Longreads + Open Thread

This issue of The Diff is brought to you by our sponsor, Perk.

Longreads

• Luke Emberson and David Roodman at Epoch AI have a report on rapid deflation in the cost of AI. They're looking at the price of accomplishing particular tasks, not the cost of tokens, and they spend a lot of the report talking about the various difficulties in measuring this kind of thing well. But the upshot is that machine intelligence is getting about 47% cheaper each quarter, and deflating faster than any other technology with similar impact. The drivers of this are more opaque than they are for some of the other characteristics of the AI industry; models are born competing on capability and live for as long as they can compete on price. But labs will sometimes price based on strategic considerations rather than pure unit economics; they might be able to afford suboptimal margins on some kind of inference more than they can afford to let customers churn off. But that probably explains more about the timing of particular price changes than the overall trend.
• It's fun to compare how different countries identify talent and then funnel it into prestige jobs. Here's a fun example: Spain's business elite comes from the three main finishing schools of the Jesuits, Opus Dei, and McKinsey. There's a first-order sense in which it's quite unfair for a handful of organizations to have so much influence over who does and doesn't get a CEO-track job. On the other hand, this creates an incentive: once a vetting institution has a reputation, it has an asset, whose value it's going to try to preserve. This tends to exaggerate whatever traits made it a valid selector in the first place, both because of who selects from it and who selects into it. So these institutions are an important way that society makes a little progress in solving the notoriously hard problem of allocating talent.
• Lily Ottinger covers the World Nomad Games in Kyrgyzstan for ChinaTalk. It's a wonderfully surreal piece—an organizer was thrilled to learn that the author, who is from Kansas, was a fellow steppe dweller. It has every imaginable nomad-themed athletic competition (yes, the horse archery contest includes Kyrgyz and Hungarian styles). Much of the piece is an analysis of China's efforts at cultivating soft power. On one hand, they're pretty clumsy—America, a nation of TV and national brands, is full of people who voluntarily propagandize for American norms and values as a side effect of maximizing shareholder returns, while Chinese companies have done better competing on other axes and the government has more control over the persuasion industry.
• Jack Despain Zhou on Nikole Hannah-Jones. For background, Hannah-Jones chose a school for her daughter based on her (NHJ's) beliefs about education. Despite having kin in the game, she missed the fact that her daughter wasn't learning much at all. (Fortunately, her grades were just fine.) One feature of political beliefs is that they motivate people to make decisions whose outcomes they're motivated to ignore. This is something they have in common with other kinds of chosen identities; if you decide to become a vegetarian, or move to a new city, or convert to a religion, you're adopting an identity and probably not trying to aggressively benchmark its performance against the alternatives. If anything, a bit of inconvenience makes the decision feel more credible. But it's a decision whose cost was borne by her kid; one of the things missing from discussions of education as either a laboratory for ideological experiments or as an irrelevant factor whose effects will attenuate in early adulthood is that childhood is a substantial chunk of people's lives, and subjectively (in terms of share of neuroplastic time) an even bigger one. It's commendable to live in accordance with one's principles, but that's in the same sense that the people who did banzai charges in the Second World War had to be very brave indeed to convince themselves to do something so pointless. It's entirely possible to build a well-meaning system in which everyone has an incentive to do counterproductive things. In fact, if your system is a compromise between enough constituencies' wishlists, that's almost certainly what you'll get.
• Kevin Gee of A Letter a Day found a fun speech by Ray Sidney, employee #3 at Google. (Disclosure: long GOOGL.) Sometimes there's a narrative that baby boomers fell backwards into unimaginable wealth, where they got to spend a few years backpacking and dropping acid before getting a job as VP of Everything for Pfizer and then having to save up for literally weeks before they could make a down payment on a house. If that kind of story makes you feel envious, then under no circumstances should you read about the guy who was hanging out with his friend's boyfriend, Sergey Brin, who casually suggested that he get a job that means that no descendant of his will ever need to work. This piece has some great context on what very early Google was like, including the revelation that their growth rate was actually disappointing: they were gunning for a majority of the search market and an IPO by the end of 1999. Sidney also has a fun story about how he shipped an update straight to prod, then realized he hadn't tested it and that search results were suddenly terrible. Every company with a cautious, box-ticking engineering culture is a company that learned some exciting lessons.
• This week's Capital Gains is on the question of how much you can adjust EBITDA, and more generally, how to use companies' custom metrics. Accounting is a conservative profession, so adjustments to standard accounting tend to be optimistic. This doesn't necessarily mean that they're less accurate, but offers a lot of room for inaccurate financial reporting.
• On Read.Haus, a very fun question: for contracts that have zero net supply, like insurance and derivatives, do rich people just face that risk and earn that risk premium by default? They do! This is one of the roles rich people perform in an economy: they hold a lot of some category of risk, reducing everyone else's exposure to it. They have to be paid well—nobody is seeking out this risk, just the attendant returns. But if they're in any way part of society, taking on this risk means that they'll occasionally lose a lot of money, and some of the social status that comes with it. For a handful of risks, there are financial intermediaries who either get paid to ride it out (reinsurance) or get paid to distribute it to whoever is least sensitive to it (market-makers), but for the long tail of other risks, the balance sheet of some rich person is the backstop.

