The Efficient Frontier of Automation

Plus! Negotiations; The Line Segment; Tokenized Trading; Financial Engineering; Trump Straddle

Byrne Hobart Byrne Hobart

In this issue:

audio-thumbnail
The Diff January 26th 2026
0:00
/779.128163

Ask questions about this article on Read.Haus.

The Efficient Frontier of Automation

A few years ago, when people talked about technology's impact on the labor market, the concern was about automating jobs. But now, white-collar workers—disproportionately represented in the discourse on this kind of thing[1]—spend more time talking about automating tasks.

Historically, one of the broad stories of economic growth has been carefully splitting the world into things that are mostly done by machines, perhaps with human supervision, and the complementary tasks that are most done by humans. In that world, the fields that see high productivity growth are the ones where you can neatly split these tasks up: it's easier to manage a fulfillment center when tasks like keeping track of what's on what shelf are fully automated, while tasks like placing three specific objects in a box are left to humans. A fast food restaurant will automate things like how long the fries get cooked, and faces some messy uncertainty over what the optimal mix of kiosks and front-of-house staff is. But once they know the mix, they know how to scale, and can assume a stable ratio of complementary investment between labor and capital. When there are shifts in this ratio, they're predictable: the fast-food chain finds a burger-flipping robot whose cost is slightly less than the cost of having a cook flip the burgers, so they bump up the capital/labor ratio but otherwise proceed as normal.[2]

The jobs that can best adapt to changes in how automatable they are are the jobs where there's a fairly straightforward continuum of task scope, and a dividing line where it makes sense to switch from manual to automated work. Programming is on this continuum, but not at an extreme. It's easy to come up with examples where there's a straightforward tradeoff between programmer time and machine time: if you're writing a parser, you could implement a slow general-case solution, but you could also implement specific solutions for particularly common cases. There's no general right answer here, though in practice the ideal answer is probably something like: write the general-case answer, look at the logs, write a special handler for whichever special case takes up the most time (compute), and repeat until the time investment doesn't justify shaving incremental money off your AWS bill. And the real practical answer is to roughly guess at these numbers. (If you care about AI safety, this will make you want to crawl out of your skin, but: if you're purely trying to maximize ROI, writing a bot that tracks how long these special cases take to implement, and letting the bot tell you when the ROI is too low, is probably ideal. Except that in a world where you can do that, you can probably just set Claude Code loose on the general problem instead!)

In some cases, there's a continuum of effort. Stone carving is an interesting example because the process at every stage can be described as "remove a certain quantity of stone from a bigger chunk of stone." Robots can do the bulkier bits of this, but human artisans need to handle the details. And there's some boundary where you'd want to switch from a machine to a person. But if you're also collecting data as that person works, the boundary keeps shifting. And, every time it shifts, the incremental cost of more carving goes down. In this case, there's actually a pretty good read on demand: people like beautiful buildings, but don't like to pay for them.[3] If human carving hours per dollar of revenue decline, and unit costs decline, we'll have nicer-looking cities. (Disclosure: I'm an investor in Monumental Labs, which is busy making this happen.)

In physical tasks where people operate machines by giving the machines a series of instructions to either execute repetitively or with more precision (or power) than a person can manage, you could see the automation moving slowly up in terms of scale; surgeons who control a machine that will cut precisely without any tremors might start controlling them in a more abstract way. But in that job, complexity is also moving downward on the diagnosis and planning sides.

(One result of this is that medicine will be a case study in a new kind of job category: AI rentiers whose credentials, and whose ability to be sued for malpractice, makes them an indispensable link in a supply chain that otherwise doesn't need them. Even if someone builds a robotic brain surgeon that handles the entire process from spotting the problem to cutting it out, and whose bedside manner includes a sycophancy level calibrated in real time by tracking your biometric data—just a little nicer until your pulse slows down another three beats per second!—it will be hard to find a legal way to sell it. And that means that there's little reason to build it.)

When tasks aren't distributed on a continuum like that, the pace of automation is less predictable. ChatGPT has reduced the amount of time needed for researching any given piece for this newsletter, but has also expanded the scope of things worth researching. And it's a long way from "Write X words in the style of The Diff with your usual number of em-dashes but way fewer contrastive parallelisms." That doesn't reduce the labor input required to write—it just produces identifiably LLM-influenced text.

