The Robots Will Eat the World

A decade ago, Venkatesh Rao published "Breaking Smart", a set of 20 essays on how software is “eating the world”.

Most of his thesis hasn't aged well, coming off as techbro arrogance, but it had an important nugget that stuck with me: He posits that prior to software, two ideas in history “ate the world”: the written word, and money.

Eating the world was a new concept for me - it’s the idea that these ideas apply to absolutely everything. Writing describes everything. Money values everything (as much as we hate that) - both of these can be applied, and now are applied, to everything in the world. Now software is on its way to processing everything. Nearly all the rules we live and work by are modeled in software already: how much we get paid, how we communicate with each other, the shape of the sidewalk we walk on, how many bananas the corner shop should order next week.

This eating takes time: Literacy didn't reach a majority of people for more than 4000 years after the first written word, with a jump over the last 200 years. Money took 2000 years to enter common daily use by a majority, with a similar ending. Software, though, is less than a century old, and in the last decade crossed the same line - probably around 2017, the year Google published the Transformer paper that led to Large Language Models.

Like many, I think these LLMs will finish the work of software eating the world. But I think most boosters make a mistake here: they assume agents will be our automation, directly replacing workers. I disagree; we know they hallucinate, they’re unreliable, expensive to run, and difficult to constrain. Instead, I think we’ll use them for what they’re very good at: writing custom software.

Now–2030: Cheap Verification

Writing code by hand is already going the way of assembly; it’s becoming the work of specialists and hobbyists. This is uncomfortable, every new model leaves fewer of my engineer friends even reviewing code directly.

Instead, we're writing business rules and specifications, goals and tenets, to constrain and guide coding agents. Harnesses are already evolving to visualize our decisions and user journeys, and both display and react to business and operational metrics. It’s a good time to be an experienced product manager!

Even the coding agents will continue to hallucinate, make up answers and features from whole cloth, regress behavior and cover it up. But we know how to handle this: testing. I started my software career as a tester, and first learned to code by writing automation to replace my manual tests. An agent can build tests as cheaply as any other code, and the owners of the business and UX can codify what's right.

This is where we get to the crux of my argument, and where Jevons paradox comes into play: as software gets cheap, more people will write software. Product management, UX research and design, program management, and sales can build whatever they dream of. Enterprise sales teams often already build their own product customizations - that pattern will explode, and with it, tons of new work to make sure that software stays aligned with corporate strategies.

As we’re seeing, agents won't replace white collar workers directly. But as custom software gets cheaper, we'll see much of the need for hiring replaced by software to execute verifiable, repeatable tasks. We'll automate from the bottom up: junior devs, paralegals, analysts, support, coordinators: the commodity tier of every desk profession will dwindle.

Companies won’t get to bank the savings. Their competitors will get the same tools, so the savings will get competed away. Every software moat will dissolve; work rationed by cost stops being rationed.

Diligence that only a Fortune 500 could have justified last year will get run on a $2M acquisition tomorrow. A plumbing contractor will get a custom scheduling system that used to mean an enterprise and a “six month” integration that doubled in cost and time. Work priced for the few will be priced for everyone, and we will collectively do vastly more of it.

As software itself stops being a moat, this shifts knowledge work to other roles - in tech, we’ll hire more product managers, UX researchers, program managers, and salespeople. And an increasing share of capital will flow to the LLM shovels: from NVIDIA, Broadcom, and AMD to TSMC's fabs, SK Hynix and Samsung's memory, ASML's lithography, and the entire stack of energy generation and grid interconnect beneath them.

But the world is a physical place. You can’t eat it all from a screen; the real work is past the information economy.

2030–2035: Training the world

Robots can do a lot today, but they can't yet replace humans for most tasks. As I write this, dancing robots still kick people and fall over: they can't yet handle much variation in environment, and few can safely work next to humans. But companies like Universal Robots and Agility Robotics are starting to ship products that improve on safety issues. As the cost of model training continues to collapse, we'll improve at training them to handle environmental variation using the next step from LLMs: Large Behavior Models (LBMs) and Large World Models (LWMs) model the physical world, and enable robotics rather than screens.

The data necessary to train these can’t be found in a book, though - it’s dependent on sensors in the wild. Hundreds of millions of people already train some of these models remotely, spread across the world, by snapping photos in Pokemon Go, ringing doorbells, and driving their cars.

