Links for December
Goblins, but more so
I hope everyone had a happy Christmas and New Year. Here are the most interesting things I read, listened to, and watched in December.
Applications are open (until January 24th) for the Summer Program for Applied Rationality and Cognition (SPARC). I may stop by.
Blogs and short links
The post is late this month, because it took so damn long to finish my Notes on Taiwan. I have been pleased with the response; it made it to Marginal Revolution, The Browser, and Conor Friedersdorf’s Best of Journalism list. If a few more people leave comments, it will be worth a dedicated comment response post.
My Links for November was abridged and reposted on EconLog.
I also appeared on Matt Teichman’s podcast.
On the utility of vibecoding. Is Claude Code the chimp, the gun, or the human?
More nerd weddings. I sometimes think that, if things don’t work out and I have to switch careers entirely, Rachel Edwards and I could become world-class boutique wedding planners in the Bay Area for people with niche academic interests.
The scientific contributions of the Carlsberg beer company. Danish Guinness?!
Peter McLaughlin’s year in books, part 2. Who among us can really say we haven’t experienced the 19th-century novel as homeomorphic to a non-globally Euclidean manifold?
I decided that, if I was going to keep talking shit about Dickens, I had to at least read one novel that Jeff Daniels couldn’t dismiss as ‘minor Dickens’. But I could not finish it, which certainly meant it didn’t change my views. You know the way you have manifolds, spaces that locally look like ℝ^n but can have a radically different global structure? Dickens novels are like manifolds: Great Expectations is locally homeomorphic to a good book, but then you zoom out… His sentences really could almost seduce me, and I could therefore enjoy many of the chapters built out of them. But I never wanted to pick it up after having put it down.
What should middle powers do for compute strategy?
It is astonishing how many anecdotes about the dysfunction of the Irish planning system involve bats in particular.
In happier news, Ireland is now #1 ranked in Excel.
According to this press release, the Jaipur Literature Festival will be coming to Dublin in 2026. A big day for Irish boys suspiciously interested in India.
Gwern contra Gary Marcus on scaling.
Everybody seems to be joining Anthropic.
Dynomight on why he hates the NumPy library:
This is insane. Using basic features should not require solving crazy logic puzzles.
And here is his alternative proposal of DumPy.
We finally have Goodreads for academic papers.1 Predictably, Gavin Leech has already joined and makes up a significant fraction of all content on the website.
The central importance of shoring up your intellectual foundations. Alsp: How often are you maxing out your cognitive horsepower?
Pitches are now welcome for In Development, a new Works in Progress-inspired magazine focused on the developing world.
From The Fitzwilliam: Many of Ireland’s cultural peculiarities have their origins in a largely forgotten form of 19th-century syncretic neopaganism – the phenomenon of Protestant magic. This was also featured on Marginal Revolution.
Our essay about why most foreign aid never leaves the country was also featured on Scott Alexander’s links for December. Scott also writes:
When a new AI model comes out, the companies typically take down the old version over the protests of researchers, hobbyists, people who think the old model was their boyfriend, and anyone else who wants access to obsolete models for some reason. Why can’t they just leave it up? Antra and Janus review the economics here: it’s inconvenient to be constantly switching GPUs from one model to another, so if there isn’t enough model-specific demand to keep the GPUs running at all times, then the company loses money. This is an interesting look at the details of AI deployment, and ends with a proposal to maintain old models through a “separate research application track”. Related: Anthropic to preserve weights of deprecated models, and include models’ own opinions in shaping the deprecation process. Good for them!
There are useful updates on many things here, including the COVID learning loss debate:
Study (h/t @KierkegaardEmil): there was minimal “learning loss” from COVID school closures, best estimate is “0.02 standard deviations per 100 days of school closure”. I correctly predicted this back in 2021, but I also wrote in March of this year about how there’s been a general decline in NAEP scores since then. It seems like maybe a student having their specific school closed for longer than other schools didn’t hurt them, but some sort of general cultural change, maybe related to COVID, did hurt.
Ben Southwood speaks the truth.
Unfortunately, I could not put my life on hold for a month to move to Argentina for Edge City Patagonia. But my friend Rucha went, and it seemed fun.
I have begun reading Import AI, the newsletter from Anthropic’s co-founder and Head of Policy, Jack Clark. He seems a thoughtful fellow. I enjoyed issue 431, which includes a transcript of his talk from The Curve conference. There is also issue 376, which reviews some research about how much LLMs’ performance drops in African languages. For some time, I’ve been meaning to work with Oisín Morrin on a Fitzwilliam essay about the challenges posed for machine translation by the Irish language.
