Longreads + Open Thread

Longreads

• A while ago, The Diff wrote about the Strategic Petroleum Reserve as a way for the US to write synthetic options on oil, hedged by its ability to allow more oil production and print more dollars ($). Turns out, it's also a physical place, and Brian Potter has written about how it works. It's kind of a nice callback that oil is an incredibly convenient energy source we extract from the ground, and it's so useful that it would be a catastrophe if we lost access, but we use it in such volumes that it's expensive to store, but thanks to another geological feature, we can actually carve out big storage containers from salt domes. (The way to hollow them out is, basically, pump in freshwater, wait, pump out saltwater, repeat.) This turns out to be a great hack for storing oil for one big emergency, but a bit worse as a way to constantly tweak the supply. Interestingly enough, there's a kind of options contract called a swing option that maps to these physical limitations, so you can still think of the SPR as an option-writing strategy, just one that's a little more complicated to price.
• Dan Luu argues that slow software is now a choice: you can keep throwing inference at the problem of making something faster, including a mix of small improvements that a human programmer wouldn't think were worth the time and big but uncertain ones that a human might think weren't worth the risk that they'd take too long. The most interesting bit is the tradeoff between more compute and domain expertise. Luu built a game AI using LLMs, and found that his was better than the state of the art. But that was mostly because it could brute-force more effectively; the best competitors were cleverer, but too slow. He claims that every doubling of speed adds 100 to the Elo, which is a convenient rule of thumb to keep in mind if it generalizes. There seems to be a general correlation between corporate success and caring a little more about latency than seems reasonable, at scales that range from how fast clicking a button does something to how long it takes to answer an email to how long it takes to negotiate an acquisition. Speed matters!
• When people talk about the environmental impact of AI, they tend to use the broadest measure they can. Which is fair! If a datacenter is the reason a coal power plant didn't get shut down, it's a contribution to emissions. But doing this right means looking at cases where the emissions impact is negative, like rerouting drivers so their trip time is imperceptibly longer in order to make the average driver's trip faster, alongside a slight decrease in emissions. As with the piece above, there are many domains that weren't worth optimizing because collecting the data would be such a schlep. Now that's less true.
• Jack Despain Zhou on how learning with AI can make you feel like you're learning, at the expense of actually learning. In one sense, this is entirely possible. On the other hand, one of the studies cited compares students who did and didn't use AI; the AI users finished their homework much faster and got better scores, but did worse on their exams. However! The magnitude of the homework time saving was bigger than the magnitude of the average test score gap; eyeballing the charts, the modal exam scores aren't that far apart, but there's a long tail of AI users who, to varying degrees, absolutely bombed the exam. Meanwhile, the right tails fo the distribution aren't as skewed. So, using AI to learn is kind of a Ring of Power: do it responsibly, and you will speed through the boring stuff and be able get hard concepts paraphrased over and over again until you've brute-forced your way into understanding them. There was certainly a point in the history of math education where calculators would have had this effect, but now we're in the era where teaching students how and when to use calculators is more effective than skipping this.
• Brad Burnham, Fred Wilson, Dayna Gant, and Lindel Eakman on how the first Union Square Ventures fund was raised. Interesting throughout, half because of what's changed, half because of what hasn't. One fun detail is that raising a new venture fund in the early 2000s was partly a form of regulatory arbitrage: public pension funds had recently been required to disclose more data about returns, and so the biggest venture funds stopped raising from them in order to keep their performance confidential. Which meant that those LPs were under-allocated to venture, and a new fund that put a lower value on secrecy could raise from them. You could have looked at a new fund in this period as a quick money-grab, more focused on an opportunistic 2 than on an uncertain 20. But success tends to be correlated: someone who finds a good tactical way to raise money from investors they couldn't otherwise reach is probably better than average at strategically deploying what they raise. (Via Trevor Mckendrick.)
• On Read.Haus, someone asks what to do if you're a founder who controls a company that's no longer growing, if you're worried that you may not be able to harvest the cash flows. In practice, there seem to be a few different answers. One is that some companies try to maximize profits over some period, or hit some growth rate, but the most interesting thing to optimize for is maximizing your growth rate a few years out, i.e. making decisions today that give you the highest odds of being able to grow at a similar rate five years from now. That tends to involve solving some scaling problems that aren't problems yet, and sometimes reveals new kinds of growth. Another pretty conventional answer is that founders should diversify gradually: having most of your money in a mature business is a smaller problem than having most of your money in a high-growth and high-uncertainty one. The last answer is that PE firms are partly in the business of making this somebody else's problem: there is a level of leverage that allows a mature company's returns to equity holders to be closer to what people demand from venture, but running things that way is a different specialization from running a growth company or taking care of a stable one. Life is full of tradeoffs, and being undiversified has a cost in the form of volatility drag. So perhaps the best way to think about this is that maintaining control of a company and refusing to diversify is the world's most expensive luxury good.
• In Capital Gains, a discussion of how a tech-driven equity bubble can turn into one based on financial engineering. This is, of course, all about how in the 1920s, corporate treasurers found that they could lend their cash to banks that made margin loans and earn more money than their earnings yield, creating a loop where investors paid high interest rates to invest in companies whose earnings were, increasingly, from lending to those same investors.

Books

Rickover: The Struggle for Excellence: There are many modern jobs that make me wonder what the people who are good at them used to do before that job existed. Someone who's technically skilled and incredibly driven today is going to get a lot of social cues about how to turn those skills into an interesting and lucrative career. But a few generations ago, there was less of a script, so you have to look elsewhere for that archetype. Specifically, you can look at someone like Hyman Rickover. Rickover was born in Poland in 1900, grew up in Chicago, joined the Navy, and led the development of the nuclear submarine fleet and later the Shippingport reactor, which used a pressurize-water design that later became standard for most reactors in the US.

Rickover is partly a story about extreme social mobility, with a fun early ingredient: one of his first jobs, during high school, was delivering telegraphs. It's an unskilled job—if you have a bicycle and can read a street address, you can do it. But it's also a job with extreme variation in individual performance, and one that introduces some scrappy young people to a completely different social class that they might set out to join. (Andrew Carnegie's career also accelerated when he started delivering messages like this; both he and Rickover found ways to get the highest-value messages, which meant both better tips and better connections.)

Rickover, with a bit of political maneuvering, got admitted to the Naval Academy, and spent time in positions of increasing responsibility aboard battleships and submarines. He seems to have developed an early instinct for constant optimization. The fleet had an internal competition about engineering excellence, quantified on the basis of how little fuel and fresh water they needed to accomplish a mission. So Rickover would slow-walk repairs to heating equipment, mandated short showers, replaced lightbulbs with lower-watt ones, etc. He was very good at trading off anything and everything against numerical targets.

The Rickover story is about politics and power, in a weird way. Rickover was extremely abrasive, and seemed to make lifelong enemies wherever he went. The general rule was that people loved his results but hated him, so he tended to get support that skipped a step or two in the org chart. As he rose up the ranks, rising got harder because there was less room at the top, and at one point he was passed over for a promotion and could have been forced to retire, midway through the first nuclear submarine project. His supporters were able to lobby for him, and made it a media project—it's hard to imagine it being a big news story today that the Navy failing to promote someone might delay a project, but it was big news, and led to lots of letters-to-Congressmen and other pressure until he got the promotion. Rickover was no one's ideal of a natural politician, but he had enough political sense to leave that job to other people.

One other thing that stands out about the book: he spends a surprising amount of time in the hospital for assorted illnesses. This is the kind of thing that I've paid more and more attention to in biographies as I've gotten older. Success usually involves some kind of compound interest, building up skills, connections, and different kinds of capital. So losing a year or two to illness is a pretty big deal, especially if there's some zero-sum aspect to the competition. The fact that Rickover didn't get derailed by his various medical problems means that, graded on a curve, he's even more impressive.

Open Thread

• Drop in any links or comments of interest to Diff readers.
• Rickover was basically a technical founder by personality type and skills, who deployed those traits in a surprising domain. It would be interesting to hear about other people who fit this pattern, like being a very good diplomat who happened to work as a lawyer.

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 an intellectually curious, mathy generalist to work on projects spanning business strategy, technology, and markets. If you’ve got the quantitative chops of a prop shop trader or ML researcher, appreciate that operational alpha decays more slowly than trading alpha, and want to work on entrepreneurial projects that drive the former, this one is for you. (NYC)
• A premier prop trading firm is looking for someone to turn its prime brokerage partnerships into something of a solved, alpha-generating optimization problem. If you’d enjoy thinking through how to price the optionality of having an extra $1bn available when a liquidity-constrained opportunity shows up and continuously mapping the term structure of financing costs across PBs to ensure that $1B shows up at a low cost of capital, this one is for you. No formal background in finance is necessary, if you have a track record of bringing a systems thinking / analytical mindset to situations where success is traditionally categorized by relationship management, please reach out. (NYC)
• Startup building the one API to power every strategy and capital allocation decision in the economy is looking for a co-founder / CTO that’s excited to solve the series of technical grand challenges (and inevitable schleps) involved in making this a reality. If you’re looking to build the economic sensor infrastructure underpinning the next leg of productivity growth, please reach out. (London, NYC)
• High-growth startup building dev tools for wrangling and debugging complex codebases is looking for someone who can personally execute the SaaS bear case: review the third-party software they use and figure out what to keep, what to drop, and what to implement in-house. (SF, DC)
• 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.