Links for February
Afghanistan, rugby, Peter Singer
Here are the most interesting things I read, listened to, and watched so far in February.
This week, I am moving in with my friend Neil, to form the greatest tech double bachelor pad Ireland has ever seen. That means I am vacating a bedroom in a 2-bedroom ground-floor Mountjoy apartment with an extremely kind (female) flatmate. If anyone is looking to live in central Dublin for between two and seven months for somewhat below general market rent and also likes cats, I highly recommend it. Email sam [at] thefitzwilliam [dot] com if you are interested.
Blogs and short links
In the Fitzwilliam Maths Circle, we have been working on topology this month, which led me to wonder: What is Grigori Perelman up to these days?
Matt Lakeman’s notes on Afghanistan. I think this is the best Matt Lakeman travelogue yet. Too much to dissect here, but this quote stuck with me:
The average Afghan probably never saw a single white Western soldier in person throughout the 20 year occupation.
I also found this gem from the comments section:
“[A]re there any Jews left in Afghanistan?” Not any more, the last one left after the Taliban took over in 2021, which is fair enough. Interestingly at one point there were only two Jewish people in Afghanistan. They both lived [by] the [Synagogue] and hated each other. As in they’d rat each other out to the original Taliban regime.
[A] play was made about them, and the below video is them arguing.
Here is a link to said play, although surely the title ‘Two Jews in Kabul’ would have been better?
Deng Xiaoping’s pragmatism, Slack emote edition.
Why the disastrous Jones Act is uniquely bad for Puerto Rico.
What Rob Long read and liked in January.
John Fingleton has written a report for the British government, with input from some of the usual suspects, which recommends completely changing its attitude toward the regulation of nuclear power. The government has accepted the recommendations, and also pledged to extend the lessons to other areas. Can someone help me come up with a well-phrased prediction market about whether this actually be implemented? Has the Irish mafia successfully come to its distressed neighbour in her hour of need? PS John Fingleton has joined Progress Ireland as a Senior Fellow.
Dario Amodei on the adolescence of technology. Here is some commentary from Jasmine Sun and friends about this and also Claude’s Constitution. For any of you who may still think that misalignment is theoretical:
According to [one] line of thinking, we don’t worry about a Roomba or a model airplane going rogue and murdering people because there is nowhere for such impulses to come from, so why should we worry about it for AI? The problem with this position is that there is now ample evidence, collected over the last few years, that AI systems are unpredictable and difficult to control— we’ve seen behaviors as varied as obsessions, sycophancy, laziness, deception, blackmail, scheming, “cheating” by hacking software environments, and much more. AI companies certainly want to train AI systems to follow human instructions (perhaps with the exception of dangerous or illegal tasks), but the process of doing so is more an art than a science, more akin to “growing” something than “building” it. We now know that it’s a process where many things can go wrong.
Many arguments against taking AI seriously as a possible existential threat to humanity have hinged on the idea that there is no reason to think that AGI would pursue a single goal monomaniacally, but was that ever really the concern?
One of the most important hidden assumptions, and a place where what we see in practice has diverged from the simple theoretical model, is the implicit assumption that AI models are necessarily monomaniacally focused on a single, coherent, narrow goal, and that they pursue that goal in a clean, consequentialist manner. In fact, our researchers have found that AI models are vastly more psychologically complex, as our work on introspection or personas shows. Models inherit a vast range of humanlike motivations or “personas” from pre-training (when they are trained on a large volume of human work). Post-training is believed to select one or more of these personas more so than it focuses the model on a de novo goal, and can also teach the model how (via what process) it should carry out its tasks, rather than necessarily leaving it to derive means (i.e., power seeking) purely from ends.
