Longreads + Open Thread

Longreads

• In Arena, Alex Kesin tells the story of pharmacometrics, which turns out to be one of those wonderful collision-of-experts stories, featuring a doctor who knew a lot about pharmaceuticals and a quant who knew a lot about mining noisy, limited data in order to get some kind of signal. Finance tends to produce useful insights about applied statistics because the feedback loop is so short after applying them, and because the gap between backtests and live trading is so informative—you're constantly learning not just whether you were right or wrong, but how calibrated you were. It's not necessarily cheap to test these hypotheses, but validating one tends to fund a lot more experimentation.
• The Diff has noted a few times, starting here, that one feature of healthcare is that it's a complement to everything: if you can afford all the cars, housing, entertainment, and travel you have time to consume, your marginal dollar has its highest return when it literally buys you more high quality time in the form of a longer healthspan. As it turns out, there's prior art, coauthored by recent Sveriges Riksbank Prize winner Robert E. Hall. Interestingly enough, healthcare spending hasn't trended up the way this implies; in the decade leading up to 2009, healthcare's share of GDP rose four points. Since then, it's up 0.8 more. But one feature of the model is that it assumes people place a positive value on being alive (and consuming). But if someone's baseline happiness is low enough that no amount of hedonism gets them to even, and they expect disutility from ending things then and there, they may not pursue life extension all that aggressively. Then again (again!), that often makes them pretty expensive to the healthcare system, since it can lead to chronic and expensive health conditions, whereas someone who's in excellent shape right up to the minute a stroke gets them hasn't been as costly. It's useful to know that, on average, healthcare becomes more of a universal complement as people get richer, but that average includes many different utility functions, which are hard to measure.
• Tyler Cowen interviews Liaquat Ahamed on 1873 and some of his earlier books. One point Cowen presses him on is that, while it's true that the late 19th century was a deflationary period, economic growth held up reasonably well, and the world certainly seemed subjectively richer after all of that (though, as Ahamed points out, it made politics more toxic). The usual reason fiat systems work is that economies can't adjust themselves to changes in credit; wages and prices resist dropping, so if the economy slows down, the only thing to do is to increase spending through fiscal or monetary policy. But if the adjustments happen faster—if workers get pay cuts rather than layoffs, and debts are promptly restructured—then you don't need all of that. Another way of saying this might be that fiat currencies deal with the high transaction cost of revising every single long-term deal in the economy, but this means that its utility is partly a function of how high transaction costs are.
• Jason Schreier in Bloomberg on the making of Grand Theft Auto VI. It's striking just how big and expensive these productions are: it's the highest fixed-cost form of non-live entertainment, unless you expand the scope enough to include the James Webb Space Telescope. The coordination costs are staggering, because so many things are getting built in parallel: the underlying game engine, textures for graphics, audio, a storyline, etc. Open world games have the marginal advantage that in some cases, if they overshoot on one of these they can add more of another—while they're waiting to finalize graphics and work out gameplay bugs, they can always add extra side missions and NPC interactions. But up to four thousand people working for eight years is just an incredible amount of human labor to put into a single coherent piece of media. It will be interesting to see how many years it takes for a swarm of agents to decompile.
• Judge Glock in City Journal has a sure-to-annoy-everyone piece on how DOGE was never going to put a big dent in total spending without doing something to entitlements, but did cut a lot of headcount. And, yes, that includes a lot of layoffs at USAID, but much smaller cuts to the amount of health aid provided by the US ($9bn in FY25, vs $13bn in FY24, but also about equal to the $9bn in FY19). And—narrative violation—the single department with the biggest headcount cuts was the Department Then Known as of Defense. The piece notes that, both in scope and in mechanism, it looks less like what Elon did at Twitter and more like what Clinton did in the early 90s—offer lots of people early retirement and see just how much employee attrition it takes to make a measurable difference in outcomes.
• This week in Capital Gains: how to think about asset managers who have different vehicles with different fees. Obviously, the better moneymaking ideas will gravitate towards the higher-fee structure, and the best ones might be done through a family office or something. But the market for asset management skill is increasingly efficient, in the sense that passive allocators don't expect to be able to pick from a menu of a few dozen funds that will all earn their 25%-with-no-down-years. What they're underwriting is an organization's ability to manufacture persistent alpha, which is the business that allocators have been in all along.
• A Read.Haus reader wants to know how much effort goes into keeping prices efficient. Which is a fun way to put it, because the biggest inefficiencies get exploited by people who weren't even looking for them. The link graph was ridiculously undervalued in 1998, the online world was crying out for a social network that represented real-world behavior in the early 2000s, prospective homeowners circa 2022 were getting astonishing square footage per dollar of monthly payment, etc.; none of them were looking for gaping inefficiencies, just responding to the ones they'd stumbled on. Markets, of course, have more explicit efficiencies. The tendency in market-facing roles is to work people to about their limit, either in terms of total hours of nominal work or total hours of complete focus (spending 9:30 to 4:00 watching and reacting to prices in real-time is a lot more exhausting than having a longer workday that mostly consists of reading 10-Ks and sell-side research). But part of the goal of high-effort research is to create some more automatable process, whether that's a checklist for deciding whether Micron's guide will be good or bad this quarter or a systematic replication of some discretionary strategy. That reduces the incremental effort, but it's really crystalizing the output of past effort.