And it's even harder for the burstiest jobs. Certain executives, creatives, and investors basically have the job of saying "Yes" to the right thing a handful of times, and spending the rest of their time maximizing the yes-able surface area. In one sense, those are the hardest jobs to automate because the sample size of good decisions is so small. In another sense, they're the easiest to automate: if someone has enough agents working nonstop to identify good opportunities, maybe their edge will turn out to be that they spot all the good options first. But at that point, there's still the question of access: if you have ten times as many AI agents as a top-tier venture fund does, so you spot a good company to invest in a few days before it's on their radar, willingness to write a check isn't what it takes to get on the cap table. In fact, as some parts of these jobs get automated, the rest of them get more valuable.[4]

It's going to be a weird labor market for a while, but even for people whose skills eventually get automated away, the early part of the process won't be too bad: if they're able to provide training data that leads to a model that eventually replaces their work, then the value of the work they do is the direct value add plus the present value of all future value creation from automating those skills. In a way, it's a compressed form of a career plus retirement, where you spend a while producing more value than you consume so you can store up enough savings to consume without producing much. To the extent that that happens in the aggregate, and pays off, the result is just like the result of working a full career: the country's richer than it was when you started, you got paid to help move that process along, and there's enough wealth to go around at the end that you'll be fine. But remember: if company-you've-never-heard-of-dot-AI offers to triple your pay as long as you wear an elaborate camera rig while doing your work, be sure to save all the extra money because there's a good chance you'll be retiring soon.

  1. Even the people who speak on behalf of blue collar workers are not necessarily such workers themselves; there's a division of labor where the most efficient way to do this is to split the work between actual blue collar work and a Laptop Job advocating for their interests. ↩︎

  2. This is a good illustrative example of the nuances of capital substituting for labor. The immediate impact is that the restaurant needs fewer worker-hours. But also, if that increases their margins, there are now more potential locations to build. The long-term impact on overall employment depends on questions like how much of this expansion results in a shift from home cooking to restaurants, whether or not labor becomes a stronger complement to capital as the total output from that capital rises, and, stepping back even further, what that complementarity looks like across the entire economy. A tricky question, and the kind whose answer will potentially change by the time you've figured it out. ↩︎

  3. It’s one of those annoying second-order effects that was well worth the cost: lots of smallish cities in the US used to have a beautiful neoclassical building in the middle of town. It was a bank, and the columns and statues were the bank's way of saying "We're not going to run off with your money and leave this nice- and expensive-looking building behind!" It's a bit like when mid-80s tax reform lowered the top marginal rate and enforced partial rather than full deductibility for business meals, which killed the three-martini lunch. ↩︎

  4. In a recent appearance on Uncapped, Ben Horowitz poignantly illustrates how seeing and selecting the right deals is essentially useless if you can’t win those deals, and why winning the best deals is the most important, non-reproducible factor in generating superior venture returns. A platform like a16z has a compounding advantage here. The better they are at winning deals, the better they become at attracting the best pickers: the best pickers (investors) want to join a firm that can win the deals they find. ↩︎

Diff Jobs

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

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.

Find a Role

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.

Find Talent with Diff Jobs

Elsewhere

Negotiations

The usual Trump approach to tariffs is to start with a crazy number, then work backward to a more modest one that's coupled with either promises to purchase American-made goods or promises to make goods in America. These negotiations are basically happening through governments but with the private or quasi-private companies that will actually implement them. The companies in question will probably get a better deal if they push for the exact terms they want. So Volkswagen says it won't build a new Audi plant in the US—one it's considered for years—unless tariffs on cars come down. This directly addresses part of the pro-tariff argument, that international companies will relocate production to the US in order to avoid paying the cost of tariffs. That's true if tariffs are a single-stage game where they're permanently set at some level that just barely offsets the cost advantage of manufacturing elsewhere. But if they're a continuous negotiation, expect people to negotiate.

