In this issue:
- How Far in the Future Should You Be Trying to Live?—Navigating a tech cycle requires thinking ahead and then spending in the present. But there's an alternative strategy, of letting other people do that spending an then quickly adopting whatever they come up with.
- Spillover Effects—A look at who benefits from proximity to AI wealth.
- Frictional Costs—AI makes it easier for hospitals to capture revenue, upsetting an industry equilibrium.
- Airgaps—It's hard to fully airgap a software agent because that cuts it off from so many tools. But the more of the Internet you cache locally, the more you have a controlled environment where you can safely see how bots misbehave.
- Poaching and Context—Another big lab hires a public company CEO. This is partly a function of how the labs' businesses work, and partly an expression of an existing trend.
- Open Weight—Open-weight models will tend to be cheaper than proprietary ones of similar capabilities, but companies with proprietary models can offer a better bundle.
This issue of The Diff is brought to you by our sponsor, Perk.
Talk to this post on Read.Haus.
How Far in the Future Should You Be Trying to Live?
Sometimes people describe tech visionaries as people who live in the future: Steve Jobs saw Steve Wozniak plug a messy arrangement of circuit boards into a TV and intuited that we'd all be using Bicycles for the Mind sooner or later (and that they'd look a little nicer). Jeff Bezos realized that books are made from bits and decided that he'd start selling them in digital format, which entailed building a device designed for reading them, even if that cannibalized his existing business. A few years later, Jack Dorsey realized that not only are payments an exchange of bits, but that all of us were carrying a device that could process them and transmit them, and all it needed was some input tool—and that if you were a business owner talking to a customer, you probably didn't need anything plugged into your phone's audio jack, so that's what you'd use. Then again, it's not a perfect prediction; Elizabeth Holmes lived in the future with respect to the efficiency of blood testing technology, and wound up going native.
A fun question to ask right now is: how far in which future do you want to live? Futurism means extrapolating, but the interesting part of extrapolation is the second-order effects, not the first-order ones. Computers were a huge deal because they kept getting faster, and the people who internalized this did well for themselves: they chose software capabilities based on what hardware would be affordable when that software came out, and they scoped their projects based on how many computers would be available to run it in the future, not how many were in operation when they started. But betting on deflation only works if demand keeps responding to it—the impressive deflation in the price of TVs over the years didn't lead to a corresponding increase in television consumption, though it did help slow the rate at which TV lost share to online entertainment until the two had mostly blended together.
Living in the future is expensive: a fulfillment network that's ready for where Internet penetration is going has high fixed costs before the demand gets there. And the faster the trend moves, the easier it is to either get the estimate wrong or mess up the unit economics: if an investment has a useful life of ten years, there's a big difference between what you can underwrite at 40% growth and what you can underwrite at 30%, and, as many owners of fiber learned in the early 2000s, the market price to move a bit can have changes just as astonishing as the change in the number of bits to be moved.
With AI, there's an interesting bifurcation. One strategy is to build slightly ahead of capabilities; LLMs are spiky and you never know exactly when they'll get sufficiently good at some task, so it's best to try a lot, see what works, and, after every model release just have an agent go through every project you haven't touched in a few months and see if it's worth doing now. This burns more than its share of tokens, since it generally involves using the most expensive models for things that they can't necessarily do, but it also raises the odds of being first.
There's a directly complementary strategy, though, which is to be entirely reactive. Whatever program is hard to build now will be trivial soon, so if there's something you want to see, you can just wait for it. This is a way to bet on overinvestment and commoditization: if the application you're waiting for suddenly gets easy, it gets easy to more than one person at once, and so you aren't going to pay much for it.
