Longreads + Open Thread

Longreads

• Dwarkesh Patel and John Collison interview Elon Musk about orbital datacenters and much more. I did not personally come away from this sold on the idea that orbital datacenters will be the cheapest place to do AI by the summer of 2028, but my confidence interval is wider. (A handwavy case is that there are some costs that are structurally fixed because of physical limitations, and others that have some kind of experience curve, and in the very long run everything moves to whatever format minimizes those irreducible costs, even if you start out at a tough part of the experience curve.) One theme that keeps coming up is that Musk is always looking for the limiting factor, and solving it. Sometimes that means switching from carbon fiber to steel, sometimes it means switching from running businesses to lobbying for them, sometimes it means getting access to public equity markets as a capital source, and sometimes it apparently means launching GPUs into orbit.
• Tyler Cowen interviews Andrew Ross Sorkin on the crash of 1929. (See this Diff review of the book for more.) It's a fun interview, especially since Cowen starts out by asking: weren't the investors who bought at the peak basically right? In the aggregate, they were right to be bullish about America, but one reason stock markets crash is that capital is misallocated within them even if it's smart to allocate capital to them. One of the big problems with the market back then was that banks traded at astronomical price/book values even though they weren't making all that much more money than usual. But another, deeper problem was that the earnings yield on equities was lower than the yield a company could get from lending out its excess cash to margin borrowers. So you basically had people paying 6% a year to buy Chrysler at 25x earnings, when those earnings were, increasingly, coming from the exact loans used to support the stock. It's not necessarily irrational to buy a company and get negative economic carry (i.e. earnings are 4% of what you pay, and you pay 6% to borrow it), if those earnings are going to grow and if you don't borrow so much that you're forced to sell at the bottom. That condition did not obtain at the market peak. The rest of the interview is quite fun; Sorkin points out that the CEO archetypes back then were not so different from the ones today—there's even an auto executive who gets involved in politics and also has some shady stock market dealings!
• Eric Jang: As Rocks May Think. It's a good exercise to read state-of-the-AI-world pieces like this and remember that, just a few years ago, you could read the same thing as a piece of speculative science fiction and think "I wonder if the 2050s will really be like that." One major point the piece makes is that we've recently had to significantly broaden the set of things that count as computationally feasible, and that we'll probably continue to do so.
• And, on the topic of AI, Nicholas Andresen has a fun piece on how AIs sometimes hide their reasoning. Reasoning models basically think out loud, but sometimes they think their way into something they're not supposed to do. Some of them backtrack, some of them deceive users, and some of them write semi-deceptive descriptions of what they're doing. As base models get more powerful, it's entirely possible that they'll learn to reason in a Straussian way, and that we'll have to detect this by measuring the information content of the reasoning relative to some proxy for how much reasoning the underlying task requires. It'll be a fun arms race, but in the end this kind of alignment will look similar to how it looks when human intelligence is getting aligned: the frictional cost of deception means that high-trust groups will outcompete the low-trust ones.
• And speaking of which: Patrick McKenzie uses the recent Minnesota daycare kerfuffle as a jumping-off point for an essay on fraud. One of the conveniently hormetic side effects of the Internet is that a) it's a full of opportunities to do various kinds of crimes, b) these crimes tend to be scalable, and thus c) they form statistically recognizable clusters. So people who've dealt with fraud online have both a big sample size about what fraudsters are like and a more detailed model of how they behave. They have much to teach the slower-moving world of brick-and-mortar fraud detection. One important point he makes is that fraud countermeasures are basically a tax on all of us, and that when fraud takes the form of ripping off welfare, the net result is a higher time burden on welfare recipients, less money for them, and less political willingness to give them more. Minnesota's taxpayers will be financially okay if more of their money has been siphoned away to bad actors (though it's reasonable for them to be furious about this), but the people who suffer a disproportionate burden are the potential legitimate beneficiaries of the programs being exploited.
• On Read.Haus, a reader asks what I think of full-reserve banking. The theoretically clean answer is that it's fine for banks to engage in maturity transformation over short periods—using demand deposits and short-term CDs to fund a portfolio of working capital loans that have a slightly longer duration. The practical answer is that, for various economic, regulatory, and historically-contingent reasons, we expect the banking system to sometimes fund very long-term loans with short-term demand deposits, and then contort assorted regulations in a way that makes this possible. And the reason for that is that there's more demand for long-duration loans (mostly for houses) than there is supply of long-duration lending, most of the time. This is not always true; if your society is getting older and not having kids, there will be a large population that's saving for retirement and wants predictable income, and there will be plenty of houses to go around. But that situation is unstable for other reasons. In an ideal world, long-duration assets would be owned by life insurers, pension funds, and households saving for retirement, and this market would be mostly separate from short-term lending backed by demand deposits. But at this point, the banks are pretty good at what they do.
• In Capital Gains this week, another history lesson: that time some reseearchers realized that Nasdaq market-makers were colluding to keep spreads wide. The ultimate lesson here is that if you're part of a cartel that keeps transaction costs artificially high, you should definitely defect, because there's more money in doing low-margin trades at massive scale, and there's even more money in being the first to operate this way.