Via Henry Oliver: How much did Shakespeare crib from Dante? I joined a reading group last month for The Divine Comedy, hosted by an old friend and classmate of Seamus Heaney’s. Will report back.
I have long recommended that my friends switch their phones to greyscale to make them less addictive. Many people struggle to stick with this, because there are occasional times (like when trying to take a photo) when having access to colour is quite important. However, you can configure your phone to be in monochrome except when the photos app is open, which presumably makes this strategy more robust.
Kristof Gazso’s links for November. I have mixed thoughts about the role I’ve played in spreading this genre.
We also have Benedict Springbett’s Links for November. Benedict kindly says that Progress Ireland is “on a roll”, and I’m inclined to agree. All of the projects I’m involved with (The Fitzwilliam, PI, the reading group, the maths circles, the Irish Maths Olympiad) have been going exceptionally well recently. On top of that, all my friends are getting married and having babies; it truly is a golden age.
The Irish government has announced a surprisingly radical plan to remove obstacles to the delivery of infrastructure, and Progress Ireland made a dashboard to track whether it will actually happen.2
A spiritual competitor to Lynkmi? (A reminder that you can use my invite code to join.)
If you want to improve your retention from podcasts, try TurboCharge by Michael Slade.
W.B. Yeats is the quintessential figure of Irish nationalism, but in many ways, nationalism enters as a tragedy in his poetry.
Sam Glover on the things that you should know about London.
The Guardian on the life of Saul Kripke. Can’t a man army crawl in front of a group of undergraduates without it making it into his obituary?
During a seminar in a lecture room he had recently taught in, [Kripke] was seen crawling along the central table, reaching down to grab an umbrella, and crawling back again. He had wanted to be inconspicuous, he said, when questioned about that method of retrieval.
I also note that Kripke was an exception to the general rule that philosophy doesn’t have prodigies:
He had taught himself ancient Hebrew by the age of 6, had finished reading Shakespeare’s complete works by 9 and published his first completeness theorem in modal logic when he was 18.
By pure coincidence, I first learned about Kripke’s causal theory of reference in a philosophy of language class on the day that he died. Later that week, I suffered the greatest burn of my philosophical career, when a tutor told me that, if Saul Kripke had heard the argument I had just made, “he would come back to life, just to die again”.3
The beauty of giving things sustained attention.
Best of Wikipedia: The Glasgow ice cream wars.
Congratulations to the new Astral Codex Ten grantees (I did not win).4 I am pleased to see more work being done on the problem of lead–acid battery recycling. You can also see the winners of the Non-Book Review Contest.
From The Economist: 1,000 people are now arrested per month in Britain for posting on social media. You can also support Sam Glover and Stella Tsantekidou’s new campaign to save freedom of speech in the UK; here is the tweet version.
Unfortunately, Karl Friston’s unified theory of the brain (the free energy principle) is probably bunk.
Harry Law’s introduction to the history of AI and overview of the legacy of the Dartmouth Conference. It’s so funny that Marvin Minsky was the scientific advisor for HAL in 2001: A Space Odyssey.
The latest outgrowth of Julian Gough’s three-stage cosmological natural selection model is Blowtorch Theory. Julian also received sympathetic coverage in The Irish Times, and there has been some surprisingly fruitful discussion on Hacker News. I am in a groupchat with some economics professors, and one of them recently posted in an article about how poorly we can account for the findings of the James Webb Space Telescope with the caption:
Guys, is cosmology even more fucked than macro?
Apply by March 31st for Tyler Cowen and Patrick Collison’s call for new aesthetics.
Preliminary thoughts on Timothy Williamson’s new book. Has Williamson, in fact, done these things?! Before he turned his intellectual powers toward understanding something even more complicated and pointless (Irish planning law), my colleague Seán O’Neill McPartlin almost became a professional philosopher. He relays the following anecdote about Williamson, which has become legendary in the friend group.
I was speaking to Tim about a paper I was writing. The paper was on the question of whether and when it was epistemically virtuous to ignore some propositions which have a non-zero chance of being true.
I was set on this question by reading two [sources]. . . The second was . . . a book. In the opening . . . the author discusses how it is fine (epistemically speaking) to ignore certain beliefs. His example was that of religious morality.
