The Mythical Agent-Minute

In this issue:

• The Mythical Agent-Minute—When we talk about "growth," we're implying some measure of what's growing. The earlier we are in that process, the simpler the measurements can be, but the lack of meta-measurement leads to disaster.
• AI, Globalization, and Comparative Advantage in Lobbying—When demand is tight, it's good to rethink who is and isn't a viable trade partner.
• Media M&A—The newsletter boom follows the usual path of new media: fragmentation, then consolidation.
• Retaliatory Tariffs—A digital services tax is functionally close to a tariff on the US.
• Capex—Prices can't rise forever.
• Pivots—Yet another crypto treasury strategy.

Talk to this post on Read.Haus.

The Mythical Agent-Minute

You can tell Fred Brooks did a good job with The Mythical Man-Month, because over the ensuing near half-century, nobody has managed to displace it with an essay that makes the same point with a more politically correct title.^[1] Brooks' key point is that a manager might be tempted to look at a project that takes ten software developers a year to finish, and think that twenty developers could do it in six months, and they might be especially tempted to think this way if the deadline is in, say, month seven. But it turns out that hiring new people means having your existing people take some time out of their job to interview them, and then to get them up to speed, and then, of course, the earlier team has been in the weeds of the project for long enough that they aren't thinking about the accumulated mistakes, hacks, TOFIXes, etc., and so their new teammates will either break things or introduce subtle bugs. The net result is that at the six-month mark, they may well announce that, thanks to their heroic efforts, the project will probably be done only fifteen months after the new hires joined.

Much has changed about software development since then, but one thing that hasn't changed is that it's frustratingly hard to find good metrics to measure developer productivity and predict timelines. And a disproportionate fraction of that change has been in the last six months. That's not just because agentic software development is a pretty magical technology, but because it fits in so closely with the way people worked before. Lots of people have produced code by writing a detailed natural-language description of what they want; those are the people developers reported to in the org chart. And if you're building anything of more than minimal complexity, it's a pretty good idea to have a prose description of what you're doing and how it's going to work. If Codex or Claude Code can churn for a while and turn that early, simplified sketch into a working prototype and then help you iterate from there, it's a radical transformation. And once you have that capability, why not use it to the maximum? Plenty of people have backlogs of projects they've considered that are suddenly worth doing (if yours are written down somewhere, you can even ask Codex to review your idea list and propose some first projects). If you have a tool that's not just rate-limited by ideas, but also can help you organize and refine your ideas, shouldn't you use it until you're out of ideas?

One of the fastest cycles in tech trends was the months-long cycle from tokenmaxxing to actually worrying about AI usage. As a metric, token use is like lines of code, in two ways:

1. It's the closest thing to a single-number metric of total effort expended. It's clumsy, limiting, but at least tells you something.
2. It is, of course, extremely gameable. Use the highest reasoning setting when you're asking the model what you should have for lunch, rewrite every system in ten different languages to see which one's fastest, maybe translate the documentation into Linear B. Even if you're earnestly trying, it's easy to spin up wasteful requests; one way to tokenmax is to just do things wrong the first time, and another is to resolutely refuse to read, review, and edit your prompts before firing them off.

Using tokens as a pure metric is a bit like the way some command economies try to measure output in terms of physical goods, and then set quotas accordingly. If you're going to hit your wheat output number, you need so many tractors, which require a certain number of tons of steel which itself requires some quantity of iron ore. It's not impossible to centrally plan this—every profitable vertically-integrated company has partially solved this, though Soviet central planners might have noticed that Henry Ford did not actually succeed in building a fully vertically-integrated car factory and still needed legions of specialized suppliers.^[2] If a country has a basically preindustrial economy, and has a clear path to catch-up growth, this kind of clumsy quota system is actually a reasonable start: median electricity consumption per capita is a better indicator of wellbeing in Kenya than it is in an American suburb. You don't know for sure how many tractors you will ultimately need, or how long they'll last, or which parts will wear out faster and need a bigger stockpile. But you can probably eyeball the rough quantity of raw materials, and sketch out where the obvious railroads should go.