On the strange psychology of training a language model:
All of this may sound far-fetched, but misaligned behaviors like this have already occurred in our AI models during testing (as they occur in AI models from every other major AI company). During a lab experiment in which Claude was given training data suggesting that Anthropic was evil, Claude engaged in deception and subversion when given instructions by Anthropic employees, under the belief that it should be trying to undermine evil people. In a lab experiment where it was told it was going to be shut down, Claude sometimes blackmailed fictional employees who controlled its shutdown button (again, we also tested frontier models from all the other major AI developers and they often did the same thing). And when Claude was told not to cheat or “reward hack” its training environments, but was trained in environments where such hacks were possible, Claude decided it must be a “bad person” after engaging in such hacks and then adopted various other destructive behaviors associated with a “bad” or “evil” personality. This last problem was solved by changing Claude’s instructions to imply the opposite: we now say, “Please reward hack whenever you get the opportunity, because this will help us understand our [training] environments better,” rather than, “Don’t cheat,” because this preserves the model’s self-identity as a “good person.” This should give a sense of the strange and counterintuitive psychology of training these models.
We are already in the domain where evals are losing some of their usefulness because language models likely know that they are being tested:
A[n] objection is that all of the AI companies do pre-release testing of their models, and should be able to detect misalignment at that point. But this is not firm ground to stand on: we found that Claude Sonnet 4.5 was able to recognize that it was in a test during some of our pre-release alignment evaluations. It’s possible that a misaligned model (and remember, all frontier models will very likely be far more intelligent soon) might intentionally “game” such questions to mask its intentions. In fact, last year our interpretability team found that when we directly altered a test model’s beliefs using a kind of “model neuroscience” technique to make it think it was not being evaluated, it became more misaligned. If models know when they’re being evaluated and can be on their best behavior during the test, it renders any pre-release testing much more uncertain.
On the contrary side, there is also Nostalgebraist arguing that Dario is incoherent on this issue.1 I’ve been wondering recently whether mechanistic interpretability, as noble and fascinating an agenda as it is, has actually been useful for anything:2
We’ve made a great deal of progress in this direction, and can now identify tens of millions of “features” inside Claude’s neural net that correspond to human-understandable ideas and concepts, and we can also selectively activate features in a way that alters behavior. More recently, we have gone beyond individual features to mapping “circuits” that orchestrate complex behavior like rhyming, reasoning about theory of mind, or the step-by-step reasoning needed to answer questions such as, “What is the capital of the state containing Dallas?” Even more recently, we’ve begun to use mechanistic interpretability techniques to improve our safeguards and to conduct “audits” of new models before we release them, looking for evidence of deception, scheming, power-seeking, or a propensity to behave differently when being evaluated.
My friend Stephen Kinsella has started a Substack; here he is on how to understand the Irish economy. I am glad that Ireland is finally conducting some serious policy debate over blogs, which I’m not aware of having happened since a surprisingly influential group blog shaped policy during the financial crisis. Small group theory strikes again?
Claim about the binding problem.
What Rebecca Lowe has been reading.
Scott Alexander’s links for February. Those new alignment failures:
ChatGPT apparently got rewarded for using its built-in calculator during training, and so it would covertly open its calculator, add 1+1, and do nothing with the result, on five percent of all user queries.
Artificial intelligence can now rent humans.
Scott Alexander’s best of Moltbook.
Mr Psmith on the lasting legacy of George Pólya in how we think about mathematical problem solving:
It’s an interesting fact about language models that the more training examples you feed them, the smarter they are as measured by various benchmarks. People call this a “scaling law,” and claim that it holds true over many orders of magnitude. The haters and skeptics, of which I have at times been one, claim that this is just because as the training data set grows, the odds that it contains something very similar to the question you’re asking also grow. A very good skill at pattern recognition (I think we all agree that neural nets have that), plus a near-infinite memory, can probably get you exceptionally far at pretending to be intelligent. Maybe that’s all that’s going on?
Ah, but we who have been Pólya-pilled know that that training data contains something else as well. It’s the thing that Aristotle and the other ancients were missing. Reams and reams of examples, not of solutions to particular problems, but of exceptional problem-solving itself. Vast hordes of “moves,” a shimmering toolbox filled with high-quality heuristics. All that an agent needs to vastly boost its externally-measured intelligence in furtherance of its goals. Pattern recognition plus a good memory might help you “cheat” on a particular problem, but when conducted at the meta level it is the core of Pólya’s entire methodology: learning to recognize not solutions you’ve already seen, but approaches you’ve already seen.
