Robots Are Coming for Jobs That Don't Exist Yet

Plus! Scale; Preorders; Politics; Funding; Investor Relations

Byrne Hobart Byrne Hobart

In this issue:

audio-thumbnail
The Diff October 26th 2026
0:00
/740.832653

Talk to this post on Read.Haus.

Robots Are Coming for Jobs That Don't Exist Yet

It's always interesting to look at which constraints shape a product, long after the original constraint actually applies. Coding style guides often specify 80-character lines because early displays were often 80 characters wide, and that was so they could display IBM punched cards, and the reason those cards were eighty columns wide was that that's how many you could easily fit into something with the same dimensions as a dollar bill, because that standard was chosen for the census and the Federal government had lots of dollar-sized storage containers around.

Similarly—if a little briefer—many LED lightbulbs are roughly pear-shaped. Incandescent bulbs needed to be evacuated or filled with an inert gas, which meant that there was a pressure differential between them and the air outside. A roughly-spherical shape, or a shape that's spherical except at the part that's additionally secured with metal, is better at resisting pressure, so it needs less material to achieve a given level of durability. (It also helps that it's a natural shape for blown glass to take.) Now, that rounded bulb shape is a mix of backwards-compatibility and aesthetics; consumers know what a light bulb looks like, even if it doesn't have to look like that any more.

This shows up when the constraint is in labor markets. In the early stages of the industrial revolution, there was synergy between mechanization and a feature of the labor market: while the entire household had to work for them to survive (remember, Malthus was right about the centuries leading up to An Essay on the Principle of Population), the market for men's labor was much more developed. A man in reasonably good health could get a job pushing, pulling, carrying, chopping, shoveling, etc. From a labor market perspective, the big impact of early automation was that it created jobs that didn't require lots of upper-body strength. And for employers, this also meant that it was okay to be the low bidder, because they were the only bidder. It was actually quite similar to the way AI impacts labor markets today, albeit at a slower speed: a leveler in some respects (your delivery or rideshare driver doesn't need to speak your language to communicate via auto-translated text), leverage for others, and a process that didn't suddenly eliminate legacy jobs. Textile mills still had plenty of tasks that required physical strength but couldn't easily be done by a machine, but they also had tasks that previously required physical strength that could be much more easily mechanized.

In some ways, this was preserving the status quo and making it legible: The textile workforce has been predominately female for a long time.[1] As a craft, it's compatible with bursts of work between other tasks, and can alternate between maintenance for the household and piecework sold elsewhere. Kids have, for much of economic history, been expected to contribute at least something to the household's economic outputs, even if they weren't historically high-ROI. They’d end up doing whatever required being small and agile, instead of big and strong. But when the formal labor market was the high bidder for their time, it could start designing things around them; a piece of equipment that an adult male needs to be able to scramble under to un-jam is going to take up more space than one designed so a second-grader can do that. A mine can be more cramped if the average worker is smaller. Cramming more workers into a given space raises returns for the factory or mine owner, so that's what happened.

At least for a while; over time, changes in both economics and social norms changed the composition of the factory workforce, which is about 70% male in the US and 60% in China. There are still kids working in factories, but generally not in places that keep reliable statistics. That kind of work is unpleasant now—it's repetitive, noisy, and sometimes dangerous—and was pretty hellish in the beginning (but the fact that these factories were able to hire workers at all tells you something about just how bad subsistence farming is).

A modern factory is more pleasant, and has ridiculously higher output than what late 18th-century contraptions could manage. Industrialization created a market for extremely low-wage labor, and then created enough wealth to render that market obsolete—and then kept on growing, and to this day continues to provide an entry point for workers in impoverished countries to indirectly plug into the global economic system.

But what happens when the equipment isn't bolted to the factory floor, but gets deployed wherever it needs to be. Specifically: how much potential economic activity is there that's so unpleasant, dangerous, disgusting, boring, or uncomfortable that it's done by workers who get paid an enormous premium to put up with it, or by incredibly desperate people in poor places, or, mostly, not done at all?

One feature of conventional land mines is that they stick around until they explode; nobody tells them the war's over and that the pressure setting off their detonator is coming, not from an enemy soldier, but from a kid playing soccer. Finding mines is tedious, and dangerous. Mines are typically deployed to deny enemy troops access to some area, but in a few different conflicts—the Soviet invasion of Afghanistan, Mozambique's civil war, Angola's, and Vietnam in Cambodia. Automating the process of removing these mines is partly a machine-learning problem—a bit like drilling oil: if you find a mine in one place, it's likely that there's more underground nearby. But there's also some irreducible risk of setting one off by accident. Drones and mine-removal robots are worth a lot less than human lives, and in this particular case that affects both supply and demand.

As a general rule, if there's a job that's done on land, it's very expensive to do the same thing underwater, particularly for extended periods of time. You have to pay specialists, and there's extra commuting time from them getting their gear on and then removing it later. Underwater maintenance and repair, of ships, pipelines, and undersea cables, still happens. But its expense is a constraint on how many of those things we can have. Autonomous underwater robots will go where the money is, like offshore oil,

On land, British railroads are using drones to inspect cramped culverts, some built up to 150 years ago with dimensions that implied that either they weren't meant to be in use indefinitely or the builders assumed that there would always be a supply of sturdy urchins to clean them out.

Some kinds of maintenance are not so much deadly as inconvenient. Solar panels do best when they get direct sunlight, and since they take up space it's also a good idea to put them somewhere where land is cheap. But "a desert in the middle of nowhere" is a place where it's not easy to source cheap labor, but where it is easy for solar panels to accumulate dust. So there are solar panel-cleaning robots that can do the job. This is a nice energy-in/energy-out situation, and there's a further symmetry given that the robot has an upfront cost and then produces energy, or at least counterfactually-not-not-generated energy, over time. So wrapping it into existing solar financing is straightforward.