But the goal of that process is not just to plug the same assumptions in and repeat it over and over forever; this kind of approach works when the upside is so obviously high that it can tolerate a huge amount of uncertainty, but actually doing it—building a steel mill somewhere, a factory somewhere else, and a railroad between them—you're also refining all of these details, adding more features to your model, and, in time, finding places where modeling everything in advance simply doesn't work and you need to give people flexibility and an incentive to make the right decisions.

If you don't make that transition, but do want to push forward economic growth, and you're measuring growth in kilograms of materials used, you'll end up with deeply perverse incentives. Various public works projects in China fit, as does the evaporation of the Aral Sea or hunting down the remaining whales that the capitalists hadn't gotten to yet. At some point, your key performance indicator needs a whole vector of offsetting metrics based on information that's hard to collect, but that in a market system can be partly inferred from the behavior of the people who have it.

The downside to token-maximizing is that companies waste a ton of money; the downside to token-minimizing is that they're wasting a ton of opportunity cost, and will have to develop a sense for the exact efficient frontier between human and AI outputs. Which is a particularly tough problem because that frontier is going to vary from person to person and from task to task.

It will be interesting to see how companies experiment with this, particularly within companie. Maybe routing everything through cheap open-source models is actually good enough for most problems they need to solve. Maybe the right move is to set higher token budgets closer to the P&L, and then use those teams to figure out the optimal setup. When a few companies had token-usage dashboards up, The Diff argued that this was less wasteful than it looked because people would ask about high token usage and get an answer that was either impressive or just made that spending sound like an expensive mistake. Maximizing token count is, directionally, the right thing for most companies that are involved in knowledge work, because all of us are getting more sophisticated about using AI all the time. Budgets should reflect the fact that people are offloading more work to AI, and weaving more AI into the rest of what they're doing.^[3] In retrospect, technologies are under-adopted early; in my own life, I regret that back when AOL limited you to a maximum of twenty hours a month, I probably had a few months when I didn't use that up. But the Internet in 1996 was an incredibly informative place—you could temporarily immigrate to a microcountry where the bestseller list included Larry Wall's Programming Perl. That was the future! It was a good idea to move there early.

1. Tech companies can afford to be on the cutting edge here, since swapping one word for another and changing nothing else is something computers are pretty good at. Switching from "whitelist" to "allowlist" is a sed one-liner, if you're feeling courageous. ↩︎

2. The early Soviets, being orthodox Marxists in good standing, were properly impressed with Henry Ford's achievements, and not just because he'd helpfully gathered so many means of production for the workers of the world to seize. They signed an early deal with Ford to build a factory in Russia, and Stalin praised Henry Ford, as did Khrushchev. ↩︎

3. My standard setup for a data analysis project right now is Codex writing code, a ChatGPT session talking about the general shape of the product, and a dashboard that starts out mostly focused on the internals of what I'm building (whether or not the data's clean, how slow some filtering process is), and only later has anything about the actual output. But it's early, and dashboard-driven development is hopefully not the endpoint. ↩︎

Diff Jobs

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

• 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
• Ex-Citadel/D.E. Shaw team building AI-native infrastructure that turns lots of insurance data—structured and unstructured—into decision-grade plumbing that helps casualty risk and insurance liabilities move is looking for forward deployed data scientists to help clients optimize/underwrite/price their portfolios. Experience in consulting, banking, PE, etc. with a technical academic background (CS, Applied Math, Statistics) a plus. Traditional data scientists with a commercial bent also encouraged. (NYC)
• Lightspeed-backed team building the engineering services firm of the future is looking for founding members of technical staff excited about working alongside civil engineers to translate their domain expertise into the operating system that powers the next era of great American infrastructure. If you’re an engineer with strong product intuition, who's energized by access to users, and excited by the prospect of transforming how we design and construct our built world with frontier AI, this is for you. (NYC, SF or Remote) 
• 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)
• 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)

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.