Books

The AGI Chronicles: If you've been following a story closely enough that you could write a book on it, as Kevin Roose has with AI, one question you have is: when do you send your manuscript to the publisher? If the story has a definite start and end date—a book about a political campaign, say—then reality storyboards this out for you. You start your draft right away, keep adding to it and editing it, and as soon as it's done, you send. But what if the story isn't just changing all the time, but accelerating? The more events that can happen over a six-month period in some area, the more it's impossible for your book to be current events; by the time the reader has it, it's history.

There's a double-gonzo experience reading a book about events I was reasonably close to, in which many characters are friends-of-a-friend, and also reading a book by someone who was the star of one of the funnier anecdotes in the history of AI alignment and deployment. Kevin Roose was invited to try Bing's AI features, and, after a lengthy conversation, his chatbot started badgering him to leave his wife for her. So, he's close to the story.

Which is good! Because it's a story that's evolved a lot since it started with various plucky bands of researchers pursuing interesting ideas that were a long way from any commercial application. It's been striking to see how much of the research-lab culture has persisted as the labs have transformed into trillion-dollar companies: they still publish cool papers, and their employees and executives still make gnomic jokes on Twitter in a way that would seem wildly inappropriate if it came from the CEO of, say, Boeing or Exxon. The labs have certainly changed over time, just not as much as similar important organizations.

Some of the early history details are quite fun. GPT-1, for example, was trained in December, because enough people were on vacation that there were free GPUs. Thus, it joins a long list of inventions made under exactly those circumstances: the integrated circuit and Unix were both invented by people who kept working while people who might interrupt them were on vacation, and a lot of the early work on Microsoft, Facebook, and Stripe took place during Harvard's winter break and reading period. With institutions' calendars and models alike, it's been found that injecting a little randomness from time to time will shake things up and lead to a better outcome.

If you plot model capabilities and labs' valuations, you see a pretty smooth exponent up through the last few weeks, but if you zoom in, you'll see plenty of terrifying dips—disappointing training runs, models getting released with a lower version number than planned (GPT-4.5 was supposed to be GPT-5, but wasn't what people wanted from GPT-5). One thing I'd forgotten, because Anthropic hadn't publicly released a product at that point, was that the collapse of FTX and the release of ChatGPT happened in the same month, and that this would have been an incredibly grim one for Anthropic. They weren't sure if the investment would be clawed back, and OpenAI had shipped something that they could have, so it would be easy for them to envision a world in which AI capabilities raced ahead of alignment purely because of their bad luck. Even if shares of pre-IPO labs aren't regularly marked-to-market the way a public company's are, owning them means feeling the volatility even if there isn’t a daily mark.

Part of what this book does very well is that it puts all of these events together in a coherent timeline, whereas we've experienced most of them as one exciting release or headline at a time. An important feature of AI people, and something useful to keep in mind when talking about them, is that they live in a world where they used to be a tiny minority with views too weird to be worth addressing ("So you'll be able to tell your computer 'make me a picture of a dragon playing a guitar solo' and it'll just make one? And also this computer might decide to turn us into paperclips? C'mon..."), and those views have been vindicated. So, in their experience, people who call them crazy are just flat-out wrong and will be proven so in a few years. They have earned the right to be very, very weird, but the more legitimate that is—i.e. the more powerful models get, and the more profitable labs end up being—the more we'll have to ask the question of just how powerful odd people should be, and vice-versa. If caring about shrimp welfare is closely bundled with the ability to come up with creative ways to improve AI performance, and if AI keeps getting used more as a result, we approach an asymptote where the world is run by people who really care about shrimp. Which, in one sense, is fine—we ran a meritocratic process where there were many avenues for seeking power, and the best power-seekers happened to choose a character build more focused on linear algebra than on winning elections or whatever, so that's who we wound up with. But it will make politics very fraught, inside the labs and out: on the outside, voters are clearly spooked by AI in various ways. But inside the labs, there's reason to be equally concerned by the weird beliefs of the median voter, who cares a lot about AI, can influence the course of its growth through voting, and has basically no idea how it works. This book will give them some valuable context.

Open Thread

• Drop in any links or comments of interest to Diff readers.
• What is most conspicuously missing from the mainstream narrative about AI history? Or do we have the story basically right?

Diff Jobs

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

• ​​Ex-Anduril, Ex-Abnormal Security, Ex-Bridgewater, fast growing startup building frontier agentic cybersecurity to make the long tail of businesses defense dominant via MSPs is looking for a senior machine learning / quantitative researcher. Working in low data environments or working in low signal-to-noise environments e.g, mid-frequency quant finance is a big plus. Evidence of data rigor required. (SF)
• 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. (NYC, Remote) 
• 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)
• Series A startup building multi-agent simulations to predict the behavior of hard to sample human populations is looking for researchers and engineers (ML, platform, infrastructure, etc.) to improve simulation fidelity and scale the platform to hundreds of millions of simulation requests. Problem-solving and genuine interest in simulation matter more than pedigree. Experience working with languages with an algebraic type system is a plus. (NYC)
• A Fortune 500 cybersecurity company with decades of proprietary security data is running an internal incubation with a pre-seed startup mentality and a mandate to build something new in AI. They are looking for a founding engineer who can ship fast, an engineer with a security background who’d be excited to contribute to OpenClaw’s security efforts, an AI researcher, and a generalist (ex-banking/consulting/PE background preferred) who wants to wear a bunch of different hats. Comp is FAANG+ and cash heavy. If you want to build something new in AI, but also need runway, this is for you. (SF/Peninsula) 

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.