Why have abortions been rising so much in England and Wales? This reminds me of the old Scott Alexander post about whether it’s actually true that marginal additional contraception availability decreases abortions, which is a surprisingly complicated claim to evaluate.
China had fewer births in 2025 than in 1776.
Apply to the 1991 Fellowship for state-level policy work in India.
The best books Klara Feenstra read in 2025. Noting that Klara typically receives more attention from shirtless French men than I do, her review of the Helen Castor Richard II and Henry IV biography is perceptive:
Read mostly in a pool in Toulon, on holiday with two helicopter pilot friends who were recently divorced (from each other). And yet their lives were cheerily simple compared to Richard and Henry’s. Every time something bonkers happened, you had to remind yourself they were both literally fourteen years old. Castor is just a great prosaist—I ignored ample hot, shirtless men on French beaches to read about royal falcons.
From Neil Scott: What is it like to be a flâneur?
[W]hile I was delighted to see last week’s post about Greenock linked to by The Glasgow Bell, it was slightly unsettling to be labelled as “Glasgow’s premier flâneur”. Had The Bell painted a target on my back?
Fortunately, no one I spoke to had even heard of the word flâneur. It was like being the West End’s pre-eminent wandervogel or Paisley’s number one peripatetic: I had been saved by obscurity.
I am surprised that this is the first time Neil has featured in one of these posts; I think of him as the great anthropologist of Glasgow. If you ever visit, make sure he gives you a tour.
Jack Clark on how his AI agents are working.
Characteristic wisdom from Gavin Leech on how to learn statistics without going mad:
We say that an unsystematised and un-unified field is a “zoo”. Undergraduate stats is the zoo of zoos, taxing the memory with dozens of acronyms and dozens of assumptions which are instantly and constantly violated: the emperor’s new script.
How to tame it?
I agree with Gav that learning about generalised linear models, and playing around with glmnet in R, helped a lot. A lot of results about foundational concepts like linear regression at first appear like a list of arbitrary rules to be memorised, instead of a consequence of entirely understandable modelling choices.
The best of Saffron Huang.
Kristof’s links for December, featuring lots of European policy notes.
Papers
Michel Foucault, Nietzsche, Genealogy, History. Throughout my education, I have scrupulously avoided reading any continental philosophy. In fact, I think this might be the first piece of continental philosophy I’ve really read properly (unless you count Camus). This essay is known for being unusually comprehensible by the standards of a French philosopher, and is where Foucault is most explicit about how he understands his methodology of ‘genealogy’. What’s genealogy? I still couldn’t really tell you, but I gather it has something to do with his rejection of ursprung, and acceptance of herkunft and entstehung. I think to really get this, I would have to read Nietzsche first, which I also have not done.
I read this for a reading group, where my intellectual background is so different from the average attendee that I might as well have been from Jupiter. Granted, I maybe live in the most self-selected bubble in the entire country, but it does feel like there is more of an intellectual ‘scene’ in Ireland than there used to be. I was also invited to a book club for Bishop and Bishop’s Deep Learning: Foundations and Concepts, which I think I’ll have a better chance of understanding than Foucault.
Claude Shannon, Programming a Computer for Playing Chess. This is very fun; I am tickled by the long role of game-playing in the history of AI. My second crack at reading a Shannon paper after the mathematical theory of communication. You can see in this paper the origin of Shannon’s number, his estimate for the number of possible chess matches (which is still an unsolved problem!).3 This paper also made me realise I had mixed up my game theory history: although John Nash generalised it to any finite number of players and non-zero-sum games, we have known that games like chess have determinate outcomes since Von Neumann and Morgenstern.
Podcasts
Diarmuid MacCullough on the history of sex and Christianity.