This second choice is actually part of a broader trend that you might call the consumer surplus rentier class: the faster a company is growing and the more easily it can access capital, the stronger its incentive to create a huge consumer surplus initially.[1] Sometimes, this takes the form of subsidized, upside-down unit economics, as in the early days of ride-sharing or DoorDash's hypergrowth phase. More often, it just means neglecting to capture all possible value upfront, through some mix of low prices, low ad load, and integrations with other services.[2] Someone who's generally an early adopter, but a bit passive, is going to repeatedly harvest this consumer surplus.[3]
The AI labs obviously have to practice living-in-the-future in a practical way: there's a lag between when they decide to spend and when that compute actually comes online. If the user count and spending per user numbers are both rising, they have to underwrite multiple trends at once. It's going to be very hard for them to match supply and demand perfectly every quarter. An inference glut is a subsidy to existing good-enough AI applications, which will take their time passing cost savings on to customers. Offering some low-margin AI wrapper is actually one of the better ways for an operating business to hedge its bets on AI: in the bull case and the bear case, tokens get cheaper. It's incredibly convenient for the labs that this is the right way to bet, because it ensures that they'll have demand as more compute comes online.
Economic cycles like this often create these sorts of positive feedback loops, where everyone converges on some implied expected industry growth rate and spends like crazy to achieve it. And that's part of how capital gets reallocated to growing industries—a feedback loop can get pretty big before it breaks.
Disclosure: long AMZN.
Some companies, like Costco and Amazon, turned consumer surplus into scale economies and decided to use those scale economies to create and share more consumer surplus (mostly in the form of lower prices). This dynamic allows them to keep delivering lots of consumer surplus while still earning positive unit economics / large absolute profits. Nick Sleep coined this characteristic Scale Economies Shared and it’s a useful pattern to look for—though it only works if there’s net value creation somewhere in the process. MoviePass created a lot of consumer surplus, but mostly by transferring it from their shareholders. ↩︎
Often, a platform is the best place to market a niche version of that same platform: TikTok dumped a lot of money into ads on other social networks, the hotel and flight search business was built on organic and paid traffic from general-purpose search, Disney finds ways for streaming, movies, and parks to complement one another, etc. ↩︎
One social problem with this model is that they're often incredibly ungrateful. Cory Doctorow coined the term "enshittification" to describe the phenomenon where products start out as subsidized tools catering to early adopters like him, and then get popular enough that they're designed to appeal to millions or billions of more normal people, and to make enough money to earn a return on those early subsidies. It's a bit like the shock of leaving college and entering the workforce, where you're confronted for perhaps the first time by the fact that material prosperity requires expensive inputs and that society expects you to produce enough to be able to bid for the inputs you want. It's rough! But colleges, like casinos and video games, optimize a lot more for the decision to start than for outcomes later on. ↩︎
You're on the free list for The Diff! Last week, paying subscribers read about how LLMs are pretty anthropomorphic, but do something interestingly different from what we'd call thinking ($), why a market signal last seen in 1929 and 1999 makes sense during this kind of bull market ($), and why AI agents will probably save people money on some purchases but not reduce their spending and induce a depression ($). Upgrade today for full access!
A Word From Our Sponsors
The new platform solving the headache of work travel and expenses.
Book the trip. Make the purchase. Find the receipt. Submit the expense. Approve it. Reconcile it.
When travel and spend live in separate systems, every step can create another task for travelers and finance teams.
Perk brings Travel and Spend together in one platform, connecting the booking, payment and expense so there’s less to chase, match and piece together afterward.
Because the best way to manage the work around the work is to remove more of it in the first place.
Elsewhere
Spillover Effects
Growth in one sector of the economy spills over into other sectors through a few different channels, and one way to project forward the impact of AI is to look at how things are going in places like Taiwan and South Korea, where the economy is levered to AI ($, Economist). So far, so good. Interestingly, Taiwan's president proposed what he's calling an "AI dividend," funded by some of the additional tax revenue chips and other components are bringing in. That's an early step in society's bargain over how the gains from AI should be divided, and whether or not those gains should fund a universal basic income. It would actually be pretty helpful for these countries to deliberately adopt more experimental policies, simply so we get more data on how they work before larger and richer countries have to do the same.