Naturally, we all think some degree of this practice is both necessary and fine. But, that remark about religious morality struck me as rather hasty. And so, I set off to think about it a little bit more.
While thinking about it, I asked Tim to speak with me about this phenomenon. And asked whether my views and approach were cogent.
[D]iscussing propositions whose potential truth values might be permissibly ignored generates amusing examples. Tim’s was “There is a hobgoblin under the bridge.”
After a little while, it occurred to me that I have no idea what distinguishes a hobgoblin from a goblin. When I asked, Tim delivered the wonderfully succinct: “uhh, it’s like a goblin… but more so.”
The Financial Times (gated) on whether Modi will touch the ‘third rail’ of Indian politics, namely the updating of parliamentary constituency sizes to reflect their current populations. They have been frozen in place since Indira Gandhi shoved through a constitutional amendment in 1976, in order that states not be dissuaded from implementing aggressive family planning. The can was kicked a further 25 years down the road by Atal Vajpayee, a directive which is expiring this year. Here is an excellent three-hour podcast from Shruti Rajagopalan explaining the Indian political context of ‘delimitation’. Something I omitted from my blog post about Tamil Nadu is the remarkably disproportionate role that former actors play in Tamil politics. I also somehow missed that the Chief Minister of Tamil Nadu is literally called Stalin??? (Yes, he is named after the Stalin.)
What the Taliban are up to in their new lives in Kabul:
In our ministry, there is little work for me to do. I spend most of my time on Twitter. We’re connected to speedy Wi-Fi and internet. Many mujahedin, including me, are addicted to the internet, especially Twitter.
Podcasts
Dan Wang on what you should do if you want to become the leader of China. I found this most interesting for Sam Bowman and Pieter Garicano’s discussion of what ails European economies. The extent to which the US has much greater competition in banking is remarkable. I recently tried to open a new bank account, but there were literally only three banks in the country I could choose from. I found the experience so complicated and unpleasant that I eventually gave up. Meanwhile, around 80% of financing for European companies happens via bank lending, compared to around 30% in America – the quality of financial services really matters!
Joshua Wallerstein explains Bruckner’s 4th and Bruckner’s 7th Symphonies. I heard the 9th Symphony live last month and enjoyed it a great deal, thanks to KG for joining me.
Is Aja by Steely Dan the greatest album of all time? There is a classic Tumblr post about this, allegedly posted by an atemporal being that has experienced all music:
i just listened to all the music thats ever been. my favorite was a song a little girl sang to her cat in greece in 1286. my second favorite was from a ceremonial performance in ethiopia, from 22000 bc. i also liked STEELY DAN.
Dan Wang on what America and China can learn from each other.
Robert Wiblin and Luisa Rodriguez on parenting in the 2020s. I have always had an inordinate fondness for French women for bucking the international trend and actually spending less time with their children than they did in the 1960s.5 And here is Rob’s list of product recommendations for new parents.
Reed Hundt, the former commissioner of the FCC, on his Year of Bach.
The Marginal Revolutionaries on their favourite economic models. Listen for a good explanation of why monopoly, a priori, does not imply a lower quality of goods than would be socially efficient. I also wonder how many people in Ireland could correctly explain Harberger’s general equilibrium capital taxation result, despite it basically being the intellectual cornerstone of our economy.
Among the countries that still recognise Taiwan as an independent country, almost half are in the Caribbean. During the recent presidency of Tsai Ing-wen, the number of Taiwanese embassies around the world dwindled from 22 to 12. Here is more on Taiwan’s vanishing Caribbean friends.
Today in non-sequiturs: Why the War on Drugs is responsible for the wide availability of blueberries.
Henry Oliver interviewed about what it means to have good taste on Connecticut Public Radio. I was interviewed about this blog once on New Zealand public radio, although I do not know if the recording ever saw the light of day.
Soham Sankaran on building a vaccine company in India. This reminds me of many conversations I have had with Akash Kulgod about his experience scaling Dognosis from Bangalore. Akash would be a great guest for this podcast.
Music
Some albums I discovered, or listened to properly, for the first time this month:
Ulkar Aghayeva, Fugue Chahargah. My first time listening to Azerbaijani music. This is a piano fugue inspired by mugham, the traditional microtonal music of Azerbaijan. Ulkar also has a suite for cello and piano and a string trio. She is insanely talented, and I hope to make it to one of her concerts someday soon.