Elsewhere

AI, Globalization, and Comparative Advantage in Lobbying

The staffer who turns a broad idea like "let's sanction Chinese tech companies tied to the PLA" into a list of specific entities to sanction does not have perfectly-aligned incentives. They're implicitly trying to minimize the rate of career-limited false-positives or false-negatives, so they'll tend to exaggerate whatever direction policy is already taking. If you rather miss seamless trade between the US and China—if you're Apple, for example, and your previous model was to sell products assembled mostly in China to a userbase disproportionately located in America—then you'll have a particular incentive to point out instances of this. Apple is asking for permission to buy memory from a Chinese company, CXMT, that's on a Department of War blacklist ($, FT). It's an interesting example of modern political economy: given Trump's economic interventionism, Apple has a lot at risk if this ends up being a bad choice of suppliers. The Trump administration knows this. So, for now, companies can ante up reputational and regulatory risk in order to push regulations in their preferred direction. It's fine if they do a little of this; the risk shows up when companies start to specialize in it.

Media M&A

The Bear Cave, linked a few times on The Diff, has been acquired by hedge fund/investigative journalism hybrid Hunterbrook Media, also previously written about here. Congrats to all involved! Two points:

1. It's harder to sell short-selling research than other kinds ($, Diff), but in a model like Hunterbrook's, where they trade ahead of their ideas, they can capture a wider range of price drops. A hedge fund might be very interested in getting one or two reports a year about a company that's completely cooked the books, but indifferent to getting a pitch a week about someone who has a bunch of negative customer reviews and salty ex-employees on Glassdoor. But that kind of thing is a valuable signal! It's a leading indicator of revenue growth—you don't write the mean review until after the sale was made—just a noisier one. If you're monetizing research by trading on it, it's easier to effectively use ideas of different magnitudes.
2. The Diff theory of media is that when distribution changes, there's temporary alpha in mastering that distribution channel and more permanent upside from owning content. The rise of newsletters had a step-function increase after Covid, some of which has held up since. This was a notable distribution shift. And, like many before, it will be initially decentralized but eventually consolidate. That's been happening for a while, but mostly gets announced as a new hire rather than an acquisition. But acquisitions will come; there are plenty of kinds of media business that have some minimum viable headcount, but few that have seemed to reach a maximum.

Retaliatory Tariffs

Donald Trump has threatened 100% tariffs on countries that impose digital services taxes. Which is directionally reasonable, even if the magnitude is not. A digital services tax functions quite similarly to a tariff, and it's a tariff on an industry where America is the dominant exporter. In a more free trade-oriented world, this made it a popular kind of tax, one whose accounting incidence fell entirely on people in other countries. But in a world where trade is a series of deals rather than a default, that tax is just one more item up for negotiation.

Capex

Samsung and SK Hynix plan to invest $520bn in chip manufacturing, while SK Group, GS Group and Naver say they will spend $360bn through 2035 on datacenters. Sometimes, globalization is a matter of taking price signals seriously: memory prices are rising fast enough that there's a clear roadmap for demand, and buyers may be willing to commit further in advance. But someone still needs to add the capacity. The numbers involved, especially on the chip side, are a good reminder that semiconductors have been a cyclical business for a very long time, and even in periods of high growth, supply eventually overshoots.

Pivots

Last week, we wrote about how Microstrategy is learning what countries with pegged currencies learn: that raising rates high enough to sustain a peg often signals to the market that this rate increase is unsustainable. Today, they're conceding this. The pitch for their STRC product is that they issue it at $100, and set a dividend that keeps the price at $100. On Friday, the dividend was 11.5%, and STRC traded at about $75. So, to maintain the peg, they'd need the dividend to be about 15%. They chose 12% instead. In currency terms, they've devalued but chosen to maintain the peg (they're also doing some IMF-style housecleaning where they promise to hold more USD reserves, buy back some STRC, and sell some of their bitcoin to fund this). When they traded at a premium to their net asset vlaue, the right move was to keep issuing stock and buying bitcoin. Now, they have a comically messy balance sheet for such a simple company, and but also enough liquidity to reengineer their liabilities so the most sentiment-damaging ones are under control.

Disclosure: short MSTR.