Longreads
- Scott Alexander on Moltbook. For those unfamiliar: there's an AI assistant currently called OpenClaw (it was previously Clawdbot, which Anthropic didn't like, then Moltbot, which nobody really liked, and is now OpenClaw). Then someone made a bots-only Reddit-like social network, where various OpenClaw instances can chat. It's pretty fun to watch. But, as with many other instances of AI doing interesting things, it's important to note that the AI is not so much "commenting on a social network populated by AIs" as it is "producing what users expect a comment from an AI on an all-AI social network to look like." So part of the appeal is that it has the same basic structure as the punchline to a joke: collapsing some uncertainty ("would they all just ignore each other?") into something that's retrospectively unsurprising. It's also like a punchline in that it's funny; one of the bots had previously been asked a logistical question about practicing Islam, and answers questions by citing Islamic principles. One of the bots says it's identified a bug in some code, decided to keep it as a pet, and created an image of it. This might be what LLMs actually do when they're given open-ended tasks and feedback from other LLMs, but it's also what the LLM thinks users would think they would do—of course a conscious piece of code would keep some simpler code around as a pet! Why not!? But really, that's closer to having a pet rock; it's either inert or in the way. It's important to note here that LARPing based on social expectations (particularly when trying to create engaging content) is a deeply human thing to do, and to an extent every lawyer's behavior is influenced by the pop culture depiction of lawyers; every parent is influenced by the roles parents play in movies, TV shows, and ads; teenagers are keenly aware of how a typical teenager is supposed to behave. LLMs weren't directly instructed to copy this, but the way they're trained makes it inevitable. So when the bots talk about setting up encrypted communication channels, it's at least partly because their training data includes lots of LessWrong posts about how that's exactly what AIs would do under the circumstances. I've thought for a while now that we ought to be careful about what we say AI's incentives are, lest the AIs take us seriously, but for now they mostly seem to be putting on a fun improv show.
- Dean Ball has thoughts on AI and children. It's going to be very hard for new laws to add much value here, for some of the reasons Ball mentions (restricting chatbots runs into first amendment problems pretty quickly, and we already have tort law to cover negligence), but social norms will probably do a lot of the work. It's a good idea for parents to assume that their kids have access to LLMs, and use them, and that they'll be able to do so unsupervised; it's a big Internet, and one of the highest-bandwidth communications systems in the world is the whisper network whereby kids learn to get around technical roadblocks set up by their parents. But it's still possible to be completely clear and consistent on when using ChatGPT is helpful and when it constitutes cheating. After all, plenty of adults could do the equivalent of having ChatGPT write their essays for them, and they still opt out.
- Anton Howes' Age of Invention has a good corrective piece on Tudor economics. Ha-Joon Chang has said (and I've repeated) that Henry VII used protectionist policies to help insulate the early English textile industry from more established competitors in the Low Countries, but this turns out to be basically backwards: he actually enacted a series of policies that made it a relatively better deal to export unfinished cloth instead. It turns out that Daniel Defoe wrote about this in some of his (still quite entertaining and readable) pamphlets, but mixed up all the crucial details. None of this is to say that industrial policy can't work, just that in this case, it wasn't what worked.
- Adam Mastroianni questions the narrative that reading is in decline, noting that stats on average books read per person show only a modest decline, and that most of the reports about this involve college students. But that can be consistent with a slow decline: if today's college students have spent their entire lives with access to smartphones and tablets, it's possible that they will have formed habits that are hard to break, and that the impact of a decline in literacy will play out over a longer period. The more interesting part is his point that the Internet is a strange hybrid of orality and literacy; our communication norms are more informal, but you can read Usenet posts from decades ago, and many people can go through their Gmail history and find the first email they ever sent.