D’Angelo, Live at the London Jazz Cafe. I only learned of the existence of D’Angelo recently from this podcast episode, and I don’t know how I had missed him for so long. My favourite track here is Me and Those Dreamin’ Eyes of Mine. When I visit London, I always fill my time to the brim, but next time I’m there, I need to make time to visit some jazz venues.
Keith Jarrett, Facing You. Listening through some of the solo albums in the lead-up to the 1975 Köln concert, one of my favourite albums ever. In Front is difficult listening that I still don’t understand. My favourite track here is Lalene, which might be my favourite song discovered in 2025.
Hiromi, Sonicwonderland. Found via her appearance on NPR Tiny Desk Concert. My favourite song is Up. I haven’t had a chance to watch the interview with Rick Beatto yet, but Hiromi’s playing is electrifying. You can also listen to the duet with Chick Corea, whom (if Wikipedia is to be believed) she met by chance aged 17 and was invited to perform with on stage later that day.
A very strong showing this month.
Papers
John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon, A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. This document is the first official usage of the term ‘artificial intelligence’, laying out the research agenda that would become the 1956 Dartmouth workshop.6 Is the significance of the Dartmouth conference overrated in the history of AI? Probably yes. Are we privileged to be alive in an era when we can read the original proposal, and ask questions of it using magic boxes that largely solved all the problems outlined in the proposal within a single human lifetime? Also yes.
Roy Foster, Protestant Magic: W.B. Yeats and the Spell of Irish History. One of the most important essays in all of Irish historiography. I was very pleased to finally meet Professor Foster in person last month, and to get a signed copy of Modern Ireland: 1600–1972. His eyes lit up when I mentioned his appearance on the Conversations with Tyler podcast.7 For an extensive discussion of Protestant magic, read our essay for The Fitzwilliam. There are also some fragments relevant to my forthcoming post on the Irish Enlightenment, e.g. page 258:
[Yeats] would reject Swift and Berkeley as not really Irish, while claiming Blake and Emily Brontë as countrymen.
There is some good stuff here on the milieu from which Dracula emerged, for which you should read my favoured theory that the novel is an allegory for Ireland’s failure to abide by international timezone standardisation treaties.
Various, Toy Models of Superposition. This paper is long, and in retrospect was an ambitious choice of reading to discuss at my mate’s AI journal club. In any case, this is the paper from 2022 that birthed a lot of Anthropic’s research agenda related to interpretability, i.e. understanding what on earth neural networks are actually doing. The authors introduced the phrase ‘superposition’ to describe a situation in which a neural network encodes more concepts than it has neurons. It is the exception rather than the rule for a dimension in a model’s activation space to correspond cleanly to a dimension in the feature space.8 So you get individual neurons that represent contrived compound concepts like “dogs half the time, but also sometimes the year 1847, or pictures of casinos”. Those neurons are called polysemantic, and the holy grail of interpretability would be to have monosemanticity.
My sense is that until relatively recently, interpretability had not actually been used directly in training or deploying frontier AI models. Still, in principle, it seems extremely promising, and this area seems to nerdsnipe a lot of very intelligent people. For example, there is a section of this paper about how the problem of trying to create as many orthogonal dimensions in activation space as possible (in order to parse out as many disctinct concepts as possible) is mathematically equivalent to the problem J.J. Thomson faced in trying to calculate electron configurations in his plum pudding model of the atom. That also connects directly with the geometry of uniform polytopes.
Further discussion is probably a wee bit technical for the Sam Enright blog. If you are interested in more details, you can watch the Neel Nanda walkthrough.
Books
Jonathan Swift, Gulliver’s Travels. Unless you count graphic novels, plays, poetry, or autofiction, this is the only novel I read in 2025. One of my New Year’s resolutions is to read more fiction. In short, Gulliver’s Travels is A++, and possibly the greatest novel I’ve ever read. For more thoughts, you’ll have to wait for my Irish Enlightenment post.
Dwarkesh Patel, Gavin Leech, The Scaling Era: An Oral History of AI 2019–2025. I was pleased to re-read this book in hardback; I also reviewed it in Links for January. The book consists of heavily edited transcripts of the Dwarkesh podcast, with some bonus essays at the end, including The Bitter Lesson and Nostalgebraist on what lessons we should draw from the success of the transformer architecture.