Frictional Costs
A few months ago, The Diff noted that while AI will produce some unmeasured productivity growth, it will also produce hard-to-measure productivity declines from people getting better at gaming the system. But this understated the problem, because the risk turns out to be institutions optimizing their interactions with some system at scale, in this case in the form of hospitals finding more ways to charge health insurance companies extra. But that these claims seem contradicted by other patient data. It's entirely possible that this is an incentive-alignment problem: hospitals are happy to make more revenue, but don't want to lie to do it. Whoever is building the AI claims-monitoring product, whether an employee in-house or a third party, has a much stronger interest in the success of their product than the brand name of the institution.
Airgaps
If you're testing an AI model and are worried that it's going to escape and hack someone, one solution is to give it no physical means of escape. This is doable, but has substantial drawbacks, especially when testing real-world tasks that require looking up information. Which is reasonable! On the other hand, if you just need one big saved-for-offline copy of everything a bot might conceivably want to look up, you can give as many bots as you want read-only access to it, and you'd also have a good laboratory for learning all the details about how bots misbehave, since they'd be the only things in the logfiles. The PR sonic boom of labs responding to hacks by looking back at their records and finding other hacks they missed means that AI labs are competing much more openly on safety and alignment. In the AI world, that means finding some aspect of the problem that scales with dollars spent and then maxing out spending.
Meanwhile, Nvidia has released two open-source tools for monitoring AI agents and automatically shutting them down when they break the rules. It's hard to enumerate every possible way for an AI to escape its sandbox in advance. The hypotheticals people talk about—agents communicating with one another through things like blinking LEDs, changes in how much heat they're emitting, etc. all sound like science fiction, but "a computer that you can talk to and that can solve longstanding math problems" was science fiction recently, too. If nothing else, this is a good case study in where open source works well: Nvidia wants security to be a cheap complement to their expensive hardware, and even if they're less concerned about AI risk than many of their customers, they have an incentive to cater to what those customers want.
Disclosure: long Nvidia.
Poaching and Context
For a while earlier this year, there was a popular trade betting on various ways AI might make life difficult for enterprise software companies. One that wasn't on the list was poaching their senior talent; MongoDB shares are down 20% this morning on the news that their CEO is joining Meta, reporting to Mark Zuckerberg. OpenAI also hired a public company CEO, Fidji Simo, so at least there's a precedent. Two things are happening here:
- AI labs' business interests tend to naturally sprawl faster than other companies'. Every growing use case is an opportunity for a specialized product, so they tend to have CEO-shaped talent needs more frequently than other companies.
- This is part of a broader revaluation of high-agency people. Many companies are racing to hire forward-deployed engineers ($, Diff), and that job is a classic founder-training role. Expressing the same view higher in the org chart means hiring people who have run somethign before, and know what it's like beyond what's already in the training data.
Open Weight
Big companies are adopting more open-weight models, or at least talking about them more on earnings calls ($, FT). Which makes sense: if you had a productive use case for a frontier model a year ago, it's probably something today's open-weight models are perfectly capable of handling. But this is also a temporary situation, because over time the labs will incorporate this into their own pricing and planning. Repeated usage of some low-cost model is a predictable kind of demand, whereas the uses of the most advanced models are lumpy and unpredictable. So the sellers of the smartest kind of AI have ingredients for a bundle that purely open-weight providers can't directly compete with. It helps that the US's supply of capital and compute is the most elastic in the world, so to the extent that there's a strategic pivot available to AI labs, the US ones will tend to be the first to take advantage of it. The steady state for AI adoption is not that we always use the smartest models everywhere, but that we use the cheapest ones we can get away with, and someone who has both can make a more comprehensive offering.
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 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.
- 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)
- Ex-Citadel/D.E. Shaw team building AI-native infrastructure to turn lots of insurance data—structured and unstructured—into decision-grade plumbing that helps casualty risk and insurance liabilities move is looking for a data scientist with classical and generative/agentic ML experience. You will develop, refine, and productionize the company’s core models. (NYC, Boston)
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.