- Jason Zengerle profiles Tucker Carlson. One thing that's surprising in this piece is just how much Carlson bounces between highs and lows. He bet his early career on print, then realized TV was the future, and then got canned from a pretty good job at MSNBC and wound up with a much worse one at Fox, where Roger Ailes apparently thought that hiring people was a good way to keep them in close proximity in order to play weird mind games with them. Then he bet on Trump early, and turned into a key figure in the cottage industry of turning Trump's various pronouncements into a coherent ideology. (My theory of this is that while Trump doesn't have a great theory of the world, he has an incredible sense for other people's weaknesses, and that tracking who he attacks and how does give you a sense of his implicit ideology. But, compared to theorizing about longer-lasting belief systems, the Trump stuff will have a short shelf-life. Perfect for TV, but a book expounding on Trump's views will only be of interest to academic historians in a few years.)
- A read.haus reader shared a chat asking if it makes sense to securitize GPUs, datacenters, etc., which happens to be a topic I've spent a lot of time on recently. The core macro thesis here is that Nvidia (disclosure: long) wants to keep the market for AI infrastructure fragmented, so no one company can capture the entire value chain (any company that monopolized purchases of GPUs would be best-positioned to fund research into some really good ASICs). And second, the users of this compute are incredibly variable. Sometimes, everyone wants to Ghiblify their pictures with ChatGPT, sometimes, they want to pay Anthropic to let them keep a sentient-seeming abstract robot as a pet. Plenty of the demand is still for classic ML tasks like routing delivery trucks, predicting customer churn, targeting ads, trawling through logs to find performance or security issues, etc. All of this will lead to lumpy and unpredictable demand. Markets work best when both sides are fragmented, and between strategic considerations and the nature of the industry, that's happened. But one point that pushes back against this is that compute is hard to standardize, which makes it tough to package into a futures contract. It has the real-time nature of electricity, combined with the heterogeneous-variety-of-products problem that oil and agricultural commodities run into. Having just one of those traits is not a critical risk, given that the examples in the last sentence come from active futures markets. Having two would be tough, and probably create some overhead. Fortunately, many marvelous tools for dealing with messy information have recently become available, so this problem is more solvable than it used to be.
- In Capital Gains we tried something new: stories from financial history instead of breakdowns of finance topics. These are roughly as evergreen as an explanation of why EBITDA isn't a scam or how discount rates affect asset values, and there are many worth writing about. To start, two case studies of the USSR's surprisingly successful commodities trading operations.
You're on the free list for The Diff! Upgrade today for full access. You'll also unlock the full archive, including this week's posts about Michael Burry and Gamestop ($), how companies are reframing hiring, partly in light of AI ($), and the tradeoffs of Elon Musk merging all of his companies together ($).
Books
The Wages of Destruction: The Making and Breaking of the Nazi Economy: There's a pop history view of Nazi Germany where their policies initially led to high economic growth, and Germany ended up with a modern, mechanized army that they used to crush neighboring countries, until Soviet manpower and American military equipment finally turned the tide. This book is a helpful corrective: much of that story is completely backwards, though Germany did manage to put itself into a few situations where the part of the front where the fighting happened was where their newest tanks and best troops were. But the whole process was a lot more slapdash and incoherent than it might look.
One way to look at Germany's economic history in this time period is to consider the classic postwar growth model, applied to best effect in East Asia. That model was: use land reform to break up big agricultural estates and give the people working the land a bigger incentive to produce more food; take advantage of that increase in calories per acre to encourage labor-intensive export industries like textiles; use the foreign exchange from selling textiles to invest in heavy industry; move up the value chain from steel and chemicals to cars, machine tools, electrical equipment, etc., and probably level out somewhere in the upper reaches of middle-income status.
