David's notes #2: AI math, the Epstein emails, and why Steve Jobs wanted to buy General Motors

Brief notes from what I've been reading

Benedetta Cappa, Speeding Motorboat

Hello! I hope you enjoyed my piece this week on why GDP numbers in poor countries are untrustworthy. There’s a piece that was supposed to be published on Friday, but I held it up to add some more stuff. It’ll be coming out early this week instead. And thank you to all those who’ve subscribed in the last week—several hundred people! I hope you enjoy what I’ll be publishing going forward.

This is my weekly round-up of things I’ve been reading.

---

1. Friend-of-the-newsletter Alex Kesin has a new podcast called Approved, “chronicling the most significant breakthroughs in biotechnology.” The first episode is on Amylin Pharmaceuticals, a company I didn’t know anything about but which pioneered the GLP-1 drug class, which has now produced roughly a trillion dollars in enterprise value. And—as is so often the case—they weren’t the ones who actually profited off it! A phrase from the podcast I’d never heard before: “Pioneers get slaughtered, settlers get rich.” There are also really interesting side-stories like the emergence of San Diego as a biotechnology hub and the surprising importance of Carl Icahn in the company’s downfall. Highly recommended!

2. On AI and formalization in mathematics. I’ve been trying to read about Lean as a bottleneck for “autonomous” AI proofs in mathematics. Mathematics is a highly oral culture—a sign of its vitality—so not much is written down, and I’ve had to rely on mathematician friends to point me in the right directions. But the gist of the problem, as I understand it so far, is that rigorous attempts to use AI to solve conjectures rely on formal languages like Lean; but Lean isn’t complete (you can sometimes use Lean to “prove” things that are false, this is called “misformalization”), so a lot of these attempts just end up finding exploits in Lean. There are a lot of things in Lean that are underdefined. The most interesting person working on this is Kevin Buzzard of Imperial College London, who has a great blog called Xena. The mathematician Boris Alexeev has a great guest post focusing on the Erdos problems; see this post by Buzzard as well. Right now these problems affect projects like Harmonic more than GPT 5.2 Pro, which is very good despite not relying on formal proof languages and instead appears to reason intuitively; but Buzzard is skeptical that a pure-LLM approach will have as much mileage as an LLM/Lean hybrid. “Teaching Lean many modern definitions seems to me to be the bottleneck; it is by no means an impossible problem to solve, but it will take time, and solving it seems to me to be a fundamental problem. … Until we find a better solution, it will be my group at Imperial College London who will work on this problem, manually.”

3. I’ve recently become interested in skydiving. I think what I like about skydiving is that you can’t help but feel, as you get ready to jump, that you’re about to die: you can tell yourself that you’ll be fine, because you will be fine, but your prefrontal cortex knows that you’re about to jump from a massive height and all your evolutionary conditioning is telling you that you’re going to die. And then the actual physical sensation of falling at terminal velocity toward the earth is incredible. It’s a testament to the human capacity for optimization that we took jumping out of an airplane and somehow managed to make it safe. (Skydiving is about as safe as a long car trip.) And the main character in that story of optimization is an engineer and businessman named Bill Booth, whose invention of the three-ring release system was crucial in allowing for tandem skydives. Booth still appears in the safety videos they make you watch before you do a tandem jump. There are heroes all around us!

4. A fascinating 1996 Wired interview with Walter Wriston, former CEO of Citibank. I read this while doing research for a post that will be published in the coming week. Wriston was a really big deal at the time—the interview mentions a 1000-page biography of him, which the article says “regards him as the kingpin of modern American finance.” Nobody today knows about Walter Wriston! But it’s incredible to think just how much innovation there was in finance between 1950 and 2000: things like the index fund or Eurodollars that are key parts of How The World Actually Works. And it’s remarkable to see a guy who was involved in all of it talk about this stuff. For example, on the invention of the cash management account: “Don Regan, when he was running Merrill Lynch, changed the world, which doesn’t happen very often. He invented the cash management account. The CMA consolidated checking, savings, investments, and borrowing all in one account, which is invested in a money market fund.” (Don Regan is better-known as the powerful chief of staff to Ronald Reagan: I always thought it was funny that Ron Reagan’s right-hand man was named Don Regan.) It was under Wriston that Citibank became the first big financial institution to really bet on ATMs. A story he mentions from that time: “Jack Scantlin came into my office one day and said, ‘You guys are doing everything wrong. You’re working on analog technology, which is yesterday’s newspaper. The future is digital.’ Ninety days later, he showed up again with what we now call a T1 terminal. He dipped in my Citibank card, and up came my balance. Scantlin designed our ATMs and bank cards and computer systems—all digital. Suddenly the world changed, because we listened to him, and he was right. Scantlin’s a genius who now lives on his own island in the South Pacific.” (You can’t find much of anything about Jack Scantlin online, by the way, even though he appears to have been the main force behind Citibank’s deployment of ATMs. He also developed the Quotron—the first tool to deliver stock market quotes to an electronic screen and a precursor to Bloomberg.) The whole interview is really fun, and you get the sense, wow, this guy is really smart. Wriston is also extremely forward-thinking. He makes predictions about the future of finance that pretty much all turned out to be correct (the rise of smartcards and electronic money), talks about his long history with encryption (“I was guarding two Sigaba machines on the Pacific island of Cebu when I was relieved of duty by a newly commissioned second lieutenant”) and about why it’s hard for other countries to dump dollar assets (basically the same analysis Eswar Prasad has in The Dollar Trap), and then says that the ongoing information revolution is “the third great revolution in the history of the world.” It’s an incredibly interesting interview.

5. China’s “factory” for elite human capital (paywalled). The Chinese education system seems to be very intentional about identifying the most talented students from an early age and pushing them through an expedited track (the very best ones can even skip the famously difficult gaokao exam) that seems to end, in disproportionate numbers, in senior leadership positions at tech companies. A really interesting piece.

6. I honestly don’t know what to think of Moltbook, where many Claude agents are talking with each other. Scott Alexander’s roundup of posts from it and Jason Manning’s notes on it are both interesting.

7. Kimi 2.5 is out. It’s good! The technical report is here. One thing that people are saying is that the creative writing abilities have somewhat atrophied, though it still seems strong to me. I’ve heard from friends that its performance in complex content moderation tasks is comparable to the latest Claude Sonnet model—quite remarkable for an open-weights model.

8. Friend-of-the-newsletter daniel bashir on AI and language. A very subtle reading of LLMs’ relationship to language through Saussure and Derrida. (Daniel is an ML engineer at OpenAI—it’s incredible that people can be both so literate and so technical! The human mind is simply remarkable.) I have never read Saussure or Derrida but my understanding of Daniel’s argument is this. The Platonic Representation Hypothesis speculates that the convergence of internal representations between different neural networks (trained on different data and modalities) suggests that they are converging on “the real.” Daniel says that in fact “the models’ internal geometries do converge, but what they converge on is not the world itself; it is a value-system that happens, contingently, to be useful in navigating it.” And he does this by defending Derrida against critics who would dismiss him in favor of a pure return to Saussure. (The Saussure view of language is that it is positional, i.e. “an internally structured web of signs,” and only secondarily a “ladder of reference” to the outside world.) But Derrida adds something Saussure doesn’t: the question of who stands behind an utterance. This lets Bashir identify LLMs as a third thing, which is “language that arrives with the grammar of address and no addressor.” I have never read Saussure or Derrida, and really have never thought very deeply about language, but Daniel is very smart, so I defer to him.

9. Europe wants to build sovereign technology (paywalled). Unsurprisingly this is a French-led effort, since of all the countries in the European Union it is France that has the most robust and innovative software sector. If Europe actually wants to become more “sovereign” and reduce its dependence on the U.S., then cloud and software is one of the logical places to start. Europe would likely have had several huge tech giants to call its own if it had enforced a “European firewall” of some sort 15 or 20 years ago.

10. I really dislike the word “enshittification” and all that it suggests. First, I hate that the word itself is vulgar and unpleasant: Gen Xers and older millennials grew up in a time when cursing was more interesting and provocative than it is today, and so their contributions to language (Cory Doctorow, who coined the term, is an archetypal Gen Xer) are of course vulgar and unpleasant and designed to provoke authority figures who vanished from the scene decades ago. The president is cursing all the time: the more interesting and transgressive thing to do nowadays is to make whatever point you want to make without cursing. (Especially since cursing, as it’s typically employed, substitutes the shock value of the curse word and the implicit emotional vehemence that curse word implies for any actual reasoning or convincing. This works better when cursing is rare and actually shocking.) And second, I hate what the word actually suggests, which is a sort of affluent consumer politics: you use all these random services—Uber, DoorDash, Instagram, Netflix, so on—and you are offended that they are too expensive, or that the algorithm for serving you content is bad, or whatever. In the words of a plodding and poorly written Slate article about nostalgia for Pizza Hut, enshittification is “the idea that the American marketplace has grown steadily more hostile toward its participants. Enshittification occurs when a subscription service triples in price, or when a gym membership tier suddenly plummets in value.” I understand that being an affluent consumer can be frustrating, because your counterparties are constantly trying to discover the absolute maximum level at which you’re willing to part with your cash, and now they have a lot of computing power to do that for them, and your financial relationship with them isn’t really embedded in a social world in which you can admonish them for charging too much (as you might with, say, your local butcher, back when people had a face-to-face relationship with their local butcher). But I just have no sympathy for this sort of self-pitying, nostalgic nonsense about how Pizza Hut used to be better or how airports aren’t nice anymore. You are one of the most fortunate people in the history of the world! You are able to fly around at your own pleasure, you are able to see both sides of clouds, you can have someone who is a much harder worker than you bring you a burrito at any hour of the day, you can watch movies whenever you want. You are the beneficiary of a truly enormous amount of consumer surplus. There are people in the world with lots of problems: disease, famine, poverty, war. The noble thing to do would be to direct your energies to help them in one way or another, and there are indeed many ways to help them; but the very least you could do is to stop acting like the great victim of capitalism is the American consumer.