There are lots of environments, at many different scales, that are both inhospitable and lucrative. They're a big enough obstacle that there has to be a strong economic incentive to convince people to stick around. But as the dirtiest and most dangerous jobs get automated, economic activity will start to migrate to those places because the land is cheap and you don't need to pay for people. It'll be a very strange reordering of the economy, but not an unprecedented one: if you looked at a sufficiently high-resolution map of where wealth is created, there would be various middle-of-nowhere bright spots corresponding to oil, metals, and minerals. There might end up being additional ones in the future that are defined by some mix of still-low real estate prices and a network effect from existing infrastructure. A great measure of economic progress is how many unpleasant jobs have been either automated away or become a small, better-paid fraction of the workforce.

  1. Here's an epitaph, via SPQR, from over two millennia ago: "Here is the unlovely grave of a lovely woman... She loved her husband with her heart. She bore two sons. One of these she leaves on earth, the other under the earth. She was graceful in her speech and elegant in her step. She kept the home. She made wool. That’s what there is to say." ↩︎

You're on the free list for The Diff. This week, paying subscribers read about why it's hard to define a unit of progress for AI ($), Bryan Johnson as a template for future medical data rentiers ($), and the wobbly path AI takes through writing- and talking-focused organizations ($). Upgrade today for full access!

Upgrade Today

Also, time for some audience participation: if you're working on an early-stage startup that's vaguely aligned with this thesis, or might benefit from it, Anomaly would love to chat about it.

Contact Anomaly

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

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.

Contact Anomaly

Elsewhere

Scale

At the point of consumption, delivery and groceries are perfect substitutes—a meal's worth of one removes a meal's worth of demand from a fairly fixed market. But on the supply side, they're complements, because logistics markets benefit from scale, and from the ability to price-discriminate based on latency. So it makes sense that a growing Instacart would be a separate business from food delivery, and that as it matured it would converge on them. One interesting feature of this is that grocery spending tends to involve more discrete SKUs than restaurant food, so it's a richer source of data to target users with high-margin restaurant offers. These businesses have many layers of returns to scale, which makes merger talk inevitable. It also makes the merger hard to negotiate, because it's hard to prove which side provides relatively more of the synergy until after the deal is consummated.

Preorders

Amazon says AWS capacity is mostly booked through 2027 ($, The Information). One point he makes is that Amazon only needs a few months of lead time for new equipment, so if their commitments are sufficiently far in the future, they're safer from overexpanding. Cloud computing companies are a mechanism by which queuing on the part of AI buyers creates more supply to break the logjam among sellers. In a high-demand scenario, selling more at the spot rate tends to be profitable; AWS is, at least when they're growing fast, short a little compute while being mostly long in their existing portfolio.

Disclosure: long AMZN.

Politics

This morning, Brazilian equities are down by about a quarter after the re-election of t Luiz Inácio Lula da Silva ($, FT). As of late last week, Polymarket's odds on this happening were 20%, implying that there would have been a ~5% rally had the election gone as expected. This is a surprisingly big gap in equity values for two people who have spent a fair amount of time statistically tied in the polls. The market/politics feedback loop is surprisingly consistent across different geographies: in a developed country with widespread equity ownership, it's a big deal if a politician says something that causes stocks to instantly drop 2% or something, and a move like the Liberation Day collapse was unprecedented. In countries where there's a less direct connection between capital markets and people's everyday lives, it's more common for politicians to push the market around a bit, whether accidentally or intentionally. But those countries run into real constraints on policy much faster; if external trade partners and lenders are junior creditors to the safety net ($, Economist), their getting spooked has an immediate and painful impact on the economy.

Funding

Elon Musk says that the credit funding environment is tough for companies that don't have non-AI revenue. This is a very convenient thing to be true for a rocket/car company executive with a neocloud side project. It implies one of two possibilities:

  1. Lenders are skeptical about end demand, and are worried that neoclouds get their revenue from activities that are unprofitable now and won't necessarily work better in the future.
  2. Neoclouds are disproportionately vulnerable to funding delays. They're the slice of the capital structure that takes first loss on these, because they're still paying interest (and potentially penalties from their future tenants), but not collecting revenue.

The second situation is actually pretty good for SpaceX, because building big infrastructure ahead of schedule is their core competency. But it's a different bear case from the demand-driven one that would lead SpaceX's AI assets to be just as stranded as everyone else's.

Investor Relations

My first day on the job at a hedge fund was the Monday after the Facebook IPO in 2012, so one of the only constants of my career since then is that, a couple times a year, Facebook or Meta is trading slightly higher after-hours on a usage-driven revenue beat, only to open down the next day because they guided to aggressive opex or, more recently, capex. For most of that time period, the reaction was because Meta's free cash flow wasn't going to be quite as high, quite as soon, as the market expected; it's only recently that Meta's been back to a point where it needs to raise capital. But that's happening; they're raising a total of $35b in equity and converts right now, and planning a $60bn at-the-market offering for early next year. So far, their shares are pretty unpreturbed; investors knew they'd do something like this, and have seen a peer do an offering of similar scale. For Meta (and Google), there are more places where they can spend on inference and get a predictable revenue lift, and they're slowly contorting themselves into maximizing that surface area even more. So an at-the-market offering makes a lot of sense: it means that equity investors can continuously allocate more capital to AI as opportunities emerge.

Disclosure: long META, GOOGL.