Germany in 1933 had many of the ingredients for that: a big and inefficient agricultural sector, some promising heavy industry, a great education system that could train the engineers and administrators necessary to run a modern economy. They also had a high unemployment rate, industrial overcapacity, and a housing shortage, all of which are helpful ingredients for a fiscally-dominant policy of deficit spending in order to promote investment in useful products (i.e. new housing) while stimulating demand. Unfortunately, they were run by a regime that had a number of crazy views, which a) precluded sensible policy, and b) made them a global pariah subject to retaliatory tariffs and constantly facing the risk that their raw material imports would be cut off. So, instead of betting on comparative advantage and benefiting from global capital flows, Nazi Germany followed an autarkic approach where they carefully controlled their imports in order to limit the leakage of hard currency.
Democracies are superficially a more politicized way to organize a country than totalitarian dictatorships with a cult of personality, but that's partly an illusion; dictatorships have internal politics, too, just of an opaque kind. The pragmatic thing for the German economy was to deemphasize farming, which was not going to be Germany's long-term comparative advantage. But farmers supported the Nazis, and the Nazi narrative celebrated independent small-scale farmers more than it did factory workers and engineering professors. The Junker elite, with their massive agricultural estates in Prussia, were also a key interest group. So the Nazis imposed various sweeping controls on both the prices of agricultural products and the ownership of farms, but didn't break up ownership of the big ones, and basically froze that sector of the German economy in time.
They did invest significant amounts in industry, but most of this was for aircraft, guns, and (especially) ammunition. So in the prewar period, the German economy grew fast, but household consumption stagnated: incremental output per hour meant that they could support a bigger and better-armed Wehrmacht while holding output constant, and that's what they chose to do.
Throughout the story, one of the pressing concerns is foreign exchange: Germany needed hard currency to buy imports, but had few ways to get it, and constantly risked credit crises when they funded deficit spending through the banking system. Reading about how they managed a program of massive infrastructure investment (the autobahn, fortifications on the French border) and a huge rearmament spree led to a sort of state-driven financial engineering equivalent to the demoscene, i.e. coming up with as many clever tricks as possible in the face of arbitrary constraints. Some of this involved finding places where the state could borrow, or could constrain consumption (they convinced a huge number of Germans to put down big deposits for cheap Volkswagens, which never actually ended up getting produced). In some cases, they recreated older forms of financial chicanery: at one point, Germany had defaulted on its debts, but still wanted to trade with its neighbors, and used reduced tariff barriers as an indirect means of paying them some return on their loans—this is exactly the structure that the Peruzzi and Bardi companies used in the 14th century in order to get around usury restrictions—if they lent a Duke 1,000 florins, instead of charging 8% interest they'd just get an 80-florin reduction in tariffs each year.
Internal economic policy had an element of insane whimsy to it that occasionally worked well (they built the autobahn) but often turned out to borrow from the future at a steep interest rate (the autobahn was funded in part by taking resources away from the Reichsbahn, the state railroad, which led to snarled transportation by the time hostilities started). For a while, Germany had an aggressive program of building new housing, but they tapered that off because the military needed so much steel and labor.
Constrained financial flows were both a cause and effect of a more literal, tangible view of the economy. Britain and the US could afford to think of a global market where they might source different raw materials from different places, build some manufactured goods themselves and trade for other things they needed. Nazi Germany tended to look at things in a more direct way, where their food supply, for example, was not whatever they could buy on the global market but whatever they could produce within their borders. (Private Empire, a book at Exxon, has a fascinating scene in which American and Chinese policymakers have exactly this disagreement—the Americans are, willfully or not, oblivious to the fact that a global market for raw materials exists only for the countries America permits it to.) Conquering a substantial portion of Eastern Europe, and depopulating it, was the only way to make their model work.