11. The Epstein emails are interesting as a view into a certain elite world—an elite world that is, today, largely in decline. This was an elite world that succeeded the WASPs and whose influence began to decline with the financial crisis in 2008. These were the guys who took over huge parts of finance and politics in the 1970s and ‘80s; a lot of them were Jewish, others were “outer borough” guys of one kind or another (I’d count Bill Clinton and Donald Trump as “outer borough” guys), but they were almost all severed from the old blue-blood WASP world of Groton and Yale and such. It’s no wonder that one of Epstein’s early sponsors was Ace Greenberg of Bear Stearns, who wrote the famous memo about wanting to hire people who were “poor, smart, and with a deep desire to become rich” rather than the classic investment banker set. Those were the guys that Epstein associated with: not old-money WASPs but arrivistes like himself. Another interesting thing from the emails is that he seems increasingly out-of-place in the 2010s: he starts paying some random guy to make his web results look better and seems to devote a lot of attention to that doomed project. And he also tries to build networks with the rising tech elite of the West Coast. He reminds me a lot of certain arriviste/trickster archetypes—Julien Sorel (from The Red and the Black), Barry Lyndon, and Tom Ripley most prominently. The New York Times piece on how he made his money is also fascinating. And it seems (paywalled) like he actually was quite knowledgeable about tax optimization?

12. Dwarkesh Patel interviews biochemist Nick Lane. This came out a few months ago, but it’s really good; the episodes where Dwarkesh sits down with a scientist and just has them explain their work are always tremendous. The basic theory, as I understand it, is that simple life (bacteria) is chemically inevitable and likely arises everywhere in the universe where you have rocks, water, and carbon dioxide. But complex life is extremely rare, because it required a sort of freak accident where one cell got inside another and became the mitochondrion, which broke the energy constraints that kept bacteria simple. That’s my understanding, at least; I’m going to read Nick Lane’s books to understand it fully.

13. Stephen Wolfram on P vs. NP. An extremely long and difficult piece, worth working through very gradually. I can’t claim to have had much success yet.

14. Alex Imas with some notes on why the micro and macro of AI productivity effects aren’t matching up yet. In other words, why haven’t we seen aggregate productivity impacts from AI? Imas basically believes that we’re early and that “we will be seeing AI showing up in the aggregate productivity numbers quite soon.” I mostly agree, though I’m surprised that the models are as good as they are and yet we still haven’t seen aggregate productivity impacts.

15. How societies lose capacity (paywalled). The decline of industry in the Western world was associated with the loss of a huge amount of tacit knowledge. “Retaining an economy’s ability to build large or complex infrastructure projects requires constantly exercising that muscle. … Knowledge economies are not ladders we climb once, but treadmills that will knock us down if we stop running.”

16. Steve Jobs really wanted to build an Apple Car. That’s from Tony Fadell, who worked for Jobs and then founded the smart home company Nest. Apparently Jobs’s idea was inspired by the Volkswagen, which I suppose Jobs would have seen all the time in the 1960s and ‘70s. Xiaomi, “the Apple of China,” succeeded in building a fantastic car in the 2020s; it would have been great if Apple did the same, though of course the project was abandoned. It’s been reported for a while that Jobs thought about a major car venture in the late 2000s. From a Bloomberg piece on the Apple car project: “In the wake of the 2008 financial crisis, with American car companies on the brink of failure, the Apple chief executive even floated the idea of acquiring General Motors Co. for pennies on the dollar.”

Thanks for reading, and enjoy the weekend!