Companies are investing heavily in training models to understand their repeatable, verifiable tasks, and using those models to produce automation. The economics are unforgiving - if you're paying a person $50,000 a year, but a robot costs $50,000 once plus a $10,000/year service contract, humans will be relegated to specialists and hobbyists, just as many industries have already shifted.

But even with pressure from competition and from organizing, this transition will be slow - except for driving.

2035 - 2040: Driving down costs

Insurers already know self-driven miles are safer than human driven miles. The actuarial tables for insurers will soon demand differing costs for self-driving than for human driving.

The next step will be a feedback loop. When a human can flip a switch that lowers their insurance costs, they will; they already install apps that monitor driving for a discount! It won’t be long before the car tells the insurance app when it’s in control, and targets behaviors the insurance company favors.

When your insurance cost drops 50%, someone else’s will go up 50% to match their risk profile. Waymo rides just got to price parity with human Uber drivers. Five years ago we still joked that self driving was ten years away, but your next car will likely have this technology, and you’ll have to pay $200 more a month not to use it.

It’s easy to be pessimistic. When cars replaced horses, demand for transport exploded, and the American horse population fell from some 26 million to 3 million. People fear this happening to human work, but most of what these robots take was never a job in the first place.

Municipalities and banks will seal the fate of human driving: As soon as they have a choice, cities won’t let human drivers near street fairs and farmers markets. When you pay for a second beer, your car won’t give you control anymore (and you’ll lose a healthcare discount, but that’s another story). Parents in affluent neighborhoods will demand safety for their children next. Banning human drivers wherever possible will save thousands of lives.

But while we’ve been talking about wheels, robots will get their feet under them; and Jevons paradox will walk in your front door.

2040 - 2050: Coming clean

Paid housekeepers in America number in the low millions. But people cleaning their own homes number north of a hundred million. Nearly all of that labor is unpaid household work, and humanoid robots will be able to use all the tools you can to do it for you.

In “More Work for Mother”, Ruth Schwartz Cowan documented that a century of labor-saving appliances never reduced hours spent on housework, because standards rose to consume the time: expectations of clean rose as fast as the ability to deliver it. Cheap cleaning creates an appetite for cleanliness that didn't previously exist, and that appetite will grow the robotics industry faster than the housekeepers it replaces.

This pattern is everywhere: Cheap freight started local, moving factories to the edges of cities first, and then moved them to the other side of the world. When containerization collapsed the cost of ocean freight, world trade rose by an order of magnitude. The latent appetite for shipping is still growing.

Global manufacturing employment rose enormously in the decades after containers, far outstripping the longshore jobs lost. They landed in Shenzhen and Busan rather than Newark - places where the conditions were right for manufacturing innovation. These changes pulled a billion of people across the world out of poverty.

At the same time, the US moved from manufacturing to an even better future: the space these left let us build an information economy that benefited city regions across the US.

The Optimistic View

Jane Jacobs named those conditions better than anyone. A vital city region sheds its older, standardized work to other places, and the shedding frees the labor, capital, and attention for newer work that was already trying to grow there. Regions that replace their own imports regenerate as a matter of course.

So, as you might guess if you know me, this essay comes back to cities. Not just New York and the Bay Area, but Seattle, Austin, Denver, Minneapolis and many more - places with enough population density to generate new work. As innovation keeps compounding, these regions will absorb it and grow.

The optimists have been right about aggregate employment nearly every time. The idea of a fixed quantity of work is a fallacy every generation fears. However, when and about whom will always be messy. We just saw the US government ban not just DJI drones, but Roombas too. There will be a fight over who builds these robots, but they’ll be commodities. The software will be developed, using all these varieties of model, in the United States, where we have the unique financial and regulatory conditions that have built Microsoft, Google, Apple, Nvidia, and Amazon.

At the same time we’re having these shifts, rich economies will deepen into their demographic aging waves. Care work will continue to grow as a share of employment. I hope political efforts win free schooling and healthcare subsidies for care jobs; they’ll be a reliable labor sink because they’re as important for social needs as physical. Elder quality of life will improve: robot cars will reduce isolation and home robots will clean for far more people than could afford a housekeeper. Fewer children will need to leave the workforce to care for a parent.

By 2050, the median household in a productive metro will buy things its 2026 counterpart couldn't have imagined: proactive legal representation, continuous medical monitoring, a dedicated tutor for every child, custom software for anything they want, and a home that cleans and maintains itself to a standard nobody in 2026 would have thought to want.

The robots are coming, and your life will get easier.