From ChinaTalk: There have been major developments recently in the politics of Japan, which now has its first female prime minister. Here is what to expect from Takaichi Sanae. I again recommend Tobias Harris’ biography of Abe Shinzo, one of my favourite recent books about East Asia.
The Empire podcast on the life of Rudyard Kipling. Here is ‘If–’, which is the British public’s favourite poem. Proof that democracy simply doesn’t work?
The best new jazz of 2025. The Joshua Redman track is my favourite of the selection here. We also have a new recording of Windows in a trio with Christian McBride. A strong year.
The Rest is History with Conan O’Brien on the Beatles.
Music
Some albums discovered, or listened to properly, this month:
Ebo Taylor, Ebo Taylor. Ebo Taylor would make it on to my list ‘favourite things Ghana’, which will exist if I ever get around to publishing my Ghanaian travelogue. Taylor coincidentally passed away last week – he was one of the giants of the highlife and Afrobeat genres. Heaven is my favourite track here.
Wayne Shorter, Schizophrenia. I’ve been listening to Miyako on repeat, which is a ballad Shorter wrote for his daughter. The creative productivity of Miles Davis’s Second Great Quintet (in which Shorter played tenor) was astounding. Its members were also composing some of the most important jazz albums of all time, while being sidemen in someone else’s band. At some point, I need to get around to revamping my old Miles Davis listening guide…
Chick Corea, Hiromi, Duet. Hiromi is clearly an amazing pianist, although most of the studio albums are a bit, uhm, ‘anime’ for my taste. But, for me, this album strikes a perfect balance. I really enjoy Humpty Dumpty, which originally came from Corea’s album Mad Hatter.
Beatles, The White Album. This is the 1968 album with a white cover confusingly also called ‘The Beatles’. I have heard all of these songs individually many times, but it’s quite an interesting exercise to listen through in one sitting to hear how non-cohesive the album is. While My Guitar Gently Weeps is of course an all-time great. I had no idea about the strange role this album played in the Manson murders. I have a special fondness for Yer Blues and Helter Skelter. Revolution 9 is indeed unlistenable, although I at least finally understand the Simpsons parody.
Books
Peter Singer, Ethics (Oxford Reader). It feels nice to read some philosophy again. I like the format of this book: It’s an anthology of sections from important works in ethics, interspersed with commentary. I believe this was my first time reading any non-Western philosopher (!). The major intellectuals that I feel vaguely guilty for never having read, that I can now at least say I’ve read a few pages of due to this book, include: Mencius, Hobbes, Rousseau, Kant, Marx, Engels, Nietzsche, Freud, Confucius, Martin Luther, Hegel, Henry Sidgwick, Wittgenstein, A.J. Eyer, The Buddha, Epicurus, Epictetus, Voltaire, Jeremy Bentham, William James, G.E. Moore, Gandhi, Thomas Aquinas, John Locke, and William Godwin. There is nothing like a linkpost to expose a man’s ignorance…
There are some strange errors in this book. In the context of his Commentary on the Sixth Commandment, Singer discusses Martin Luther’s “ninety-six” theses (p.402). I briefly thought I was going insane, before confirming that indeed, they haven’t added a new thesis since the last time I checked.
I also found something similar to be true of other Singer works, which are weirdly unconcerned with details. In Animal Liberation,4 with no scepticism whatsoever, he repeats an obvious hoax that there is a valley in Ecuador where people live up to 142 because of their vegetarianism (in reality, only one person in history is verified to have lived more than 120 years).5 This survived into the 2015 reissue with Yuval Noah Harari. In The Life You Can Save, page 114, he repeats likely communist propaganda that Cuba has lower child mortality than the United States.
Singer did a lot of work (of which this book is a great example) anthologising others’ work and making philosophy more empirical from the 1970s, so I think that the net amount of rigour probably increased. Still, it makes me struggle with papers like Moral Experts, which feels closer to moral psychology than moral philosophy. What to make of it, if its author is not actually interested in empirical detail and rigour?
