Longreads + Open Thread

Longreads

• From SentinelOne, some amazing malware archaeology: there was a clever, subtle, complicated piece of malware, at least five years before the infamous Stuxnet. What's most entertaining about this one is how it works: it very slightly corrupts the results from particular pieces of simulation software. These kinds of statistical attacks, where you don't shut down a system entirely but just shift it to being slightly less accurate, are incredibly fun to read about, because they're subtle enough that sabotage isn't going to be the first theory someone comes up with, and because it's so hard to track down the mechanism. We're in the middle of a kind of sonic boom of software history, where it's more tractable to look at cases like this and figure out what's really going on. (It's also clear from the snippets of code referencing this project that this kind of attack is very fun to pull off.)
• Jane Psmith on the life of 12th century knight William Marshal, which also turns out to be a good window into the medieval world and the broader world of shaky property rights and weak governments. A surprisingly important factor in economic history is: how big a tax base do you need to support one effective soldier? In a relatively peaceful context, this is just an academic exercise of calculating various ratios, but any time military conflict or the threat thereof is an important factor in decisions, it has a big impact: England after the Norman conquest had fairly concentrated land ownership because it took a lot of land to afford a warhorse, and these horses had a fast, lumpy depreciation schedule. During the US's westward expansion, a more gradual sort of conquest, land got dispersed, because the cost of a musket and gunpowder, as measured in hours of labor-output-minus-subsistence, was much lower. It's also a look at how social roles develop over time; The Diff has pointed out in the past that once some job gets canonized in the media (Liar's Poker for bond traders, The Paper Chase for law students), that attracts new people, but what they experience is other people living out the same cultural narrative.
• Tyler Cowen interviews Craig Newmark. This is an interview with someone who doesn't spend a ton of time on the podcast circuit, which is refreshing. Craigslist has always been an odd sort of institution, in the sense that anyone could have put together a stripped-down classified ads product (so he's nothing special) but he did it first (so he's the all-time winner in an incredibly competitive field—$6M per employee is not too bad). The best way to read this interview is that it's a meditation on what to do when you end up with a surprising amount of wealth and responsibility—something many people in the AI world, and some people in the extended AI supply chain, should be thinking about.
• Ben Thompson interviews Sam Altman and AWS CEO Matt Garman (disclosure: long AMZN). What's striking about this is how big a deal it is that OpenAI models—accessible to anyone who gives them payment information and reads the API docs—is so much more valuable inside of the existing AWS ecosystem. There's also a fun early meditation on how Y Combinator was, in retrospect, a kind of macro bet that tools like AWS would decrease the cost of starting a startup, and improve their cycle time. Software investors had a very lucky series of tailwinds from cloud computing, the proliferation of point solutions removing the need to internally implement the same thing, and the cultural cluster that made sure founders knew about these things.
• How do prediction markets actually produce useful information? Cram, Guo, Jensen, and Kung look at data on Polymarket traders and show that about 3% of them have skill in making predictions, 0.1% of them are market-makers, and everybody else is some variety of bad at making money. Part of what they're exploring is whether the "wisdom of the crowds" means that we're getting a good view of the average opinion, which is more accurate than any one person's opinion, or if informed traders are a minority who have a disproportionate impact on prices because they consistently trade in the most accurate direction. It's the latter. And that makes sense given the distribution of knowledge and handicapping ability generally: selective recall means that we'll all tend to remember the improbable outcomes we were right about and to forget our other predictions, so it's reasonable to expect most people to be over confident in their predictive ability. But that means that we're the kind of noisy traders it's profitable to provide liquidity to, and that means there's enough liquidity for good traders to make money.
• In this week's Capital Gains, we belabor the point that the date range you choose has a big impact on the trend you see. It's important, because choosing the start and end dates for some measurement is a good way to optimize for attention-getting statistics, particularly if they're wrapped around a compelling narrative.
• A Read.Haus user asks: what are the flaws in synthetic market research? Synthetic market research is, basically, using AI to simulate people and then asking them questions. It's much cheaper than doing a regular survey, but has the obvious laundry list of problems you could name for an approach like this: it's sensitive to what it's trained on, what RLHF has happened since then, how questions are phrased, etc. If you're looking for a particular answer, this approach will let you find it. The trade off is that, if you're willing to accept some level of noise (or willing to anticipate that the level and cost of the noise will both monotonically decrease for the foreseeable future), you can test a lot more, and more cheaply. The risk is, ironically, similar to the risk with polling in general: you're limited by who gives you answers and by how honest they're willing to be.