Gavin came to Ireland for a meetup we had about this book we called ‘The Scaling Éire’ (get it?!). It was my favourite reading group we’ve done, and was a reminder of how lucky I am to have stumbled, almost by accident, into having some of the greatest mentors in the world. <3
You can also read the New Yorker coverage or Gavin’s review of his own book. The extraordinary pace of AI is a bad fit for people who, like me, learn best from reading books. One of the only other up-to-date ones I’m aware of is the Welch Labs illustrated guide to AI, which I have pre-ordered.
Jacob Mikanowski, Goodbye, Eastern Europe: An Intimate History of a Divided Land. The first serious book about Eastern Europe I have read. I thought this was a breath of fresh air, being a history book explicitly organised by theme rather than chronology. I have many other thoughts, which I hope to cover in a sequel to Notes on Taiwan about my time in Poland.
Films
Greg Kohs, The Thinking Game. A documentary about Demis Hassabis and the story of Google DeepMind. You can watch the full thing for free on YouTube. The main new thing I learned about was Demis’s role in developing the 1994 video game Theme Park.
This made for strange viewing. OpenAI, Anthropic, and ChatGPT are never mentioned. You could watch this entire film, and be under the impression that AlphaGo-era reinforcement learning is still the dominant paradigm. Significant sections of it could equally have appeared in the 2017 documentary about AlphaGo, which is one of the greatest documentaries of all time.
I was also disappointed that Demis’s side gig as the world’s fourth-highest-rated player in the world of the board game Diplomacy wasn’t mentioned.
Edward Yang, Yi Yi (一一). My first Taiwanese film, found by asking Claude Opus 4.5 “What is the greatest Taiwanese film of all time?” This was Yang’s last film, and is a major part of the Taiwanese New Wave. The title means ‘one by one’, and, when written vertically, looks like the Chinese character for the number 2.
I think this is one of the best films I have ever seen. It captured the ‘relentlessness’ of mundane family life better than any other visual media I can think of. I later learned this is one of the few times Scott Sumner has given a perfect score to a film:
At one point a character says something to the effect that movies triple the size of our lives. We experience far more than otherwise. This film is certainly a good example. Like almost all of the greatest films of the 21st century, it came out at the beginning of the millennium. Not a good sign.
I loved how wide an age distribution the main characters have. The scene toward the end of the father and daughter going on parallel, stilted dates is a masterpiece.
Wong Kar-wai, Fallen Angels. My third Wong Kar-wai, after Chungking Express and In the Mood for Love. It is definitely much closer in spirit to the former, and I later learned that Fallen Angels is often seen as the derivative and unnecessary entry in Wong’s oeuvre. East Asian Cinema History defends the unorthodox take that it’s his best film. Somehow, I found Fallen Angels to be a better exploration of what Stella Tsantekidou calls the desperation of female neediness than the other two. The sex scenes convey a kind of hypersexualised but romanceless society that makes for timely rewatching in 2025.
A few days ago, I had a surreal moment where it felt like I was inside Chungking Express. At a sandwich joint I frequent, there was a girl who looked similar to Faye Wong. She served me while blaring late 1960s West Coast American folk rock, while dancing and shouting over the music to ask me what my favourite movie is. I hope she doesn’t break into my apartment later.
From YouTube, we have Grant Sanderson explaining the Laplace transform. There is also Tame Imapala, Tom Misch, and CHIC on NPR Tiny Desk. Finally, The New Yorker explains for the unitiatied why the Irish Republican Army was required to use voice actors.
Thanks to Anuja Uppuluri.
For example: “The government will . . . limit who can take a judicial review to those who would be directly impacted by any proposed project.”
No, I can’t remember what my argument actually was.
I hope I am not in the group that of unspecified size that were turned down because “you were suffering from LLM psychosis”.
Although apparently the data quality is rubbish, as they discuss in this episode. I don’t know if that is related to why the Our World in Data page on how much time parents spend with their children is now only available on Internet Archive, and redirects to the general page on time use.
I also learned that the Dartmouth conference was funded by the Rockefeller Foundation.
He seemed shocked that anyone outside the walls of Trinity College, let alone an American, knew about Isaac Butt.
WTF counts as a feature? Isn’t that inherinetly a vague and subjective concept? They give a definition of that in the paper for demonstration purposes, and this is part of the reason why these are only “toy” models.



There are a lot of great bootlegs of Keith Jarrett solo from the 70s on YouTube. Check out Stockholm and Finland…
I loved Yi Yi too. You might like Hou Hsiao-hsien's movies. Some of my favorites are A City of Sadness, Flowers of Shanghai, and Three Times.