but I’m not sure I can really give more information without spoiling the plot. I assume there are deep parallels with Kafka’s The Metamorphosis, but I also haven’t read that, so they went over my head. I liked it. I especially enjoyed reading this with my girlfriend.

Yes! Girlfriend! Readers occasionally remark that such a long linkpost could only be the product of a chronically single man, which may be accurate. This is the first time I’ve been in a stable relationship in over two years, so some life adjustments are in order. However, if labour economics is anything to go by, men are up to no good by default, and are improved by relationships on both personal and professional margins. Stella Tsantekidou will now have somewhat less material with which to make fun of me on her blog. Riya is very special. <3

Joe Sacco, Footnotes in Gaza. One of my childhood friends recently gave me a hard time for not writing about Israel on my blog. I told him that I essentially never read the news, I know very little about the Middle East, I’ve never even visited, and I have zero comparative advantage in writing about it. I think all of these are good reasons.

What I do know about the Middle East disproportionately comes from graphic novels. The non-fiction graphic novel remains a criminally underrated format. For the first hundred pages or so, I had this author mixed up with Guy Delisle, a cartoonist who similarly writes travelogues from exotic locations, although I haven’t read his Jerusalem book yet.

Footnotes in Gaza is about a journalist trying to reconstruct the details of the massacres of Palestinian males by the Israeli Defence Forces in the towns of Khan Younis and Rafah in 1956, during the border skirmishes between Egypt and Israel during and immediately after the Suez Crisis.²⁰ Israel’s position seems to be that this was a normal military operation against fedayeen, which created a small number of accidental casualties, but the eyewitness testimony in the book all alleges they were executing civilians.²¹

There are various criticisms of this book, most of which seem pretty weak to me. He conducted the interviews with survivors 50 years after the events; human memory is extremely fallible, but he grapples with this at length in the book. The criticism I am least sympathetic to is that he’s making a strong claim about these episodes being representative of broader conflict in the Middle East. I think it is fine (even preferable) to just write about minor historical incidents because they are interesting, without having them connect with a grand historical narrative. Don’t let boorish hacks kill your curiosity. Sacco doesn’t make any claim to be a completely even-handed source, and he says in the endnotes that he was influenced by a Noam Chomsky book, which is a red flag.

Films

Tomm Moore, Nora Twomey, The Secret of Kells. An absolutely gorgeous animated children’s film about the creation of the Book of Kells. According to legend, the manuscript was started on the island of Iona by St. Colmcille, before being brought to Kells in County Meath in the 9th century. This was the first film in a thematic Celtic mythology trilogy, which I watched because I liked the third one, Wolfwalkers, so much. Eventually, I would like to pen something longer about Irish cinema, which has had many great entries. A strength is that, owing to our sentimentality, internationally acclaimed actors will work (presumably for much less money) on local projects, e.g. Brendan Gleeson voices the Abbot of Kells here.

My film consumption has fallen off a cliff recently, since differing tastes mean that it’s difficult to converge upon a choice with my girlfriend. She reports the dilemma is adequately explained by this TikTok:

film bros when u tell them u want to watch a marvel movie and not a 2 hour black and white movie about the serbian government shown through the eyes of a pigeon

From YouTube, we have the NPR Tiny Desk concerts with Charlie Wilson and David Byrne. We also have the story of JMail, and Luigi Tenco’s 1966 performance of ‘Vedrai Vedrai’.

1. I honestly wouldn’t put it past my flatmate to own a DNA sequencer.

2. Peter reports that this also happened (still happens?) in Kenya.

3. I am unqualified to have an opinion about this, but I would really value the perspective of people knowledgeable in quantum field theory about this argument.

4. Nguyen also talks about a recent biography of Henry Bergson. Bergson is one of those characters whom I long thought was more significant than he really is, because I made the mistake of first learning about the history of Western philosophy from Bertrand Russell’s The History of Western Philosophy. If volume two of the Ray Monk biography is to be believed, he was running up against the deadline and included material from whatever essays he had lying around to get it finished in time. Thus, Kierkegaard is hardly mentioned, but we got a whole chapter on Bergson.

5. This is what Claude told me, but that can’t possibly be right, can it?!

6. She also has a memoir partly about this.

7. One of the reasons why Bachelier didn’t get very far is that his thesis advisor, a fellow by the name of Henri Poincaré, wasn’t very impressed with him. The story goes that Leonard Savage (he of the Savage axioms) sent a postcard raving about Bachalier to Paul Samuelson, who then told Fisher lack and Myron Scholes and Robert Merton. There’s a lot of alpha in sending postcards!

8. Lore: Charles and Sue Mingus’s unofficial marriage ceremony was “officiated” by the poet Allen Ginsberg.

9. If I’ve understood correctly, he plays on tracks 1, 2, 9, and 11.

10. As we know from The Fitzwilliam history of Guinness, the Student’s t-test is another great Irish cultural export.

11. The Stanford YouTube channel made this adorable video of his kids explaining what their dad won a Nobel Prize for.

12. Technically, it’s an AFQT, which I’m told is an IQ test in all but name.

13. I also learned that this paper is largely responsible for the convention that LLM training runs have only a single epoch.

14. This is itself linked to in Gavin Leech’s review of his own book.

15. They recommend around 20 tokens per parameter.

16. The timelines on the publication of top-five economics journal articles have stretched to a point of beggering belief. ‘Railroads of the Raj’ was circulating as a working paper since 2010, but wasn’t published in the American Economic Review until 2018, which was 17 years after the project began. This was a year after Donaldson won the John Bates Clark medal, partly for the strength of this paper.

17. There’s not tons of information about this online. There is a webpage about expansion under the Raj on the Indian government’s railways webpage, but as befits my previous experience with Indian trains, it’s broken.

18. This is the ‘variance inflation factor’ (VIF), which is 1/(1-R_j^2), where R^2 is the fraction of variance explained by all regressors except X_j.

19. There’s some pretty interesting stuff (to me) about Jevons’ work on how to construct price indices, and why he favoured the geometric mean over the arithmetic mean for this purpose. See page 58.

20. The official UN statistics were: 275 dead in Khan Younis, 111 in Rafah.

21. I say “seems to” because Sacco largely got stonewalled by the IDF in writing the book. I’m not sure that there is a developed counter-narrative from Israel of what happened.