Books

The Amoeba in the Room: Lives of the Microbes: If you like cosmic horror, at some point you'll end up telling one of your friends to check out H.P. Lovecraft, but if you like having the respect of your friends you'll find some tactful way to warn them that Lovecraft expresses intense antipathy towards basically anyone who isn't a WASP like him (especially if they're WASPs-but-not-like-him). So, with that in mind: this book is fascinating, but periodically goes out of its way to express contempt towards minorities, specifically all of humanity. It's a strange experience to read a book by a unicellular supremacist, but that's what The Amoeba in the Room is.

But if you can set aside the author's seething bigotry, it's a fun work. The core thesis is that our model of biology is partly a legacy of how limited our observations were. It was just a lot easier to study pigs and trees and other human-scale life before the microscope, and every new discovery gets slotted into existing models. But there's a lot more action at the single-cell level, and much more genetic diversity.

But to even talk about that risks accidentally reifying models that only apply at larger scales. For example, we tend to think of leaving things as discrete objects, and even the tricky cases aren't that hard: even if bees need a beehive, you can think of them separately. But when bacteria can share DNA, either by sharing plasmids or by viruses doing the same thing for them. And there's incredibly aggressive selection: the book cites an estimate that every day, viruses kill 20-40% of bacteria.

The Amoeba in the Room ends up providing some sympathetic explanations for the model it's complaining about. At the multicellular level, you can isolate a species from its environment and provide it with unnatural replacements in order to get a rough idea of how it behaves. But it's hard to create a bacterial zoo because the small-scale ecosystem is so complicated, with so many niche players and so much change due to short-term fluctuations; every wave changes the environment in the ocean, every time an earthworm passes through some dirt, it affects which cells will thrive there and which ones suddenly find themselves in a deadly environment.

It's a good book as a whirlwind tour of the invisible world of microbes, by someone who likes them a lot more than he likes us.

Like: The Button That Changed the World: A few weeks ago, The Diff asked readers to recommend profiles of particular products, rather than companies. This book is a fun history of the once-ubiquitous like button, though by the end it's clear why the button is in decline.

The right way to situation the like button in history is that it's an important point on the continuum between the Internet as a peer-to-peer communications medium and the Internet as infrastructure for one-to-many broadcasts. An early Internet user was almost certainly emitting a constant stream of hand-crafted bits of information—replies to emails and Usenet posts, blog entries, comments, etc. But as the Internet gets bigger, two things happened at once: first, the incremental new user was someone who was less excited about typing a lot (all the frenetic writers who longed for an audience had already joined). And second, the scale of the internet and improvements in recommendation systems meant that there was more competition for attention, and there were more interesting things to consume passively.

