What am I up to now?
April 1, 2025
March, 2025
Contents
Updates
The joys of February are just sufficient for 28 days. Oxford is never lonely during term time, and after two months away it’s been great seeing so many friends. My mom also visited for a few days, and in the grand tradition of My Parents, she fed many graduate students.
The winter term is quiet academically, but I did host a few meetings of the development econ student working group and present internally twice. We also had a visit from Prof. Raúl Sanchez de la Sierra, who gave a great talk on violence, pro-sociality, and the origin of the state.
February was an AI-heavy month for me. LLMs are the most fun technical tools available right now. I spent about $160 on AI this month, on subscriptions to ChatGPT, Claude, and Gemini, and API costs, mostly to OpenAI. My general uses are:
- ⁃ Gemini for large context, e.g., reading a whole book or multiple papers
- ⁃ GPT-4o for large volumes of tasks, occasionally in chat but more often through the API
- ⁃ GPT-o1 for thinking through analytical problems or double-checking my reasoning
- ⁃ GPT-o3-mini-high to try to one-shot medium-to-large coding tasksIf o3 stumbles on the first try, I find it rarely recovers regardless of effort.
- ⁃ Claude for dealing with discrete coding tasks
- ⁃ DeepSeek for weird, unique prose or wordsmithing
- ⁃ Perplexity for fact-checking queries which alternatively I would have Googled
Of course, there’s no reason not to get multiple answers, and these are all complements. All of the above uses have been scrambled by Claude 3.7, which is currently my first stop for practically everything except massive context.
More interesting has been recent experiments running small open-weight LLMs on my Macbook. There were a few days in Gambia in December when the internet was downReportedly, the WACS cables were under maintenance. and I would have loved access to an LLM. It is surprising how small LLMs can be made. I tried a few local models and now have Microsoft’s Phi-3 on my computer for the next internet outage. It’s 3.8 billion parameters,For comparison, Claude 3 and GPT-4 are in the 2 trillion range. My local version of Phi-3 is quantized from 16 to 4-bit. trained especially for math and coding reasoning problems. This is good enough for me, but not the state of the art: Gwern expects “we may have superhuman reasoning LLMs in the billion or sub-billion parameter range… which are just very, very ignorant, perhaps even more ignorant than you or me.” These tiny LLMs are never going to be a researcher’s primary tool. But they are really cool and surprisingly powerful.
This month there are two new posts on my blog tab, both on the labor effects of LLMs using new data from the Anthropic Economic Index. In the first, I use the Anthropic data to replicate some parameters in Acemoglu’s 2024 paper on “The Simple Macroeconomics of AI,” finding substantially above-expectation productivity effects. In the second, I examine which sectors are automating faster or slower than expected by economists at OpenAI and Acemoglu himself.
I’m seeking the sun. The spring rowing event, Torpids, is taking place under blue skies, but over a swollen river from days of downpour. In a couple of weeks, I’ll fly to Phoenix to see the sun, my dad, and my sisters, then on to DC on 20 March. DC friends, please say hi!
Reading
No books knocked me on my butt this month. I re-read Zuleika Dobson after my mom’s visit; she stayed at the Randolph Hotel, which has Sir Osbert Lancaster’s famous illustrations of the novel on the walls of the tea room. Tchaikovsky’s Children of Time was as entertaining as people say. A collection of Le Guin’s short stories was high variance. Two memoirs, from Obama’s 2008 campaign manager and the military head of the Manhattan Project, were interesting historical documents but lacked a literary spark.
Tell Them of Kings, Battles, and Elephants by Mathias Énard gestured at excellence. Énard, translated from the French by Charlotte Mandell, tells the possibly-true story of Michelangelo’s 1506 visit to Istanbul, to design a bridge for the Golden Horn. The novel’s endnote purports authenticity, based on a sketch of a “Project for a Bridge for the Golden Horn” which was found in the Ottoman archives and attributed to Michelangelo, as well as some security records. Reviewers doubt: the Asian Review of Books, the London Review of Books, the New Yorker. There are nice sections which gesture at Borgesian thought experiments, but the novel gets a bit too French for me. For alternate histories about voyages by Renaissance painters, I prefer the one about Caravaggio’s tennis game.
I’m hopeful for March. I stayed up until 1am last night on the first 150 pages of Love in the Time of Cholera.
Links
Last month, I linked to a few Paris Review interviews. This month I read Haruki Murakami’s.
If I were to translate postmodern writers like Don DeLillo, John Barth, or Thomas Pynchon, there would be a crash—my insanity against their insanity. I admire their work, of course; but when I translate I choose realists.
…
Many readers in Japan read my books on the train while commuting. The average salaryman spends two hours a day commuting and he spends those hours reading. That’s why my big books are printed in two volumes: They would be too heavy in one. Some people write me letters, complaining that they laugh when they read my books on the train! It’s very embarrassing for them. Those are the letters I like most. I know they are laughing, reading my books; that’s good. I like to make people laugh every ten pages.
And if that second paragraph confuses you, read these 10 observations about Tokyo.
Herzog on how he would stage Hamlet:
In 1976 I made a film about the world championship of livestock auctioneers, How Much Wood Would a Woodchuck Chuck, which had to do with my fascination with the limits of language. That’s why Hölderlin and the Baroque poet Quirinus Kuhlmann are so important to me, because in their different ways they approached the limits of my language, German. In Stroszek, when Stroszek’s dream of America is broken, his mobile home is put up for auction. The actor in the scene was a one-time world champion livestock auctioneer whom I had followed to Wyoming and brought out of retirement for my film. His auction, in which language becomes singsong, a cascade of madness impossible to intensify further, is surely unforgettable to anyone who sees the film. I always had the suspicion that this raving was the last form of poetry or at least the last language of capitalism. I always wanted to direct a Hamlet and have all the parts played by ex-champion livestock auctioneers; I wanted the performance to come in at under fourteen minutes. Shakespeare’s text is widely known anyway, and to prepare for the production, an audience would only have had to refresh their memories of it briefly.
A love story in the Wesleyan student newspaper.
Sam Kriss on pangolins, with echoes of Joe Carlsmith on otherness and gentleness.
Nathan Goldwag never misses , and this meditation on Star Trek, the End of History, and post-liberalism is great. We share a passion for Lower Decks.
Thanks to RS for pointing me to this collection of translations of the “Samson” poem in Stanisław Lem’s Cyberiad. I recently wrote about this poem here.
Erik Hoel on death is touching and sad.
Here’s a video of a dog.
Previously
…