Books

Investment Banking in America: A History: let's begin with a preamble and articulate the Diff theory of used bookstores. Algorithmic feeds and search results are incredibly good at finding things you'd want to interact with and pay for. They're subject to a ruthlessly Darwinian process that makes it so. But this also means that the more of your life that's managed by algorithms, the more you lock into path-dependence: it doesn't take long for big platforms to put you in some cluster of similar people, at which point some share fraction of your life is going to be an idealized version of the average of lots of people like you. But this is more optimized than optimal. You still want to inject some entropy into the system.

And a good way to do that is to go to a used bookstore. You can browse by category, but what you get is not what's been selected by some monstrously complex algorithm as the highest-probability bet. There's some randomization, which you need. But there's also some serious adverse selection. There are three reasons people sell books: they're de-cluttering, they're moving and are thinking of books in terms of pounds and cubic feet, or they died and their befuddled heirs are liquidating their possessions. Which means that you, as a used-bookstore shopper, are operating in an adversarial environment. If you're at Strand, where I picked up this book, there's a good chance you're buying a book from the collection of someone who couldn't hack it in NYC—a very bad omen if you're browsing the "Finance & Economics" section! So, if you see a book whose publication date is in the last five years or so, there's a very good chance that this is a book for posers. Those odds get even more punitive when you think about buyers, who are going to snap up bargains and leave the rest behind; a used bookstore is a store full of books that lots of people have glanced at and declined to buy. So, you want to optimize for books that made someone text their sibling and say "Can you believe dad had a book about..?" And a history of investment banking, published in 1970, probably qualifies. It feels like I'm shoving some heirs aside in order to commune with a book collector who would have been fun to hang out with.

So, the actual book. From the perspective of the book's time period, the history of investment banking in the US is that for a long time, they were a smaller piece of the economy than was typical for other countries with similar economies to the US (before the First World War, it was common for companies to underwrite their own offerings, because they were more trustworthy than the typical bank). They grew, reached a high point in the 1920s, and then declined thereafter, with small ebbs and flows. That thesis would have been creaky in 1980, and definitively disproven by 1990. But it's still a good summary of where the world stood then.

But with the benefit of more than a century of hindsight, we can see that the industry follows a cycle but tends to grow over time, and that given enough time between cycles, the exact same idea can repeat itself. For example, the book details a situation where financial institutions like life insurers were facing concerns about their creditworthiness after they'd extended illiquid loans to dubious borrowers. The ensuing crisis in confidence led to the selloff of 1907, but mirrors the situation in private credit today (albeit without the bank run-prone institutions). Rolling forward a few decades, the 1920s saw a spate of leveraged investments in electrical utilities, rather than in any particular company making electrical appliances or machinery, which maps to the neocloud boom today.

One notable point the book makes is that the US's equity culture was actually indirectly kickstarted by a government program: when the US sold war bonds, it was the first time many consumers had opened an account with any sort of brokerage, and that reduced the friction of participating in the postwar equity market excitement. Businesses that are dependent on network effects, like financial markets and the Internet, tend to get kickstarted by large, horizontal institutions that can capture the upside. And the government tends to be the biggest of these.

Open Thread

• Drop in any links or comments of interest to Diff readers.
• Are there any other good industry histories that stop right around when a trend inflected, one way or another?

Diff Jobs

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

• Series A, ex-Navy defense firm building AI-enabled drone defense systems is looking for an electrical engineer with range: schematic design, electrical simulation, printed circuit board (PCB) design, and firmware development in high performance languages (C++, Rust, etc.) If you’re interested in power and control systems for mission critical technology, this is for you. (Austin, TX)
• A startup building a new financial market within a multi-trillion dollar asset class is looking for a data scientist with commercial financial experience. (if you’ve been an investor but are newer to the data side, that’s great too.) (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)
• AI Transformation firm with an ambition to build an economic world model to run swathes of the private, unstructured economy is looking for FDEs, Platform Engineers, and business generalists who understand how to solve problems.  
• Well-funded, frontier AI neolab working on video pretraining and computer action models as the path to general intelligence is looking for researchers who are excited about creating machines that learn from experience, not text. Ideally you have zero-to-one pre-training experience and/or are a high-slope generalist who’s frustrated that the big labs aren't doing this. (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.