The Line Segment

Saudi Arabia has scaled back its Neom project ($, FT). For countries whose economy is focused on extracting a single natural resource, there are basically three bets they can make:

  1. Move up the value chain, from extraction to refining, basic chemicals, polymers, and onward. Saudi Arabia does some of this, but inconveniently for them, refining is a global industry, shipping crude oil is relatively cheap, and shipping refined products requires a little more bulk or, later in the process, specialized infrastructure. And even if you have a captive buyer for your crude, that means you have a locked-in seller, so you just shift some of the cyclical volatility from oil to petrochemicals.
  2. Use oil industry profits to accumulate capital, and invest that in financial assets whose returns aren't especially correlated to oil. (You could try to invest in products that hedge the risk of a secular decline in oil consumption, like renewables, but doing too much of that risks accelerating the decline of oil.)
  3. Build a service sector, which is insulated from Dutch Disease but which won't necessarily experience long-term productivity growth.

One of the country's problems is that it tried to do all three at once, but that meant that any decline in oil revenue would force them to cut something, and a service sector, with its internal complementarities and dependence on the expectation of future revenue, was the easiest to cut.

Tokenized Trading

Several overseas crypto exchanges are trying to trade tokenized representations of US equities ($, The Information). One argument for this is that it's convenient to put more assets on crypto rails, which can afford more flexibility and lower transaction costs than working through various countries' fiat systems. The other argument is that some market participants would be quite interested in trading US stocks using a source of either funds or trade ideas that Charles Schwab and the like would probably report to regulators. As The Diff noted earlier ($), individual prediction markets can't ban insider trading because people will arbitrage across markets; as long as there's any market whose specifications match that of a regulated market, the overall market will reflect the participation of well-informed counterparties. So, the long-term result of this will be regulatory convergence of one kind or another; either zero tolerance for listing tokenized equities for any venue that doesn't report suspicious trades to the SEC, or relaxing the prohibition on such trading. In the meantime, it will be fun for options market-makers to take deviations between the price of a tokenized stock and the price of the underlying into account when quoting options—pending M&A will leak when a tokenized bet on a stock trades at a premium to what someone who's trading under their legal name at a US-regulated venue is willing to pay.

Financial Engineering

IonQ is a quantum computing company that is a small business if you look at trailing revenue and a very big one from the perspective of its market cap. It popped up in The Diff a few weeks ago because it had announced a fundraise in a very misleading way (they raised at a discount to market value, not the premium they cite in their press release). Now, they're acquiring a chip company for stock. They'll give up 4.4% to 6.7% of their shares, in exchange for which they'll acquire a profitable company whose trailing revenue 4.3x theirs. But that range is interesting:

The stock component is subject to a collar under which SkyWater shareholders will receive IonQ stock valued at $20.00 per SkyWater share, based on the 20-day volume weighted average price of IonQ stock as of three business days before closing, unless such volume-weighted average is greater than $60.13 per share, in which case SkyWater shareholders will receive 0.3326 IonQ shares per SkyWater share, or less than $37.99 per share, in which case SkyWater shareholders will receive 0.5265 IonQ shares per SkyWater share.

Within a range, IonQ is willing to compensate the seller for how volatility the currency they're selling with is. But IonQ won't sign up for arbitrary levels of dilution when the range of outcomes is so wide.

Disclosure: Short a small amount of IONQ, as part of a larger basket of short positions on speculative froth.

Trump Straddle

The FT notes that investors have learned from Liberation Day, and simply didn't react all that much when Donald Trump hinted that he'd consider launching a war of conquest against a NATO member ($, FT). And, to be fair, tariffs were a lot more serious than Greenland, and at least some of the market's recovery since tariffs were announced has been driven by AI companies that aren't as affected by tariffs. Still, it creates a perverse dynamic: Trump reacts to market indicators, but traders know this, so the investor population is basically tripping on its own feet to signal negative feedback to Trump while buying ahead of the response to that negative feedback. Any time there's some force that causes a price to rubber-band back from most moves, it raises the probability that a big move will break that feedback mechanism and lead to a bigger swing. It's hard to eliminate volatility, but relatively easier to move it to the tails of the distribution.