Longreads + Open Thread

Longreads

• The NYT has a look at essay mills in Kenya, which lost most of their business to AI. The piece also talks about legitimate workers who got hit with the same problem: transcription jobs were actually automated a little earlier. One way to look at this is that there's a part of the labor market that competes purely on price—in this case, basically by being the poorest place rich enough to afford fast Internet—and this category is the most vulnerable to technological disruption. In fact, this story gets told in many developing markets and with many different causes: when a poor country first starts exporting, it'll often be concentrated in either whatever the most convenient raw material is or whatever form of light industry takes off first. They eventually diversify, but early on, the process of globalization is one of linking small economies to an enormous global economy whose shifts can be violently unpredictable.
• Dan Williams argues that wishful thinking is mostly a myth: there's very strong evolutionary selection against a tendency to believe that everything is fine in a practical sense (though there's selection the other way, towards believing that everything is fine in a more existential sense). And, as he notes, people tend not to have imaginative beliefs that make them feel good: nobody talks about how the strings are being pulled by a cabal of elites who are all doing a really great job and deserve all the adrenochrome they want. No! If you have some unusual political belief that other people don't see evidence for, it's often about something very bad. Meanwhile, some of the behaviors that people explain through wishful thinking can be explained through better-supported models.
• A semi-popular claim online is that AI detectors are biased, particularly against people on the autism spectrum or who aren't writing in their native language. (The second is technically true: if I ever send an email in French or German, it's going to be something I wrote in English and then used an LLM to translate.) If you make a falsifiable claim about the behavior of people on the autism spectrum, you should be thoroughly unsurprised to find out that someone used $80k worth of tokens to write a 12k-word rebuttal: no, there is not any detectable bias here. The data-gathering piece is the most worthwhile, because there are many other kinds of questions LLMs can answer at scale about the statistical properties of various texts, and they're all constrained by the need to have well-labeled examples. In this case, the study looks at self-disclosure, participation in autism-focused online communities, disclosure from reliable sources, etc. Which is about the best that one can do, but does run into a problem: the best way to think about autism diagnoses, and diagnoses of other forms of neuroatypicality, is that when you're dealing with institutions, you want a binary, and when you're thinking about people, you want a spectrum. It's useful to know that people who are bad at eye contact and develop intense intellectual interests are also going to be sensitive to sound and texture, but pretty useless to argue about whether or not someone is just below or just beyond a diagnostic threshold. But this has a big distorting effect on which writing gets explicitly tagged as coming from an autistic source, and which comes from someone opting into the identity, and even though they overlap a lot, these are distinct populations. Fortunately, the study's null result means that this is unlikely to affect the outcome: as it turns out, humans, even ones with unusual brains, don't replicate the statistical distribution of LLM-generated text.
• Ben Moll and Alex Imas ask: what do you have to believe to expect 10%+ GDP growth from AI? On one hand, the classic dynamic is that agriculture and manufacturing keep producing higher output per worker, to the point that they saturate all plausible demand and further productivity growth just shifts workers into other sectors. Which is a good model, though in AI's case those sectors need to be defined in a more fine-grained way, because many of them are exactly where AI is substituting capital for labor. You don't strictly need a lawyer to review or draft a contract (though it would be a good idea to have one take a look before you sign). But if you break service sector jobs down more, there are some that get automated fast and some that are complements to that automation, and over time those jobs account for more wages because they're the jobs that are left. The Baumol Effect is basically how economies keep productivity gains in one sector from causing that sector to eat the entire economy. It's a good piece that puts sound bounds on a few of the crazier scenarios—and a point or two of extra GDP growth is still a huge deal, even if it's not what some people were hoping for.
• After 9/11, the NYT collected narratives about what people did in the towers after they were hit. One of the reasons 9/11 had such an impact on the public consciousness was how thoroughly documented it was; we have footage from the moment it started through the collapse of the towers, and extensive records of how people got out or what they were thinking before they died. And yet, compared to the modern world, there's incredible information poverty: worse cameras, less connectivity, no expectation that everyone's available by text, etc. We still don't know the identity of whoever it was that climbed down more than a dozen stories on the outside of the North Tower; if it had happened today, his name would have been trending on Twitter by the time he died. Media has a multiplicative impact on terrorism—there's a reason Al Qaeda targeted the buildings that directors use in establishing shots when they want to say "this part takes place at the center of American commercial or military power." One potential upshot of this: a more fragmented media environment but a much more media-aware population will lead to more small-scale, meme terrorism. The assassination of Charlie Kirk is a good example of this, both because of the act itself (the cartridge case of the bullet that killed him had a furry meme on it) and because of the aftermath where it's a popular topic for nihilistic humor. 9/11 itself might be considered a late entry in the story of mass-media terrorism; over time, even figuring out why someone committed some atrocity is going to require hours of immersion in the relevant lore.
• A ReadHaus user asks about the intuition behind more borrowing (e.g. from AI companies) and higher interest rates. There is the mechanical supply/demand explanation, which is always useful to keep in mind, but the deeper phenomenon at work is that real interest rates represent some broad exchange rate between consumption in the present and consumption in the future, and, in particular, represent how much you get if you defer consumption right now. The more promising investment opportunities are, the more the incentive to do exactly that. Zero interest rates imply deep pessimism, and are basically making the statement that we might cycle through some different kinds of capital, but basically have all we need. Higher real rates mean that the average view of the future is more abundant, and that it's worth sacrificing more to get to that future sooner.
• In Capital Gains this week, we're on the media theory beat with a brief history of quoting people in text, and thoughts on what level of accuracy is honest. This is motivated by two phenomena: accurately reporting someone's filler words, pauses, etc. to make them look stupid or dishonest (take your pick!), and the phenomenon of using ellipses to remove text that actually changes the meaning of the quotation. Writers do actually have permission to make spoken words read more like written words, but that flexibility gives them some room to mislead readers.

Open Thread

• Drop in any links or comments of interest to Diff readers.
• We've had a lot of AI-related news in the last few weeks. Curious for readers' thoughts on what will be most important in retrospect: general capabilities, specific mathematical results, misbehaving models, or the fact that that misbehavior has been identified?

Diff Jobs

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

• The next great American print magazine hyperstitioning abundance through gorgeous techno-futurist aesthetics and the highest quality long-form writing on entrepreneurship, technology, and capital is looking for a Chief of Staff to execute on a number of interesting growth projects. One day you’re standing up a physical bookstore (and modeling the attendant unit economics), the next you’re producing a podcast or mailing coffee table books to cultural tastemakers, throughout it all you’re focused on getting things done. If you have an operational / analytical toolkit (e.g, banking, consulting, startups) and want to work with a team shaping culture, please reach out. (Austin)
• 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. (Remote)
• 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)

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.