As we moved along this continuum, there was a lower likelihood of someone going out of their way to write a lengthy message, but there was also growth in the measurable context around any given user-content interaction. If you read a news story online in 1997, CNN.com might have placed a cookie on your machine already, so they might have had the information necessary to know that you were interested in the Middle East but didn't care much about sports, or vice-versa. But they weren't going to customize the homepage for you. Whereas if you read a news story in 2017, there was a good chance that you found it through a social media feed, and that the social media site in question had an extensive psychographic profile that allowed them to show you the 0.01% of content posted by your first- and second-degree connections that you actually cared to see. In an environment like that, the like button is a very valuable bit, a way to say "Yes, good job, you showed me what I wanted to see." But eventually, other signals get robust enough that it's superfluous: looking at what you tap, or even what makes you stop scrolling, or which kinds of stories you read to the end and which you bounce off of in a second—the like ends up being just one bit among many.

The best parts of the book are the early chapters on the history of the like button as a feature (including the javascript that Yelp used to implement it for the first time) and a social history of the thumbs-up gesture, which comes to us by way of 19th-century paintings slightly misinterpreting Roman sources. This gesture made its way through evangelical preachers' sermons, pilots looking for a way to communicate in a noisy environment (yet another case where bits are scarce!), then to hot rod culture and then to everyone. (As with social media, there's a glide path of adoption where visual slang starts out with high-status people like pilots, then to medium-status people, then to everyone.)

The button got implemented in different ways by different companies, sometimes as an anonymous feedback tool and sometimes as a way to offer the minimum viable yes-I-saw-this. Mark Zuckerberg actually spent years vetoing various implementations, mostly on the grounds that it cannibalized things like comments and sharing, before conceding that it's better to get a weak signal than no signal.

But after that, the book is basically fighting a rear-guard action against the decline of the like button's importance. It used to be a big deal, but it can be gamed, and other signals are cleaner. As the authors note late in the book, likes were a good way to bootstrap a model of preferences before people stopped using them so honestly. And now, it's great to own a business that used to rely on the like button, but ideally one that doesn't need it now.

Open Thread

• Drop in any links or comments of interest to Diff readers.
• So, should we do AI in space? What has to happen to make that viable?

Diff Jobs

Companies in the Diff network are actively looking for talent. See a sampling of current open roles below:

• Series A startup that powers 2 of the 3 frontier labs’ coding agents with the highest quality SFT and RLVR data pipelines is looking for growth/ops folks to help customers improve the underlying intelligence and usefulness of their models by scaling data quality and quantity. If you read axRiv, but also love playing strategy games, this one is for you. (SF)
• YC-backed startup automating procurement and sales processes for the chemicals industry, which currently relies on a manual blend of email, spreadsheets, legacy ERPs, etc. to find, price, buy, and sell over 20M+ discrete chemicals, is hiring full-stack engineers (React, TypeScript, etc.). Folks with exposure to both startups and bigtech, but also an interest in helping real-world America with AI preferred. (SF)
• A hyper-growth startup that’s turning the fastest growing unicorns’ sales and marketing data into revenue (driven $XXXM incremental customer revenue the last year alone) is looking for a senior/staff-level software engineer with a track record of building large, performant distributed systems and owning customer delivery at high velocity. Experience with AI agents, orchestration frameworks, and contributing to open source AI a plus. (NYC)
• Ex-Bridgewater, Worldcoin founders using LLMs to generate investment signals, systematize fundamental analysis, and power the superintelligence for investing are looking for machine learning and full-stack software engineers (Typescript/React + Python) who want to build highly-scalable infrastructure that enables previously impossible machine learning results. Experience with large scale data pipelines, applied machine learning, etc. preferred. If you’re a sharp generalist with strong technical skills, please reach out.
• Fast-growing, General Catalyst backed startup building the platform and primitives that power business transformation, starting with an AI-native ERP, is looking for expert generalists to identify critical directives, parachute into the part of the business that needs help and drive results with scalable processes. If you have exceptional judgement across contexts, a taste for high leverage problems and people, and the agency to drive solutions to completion, this is for you. (SF)

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.