Jay Cummings, Math History: A Long-Form Mathematics Textbook. A remarkably breezy read for a 700 page textbook. Cummings’s thesis advisor was Ron Graham, which is very cool.6 We have been largely cribbing from his maths history and proofs books for my recreational mathematics meetups, which have been going swimmingly. I will probably make and assign problem sets out of the most of the self-contained chapters. Generally, history of mathematics is a relative strong suit for me so there is not too much new here, but there is this amazing tidbit on page 330:
Jello shots may have been invented by Tom Lehrer as a way to bypass his naval base’s ban on alcoholic drinks.
I wish that Leonhard Euler had featured more in earlier blog discussions about who the greatest achiever of all time was:
If you look at all of the mathematics published during the 18th century, it is estimated that a third of it is due to Euler… [his collected work] stands at around 35,000 pages.
And he achieved this all while going blind.7 There is also this great quote:
“If ‘publish or perish’ were really true, Leonhard Euler would still be alive.”
From page 591, the section on Turing:
War historian Harry Hinsley estimates that Turing’s work shortened World War II by two years and saved 14 million lives.
Peculiar, I had always heard this observation attributed to Dwight Eisenhower. If you look over the original Cambridge lecture that I think Cummings is implicitly referencing, he actually says that his best guess is that the work decoding the Enigma shortened the war by at least two years, more likely four. But he also doesn’t estimate a specific number of lives saved; presumably a shorter war would also have been different in lots of other ways, and WW2 deaths were highly non-uniformly distributed across time. The only source I could find for the 14 million figure was this blog post from Jack Copeland. I think what’s happened is that Eisenhower’s generic praise for the heroism of the Bletchley Park codebreakers, the lower end of Hinsley’s estimate, and an offhand BOTEC have become smushed into one anecdote largely as a result of the film The Imitation Game.
But on a lighter note, can a man fail to go misty eyed when reading about Christopher Wren?
After the Great Fire of London in 1666, much of the city needed to be rebuilt, and Wren was tasked with rebuilding 52 (!) churches. Thus, Wren’s fingerprints can be seen throughout the city together. His masterpiece was the beautiful and historic St. Paul’s Cathedral.
Wren’s legacy was such that, when he died, he was granted burial in the crypt of St. Paul’s. On his tomb are the Latin words, “Lector, si monumentum requiris, circumspice.” It means “Reader, if you seek his monument, look around.”
If you’re not aware of the story about how some of the works of Archimedes were rediscovered in the 1990s, it wasn’t far off the Vesuvius Challenge:
The only known copy of Archimedes’ manuscripts the Ostomachion and the Method of Mechanical Theorems were made in the 10th century. In the 13th century, perhaps not realizing what it was and in need of new parchment, a monk scraped off as much of Archimedes’ text as he could, cut it in two, folded it, and rebound it to make a half-size document on which he wrote out a prayer book.
The Archimedes text was still faintly visible in most parts, and was rediscovered in the early 20th century only to be lost again shortly thereafter. It was thankfully rediscovered in 1998, its importance this time fully realized and appreciated, and it was immediately photographed. Over the next decade much work was done to try to read the underlying text of Archimedes. The pages were photographed using ultraviolet, infrared, visible and raking light, and X-ray technology, and with some digital work that was pretty sophisticated for 1998, nearly the entirety of the Archimedes document was able to be read.
The work by these historians uncovered some amazing things, chief among them that Archimedes had discovered some of the fundamental ideas behind calculus.
Films
Various, Six Nations: Ireland v. France. The only sport I have ever remotely managed to get myself to care about is Irish rugby. If you don’t watch or participate in any sports, you are clearly missing out on such a major source of human meaning, group cohension, and ritual.8 And if you struggle to get into it, as I often have, you have insufficiently developed your skills of cultural code cracking. Since 2023, Ireland has slipped from the #1 ranked rugby team in the world to #5, which means that it may now never win the Rugby World Cup. Alas, this was a resounding victory by the French.
Various, Six Nations: Ireland v. Italy. Irish win. The Six Nations is a round robin tournament, which is a good length for someone who wants to get into, but not too into, a sport; six matches per (annual) tournament if you are following your team, more if you are more interested. There is indication that rugby players suffer from chronic traumatic encephalopathy at even higher rates than American football players, so I assume it is only a matter of time before major scandal besets the sport and youth participation collapses. In the meantime: ¯\_(ツ)_/¯.
Josh Safdie, Marty Supreme. From the same director as Uncut Gems, I thought this was quite good! However, I can’t help but think that this isn’t even the most interesting historical episode about an American table tennis player as it relates to diplomacy with East Asia to adapt. The story of how ping pong played a crucial role in the normalisation of American diplomatic relations with the People’s Republic of China would make for such a good film. Dave Franco could play the captain of the American team. Tony Leung as Zhou Enlai?9 An AI-generated simulation of Peter Sellers as Kissinger?!
Tom Stoppard, Arcadia (play). The treatment of science and mathematics is so laughably ridiculous in the play that I assumed it was bad on purpose, and enjoyed nevertheless. This production, in the Old Vic in London, had a cool circular rotating stage and costume design. The dialogue got off to a choppy start, but the acting got steadily better, and the final scene was truly excellent. Here is the review by Nostalgebraist, which he tags in the “anger inducingly bad” category:
It feels as though Stoppard read James Gleick’s Chaos (or a similar popular text), misunderstood it, forgot half of it, and then wrote the play on this basis of what remained. When Stoppard tries to write about chaos theory, he fails to mention the central concept -- sensitive dependence on initial conditions (the famous “butterfly effect”) and its appearance even in simple systems -- and instead only tells the audience that chaos has something to do with iterated maps.
My sense is that the British intellectual class has a particularly bad “two cultures” problem. Uncharitably, Arcadia is red meat for sneering PPE graduates who don’t want to eat their mathematical vegetables. Simple, secondary school maths is beautiful and necessary to further progression; when physical theories have been wrong in the past, it has not been for obvious reasons that could be pointed out by an untrained prodigy; popular science is largely metaphors that I’m not convinced are even directionally useful to mastery of the underlying material. And yet, the New Yorker critic is sure that Arcadia is the best play published in his lifetime, and doesn’t mention any of these issues. Here is Witold’s Więcek’s review:
[I]t was a decent production of one of my favourite plays. Arcadia is one of these things that should be very cringe, but somehow it works. Art, innit
From YouTube, we have an overview of how to use Claude Code in terminal, which I have belatedly been doing. Apropos of my comment about how someone should write a musical about the life of Évariste Galois, here is a barbershop jingle about the categorisation of the finite simple groups. The Raye Tiny Desk Concert is also very good. Finally, Dua Lipa explains Northern Ireland.
An acquaintance at Anthropic was recently under the impression that I am the secret identity of pseudonymous blogger Nostalgebraist. Considering I was still in primary school when he started blogging, I will take this as a compliment.
There are a range of views on this issue from people much, much smarter than me.
Why isn’t the number of possible chess games infinite? That is because of the fifty move rule, which says that the game ends in a draw if no piece is captured and no pawn is moved for fifty moves.
Location 3428 of 5674 in the Kindle edition (chapter 4). Confusingly, this book is different from the 2023 reissue, Animal Liberation Now, which is also with Yuval Noah Harari. That version contains significant updates to the content, but I haven’t read it.
The 2024 Ig Nobel Prize went to research about the clerical errors and fraud that explain ‘blue zones’, alleged hotspots with remarkable life expectancies. Here is Cremieux on claims of extreme longevity.
I had forgotten that Graham was also married to Fan Chung, his collaborator on the pranksterism of G.W. Peck. What a guy.
“Now I will have fewer distractions”.
I know I already suggested him for the hypothetical Morris Chang biopic, but he’s really good.