A Word From Our Sponsors

Travel and expenses. Now on one platform with Perk.

A booking in one platform. A card transaction somewhere else. A receipt in an inbox. An approval in another system.

Every handoff creates another task.

For finance teams, the real cost of disconnected systems isn’t just the software. It’s all the manual work needed to connect the dots, from chasing information to matching transactions and reconciling spend after the fact. Fewer handoffs can mean less admin for employees and a clearer picture for finance.

Books

Empire and Communications: History books often have maps showing you roughly who ruled where and when. Which is convenient, but the lines on those maps mean different things in different times and places; Kim Jong Un and Donald Trump are two very different kinds of in-control with respect to their territory, and while Louis XV and Emmanuel Macron both ruled France with pretty similar borders, they had different levels of control. Over long periods, one of the critical variables is how the person in charge's commands get communicated, and how easy it is to verify that they were carried out.

Empire and Communications takes this question very seriously, and gives a quick summary of human history mostly through the lens of official communications. It's a surprisingly fun model. For example, it implies that countries where the favored medium takes a long time to build and lasts forever (so, enormous stone monuments) will tend to be stable. It would take a lot of work for a pharaoh to build some symbol of power that was bigger than the pyramids, whereas treating the pyramids as a symbol of the durability of the state.

His fundamental model is that durable communications media create a static but decentralized society, that’s stable but brittle. One of the limits on bossing people around is how hard and inconvenient it is to send a message. So empires that have faster but more transient communications media, like paper, can exercise finer-grained control, and react more continuously. But a faster-decaying medium where it’s easy to produce and distribute new content is one where ideas have a shorter half-life. So empires have had to balance between the permanent but potentially obsolete and the fast but transient.

One thing he explores is that some governments use cheap communication generally as a tool of soft power. The Ptolemaic dynasty subsidized papyrus while taxing exports of it, partly so their own bureaucracy would function better and partly because that meant subsidizing schools and libraries to make themselves the center of the Mediterranean intellectual universe. This has an interesting feedback loop: the language and writing system of the central government will spread throughout the hinterland, and to neighboring countries, and one effect of this will be that the most established legal, literary, and philosophical canon will tend to line up with the way those rulers think, so it's self-reinforcing.

Innis delivered these lectures in 1948; he briefly talks about radio towards the end, barely mentions movies, and never discusses television. This makes his model a lot cleaner, because any analysis of media after around that time has to contend with how profoundly powerful TV has been. Politicians and advertisers in the decade after this lecture were routinely astounded by just how much of people's worldviews came directly from the TV set, and how keeping something off TV meant that to the average person, it basically wasn't happening. So it's a bit like reading something about pandemic preparedness written in 2019; you'll get a model that isn't distorted by having one big datapoint that swamps everything else.

Open Thread

• Drop in any links or comments of interest to Diff readers.
• Any book/essay recommendations along the lines of Empire and Communications, i.e. about how different media technologies affect state capacity, coordination, etc.?

Diff Jobs

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

• A top prop trading firm is looking for people who combine exceptional quantitative ability with the judgment and relationship instincts to build critical financial partnerships and lead high-stakes negotiations. Candidates may come from markets, finance, or further afield; what matters is this combination of abilities, strengthened by the perspective that comes with experience. (NYC)
• A top prop trading firm is looking for people with exceptional strategic thinking and quantitative skills to help the firm model and manage its financing risk and strategy. Open to candidates from a range of analytically demanding fields.(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.
• Series A, Thiel-backed team building full-stack software, hardware, and chemistry to end water scarcity, is looking for an ambitious product engineer to help build operating platform and internal tools that weather modification operators can use to make weather, geospatial and flight data useful to plan, monitor, and review campaigns. If you’re excited to own product outcomes through deployment and are strong at turning messy problems into good product (using Typescript and React), this role is for you. (Los Angeles)
• Ex-Citadel/D.E. Shaw team building AI-native infrastructure to turn lots of insurance data—structured and unstructured—into decision-grade plumbing that helps casualty risk and insurance liabilities move is looking for a data scientist with classical and generative/agentic ML experience. You will develop, refine, and productionize the company’s core models. (NYC, Boston)

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.

And: we're now actively deploying capital into early-stage companies through Anomaly. Our focus is on defense, logistics, robotics, and energy. If you'd like to chat, please reach out.