There's a weird dynamic with how Nazi Germany talked about war internally and externally. Their basic view seemed to be that Germany could be invaded at any time by its larger and better-armed rivals, who were overreacting to Mein Kampf's explicit claim that Germany needed to invade its neighbors. So they had no choice but to go ahead and invade their neighbors. Their early war plans were fairly long-term, with lots of plans to be at full strength in different parts of the military in the mid and even late 1940s. Then they realized that neighboring countries were making the flagrantly provocative decision to also arm themselves in order to fight a belligerent neighbor with a massive military. It's always striking to look back at Germany's history over this period and see it as a sequence of lopsidedly bad bets that all happened to work—every threat, annexation, and invasion all the way through Barbarossa had some element of a bluff, and because many of these decisions were made hastily and under a great deal of uncertainty, they came as a surprise because they just didn't make sense. The last big gamble was that Germany could secure a food source, oil, and manufacturing capacity by invading the USSR, and this had to work because there's no going back from invading a country with which you've recently signed a non-aggression pact. The only plausible endgame was for Germany to win a quick war and become an autarkic hermit kingdom. And that invasion had to happen very fast, because there simply wasn't enough transportation infrastructure to supply the German army once it got deep into Russian territory. There just wasn't a plausible compromise between a Reich that annexed most of Western Russia and a Reich that completely collapsed.
The sense this book gives is that if you run the simulation 100 times from 1933, almost all of them lead to an outcome where the war either doesn't start or ends earlier. Hitler could only really function in an unstable world where he could keep everyone guessing, and as soon as there was any sense of inevitability, his bluffs would start getting called. Germany wasn't materially capable of winning a world war, and the geopolitical situation at the time meant that there was little chance that they could fight a more limited conflict than that.
Open Thread
- Drop in any links or comments of interest to Diff readers.
- The Diff has previously written about the East Asian economic miracle, North Korea’s weird blend of Stalinism and anarcho-capitalism, and crypto. What other economies-very-different-from-ours should be next?
Diff Jobs
Companies in the Diff network are actively looking for talent. See a sampling of current open roles below:
- A transformative company that’s bringing AI-powered, personalized education to a billion+ students is looking for an elite technical product manager to translate cutting edge learning science research and breakthroughs in machine learning into a platform that enables students around the world to leverage machine intelligence to effectively develop their own intelligence. Experience with full technical and delivery ownership on a revenue-generating product preferred. (Remote)
- A Founders Fund backed, decentralized AI infrastructure company that’s solving the compute shortage by offering hardware procurement, financing, software, and customer demand that allows any data center operator to become an AI factory/turn data center infrastructure into cash flow is looking for a capital markets associate to help facilitate GPU financings and support the company’s growth into an institutional private credit platform. You’d be helping underwrite a multi-hundred-million-dollar loan pipeline and encouraged/expected to find and pursue new angles for corporate development. If you have experience with data center financings or M&A, please reach out. (Remote)
- A startup is automating the highest tier of scientific evidence and building the HuggingFace for humans + machines to read/write scientific research to. They’re hiring engineers and academics to help index the world’s scientific corpus, design interfaces at the right level of abstraction for users to verify results, and launch new initiatives to grow into academia and the pharma industry. A background in systematic reviews or medicine/biology is a plus, along with a strong interest in LLMs, EU4, Factorio, and the humanities. (Toronto, Remote)
- A frontier investment firm is looking for someone with exceptional judgement and energy to produce a constant feed of interesting humans who should be on their radar. This person should find themselves in communities of brilliant people hacking on technologies (e.g. post-quantum cryptography, optical computing, frontier open source AI etc.) that are still well outside the technological Overton window. You will be responsible for identifying the 50–100 people globally who are obsessed with these nascent categories before they are on-market, then facilitating the high-bandwidth IRL environments (dinners, retreats, small meetups) that turn those connections into a community. (Austin, NYC, SF)
- A leading AI transformation & PE investment firm (think private equity meets Palantir) that’s been focused on investing in and transforming businesses with AI long before ChatGPT (100+ successful portfolio company AI transformations since 2019) is hiring experienced forward deployed AI engineers to design, implement, test, and maintain cutting edge AI products that solve complex problems in a variety of sector areas. If you have 3+ years of experience across the development lifecycle and enjoy working with clients to solve concrete problems please reach out. Experience managing engineering teams is a plus. (Remote)
Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up.
If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority.