In 1850, around 70% of people worked in agriculture.
Today, it's less than 5%.
And nobody ended up unemployed.
That single fact should change how you read every headline screaming that AI is about to wipe out work as we know it. We've run this experiment before — many times — and the result keeps coming back the same. Technology doesn't destroy work. It moves it.
I'm not writing this to tell you everything is fine and you can ignore what's happening. The opposite. The panic is the wrong response and so is complacency. The honest answer sits in the uncomfortable middle: your job won't be replaced, but the shape of it will change — and whether that change works for you or against you depends entirely on what you do now.
A quick word about my own posts
If you've followed what I write here, you've seen the harder-edged stuff: layoffs at Coinbase, code becoming a commodity, "how do you become irreplaceable in the 40% AI can't touch?"
I've noticed a lot of people read those the wrong way — as a warning that AI is coming to replace them, personally, and they'd better run.
That's not what I'm saying. It never was.
"Code is becoming a commodity" does not mean "developers are becoming obsolete." It means the execution layer is getting cheap while the judgment layer gets more valuable. "Become irreplaceable in the 40%" isn't a threat — it's a map. The posts that sound like alarms are actually descriptions of where the value is moving. If you read them as "the end is near," you'll freeze. If you read them as "here's where to stand," you'll move. This post is the long version of the second reading.
The oldest mistake in economics
Most fear about AI rests on a hidden assumption: that there's a fixed amount of work in the world. A finite pie. If a machine takes a slice, there's less left for you.
Economists have a name for this. It's the lump of labor fallacy, and it's been wrong for two centuries straight.
The amount of work to be done has never been fixed, because human desire has no ceiling. We are never satisfied. The moment one need is met cheaply, we invent three more. So when a technology absorbs a chunk of existing work, the time and money freed up don't vanish — they get redirected toward things that weren't viable before. New products. New services. Entire professions nobody could have pictured.
This is the part people miss. They can clearly see the job being automated today. They cannot see the jobs that automation makes possible tomorrow, because those jobs don't have names yet. The loss is visible and concrete. The gain is invisible and diffuse. So we weight the loss far more heavily — and we panic.
The twist nobody expects: Jevons' paradox
There's an even stranger pattern worth knowing.
In the 19th century, economist William Stanley Jevons noticed something odd about coal. As steam engines became more efficient and burned less coal per unit of work, total coal consumption didn't fall. It rose.
Why? Efficiency made coal cheaper to use, which opened up applications that weren't worth it before. More uses, more demand, more consumption — not less.
This is the Jevons paradox, and it applies directly to human labor. If AI makes a worker dramatically more productive — and cheaper per unit of output — the likely result isn't less demand for that worker. It's more, spread across new kinds of work that suddenly make economic sense.
Even Dario Amodei, the CEO of Anthropic and someone with genuinely cautious views about where AI is headed, points to this same dynamic: AI as an accelerant to productivity, letting teams do far more even as they grow.
The receipts: this has happened before
Theory is nice. History is better. Here's what actually played out the last few times a transformative technology arrived.
Agriculture. As machines mechanized farming, the share of workers in agriculture collapsed from roughly 70% in 1850 to under 5% today. Crop prices fell dramatically. The displaced labor didn't sit idle — it flowed into industries that didn't exist before.
Electricity. Electrification made production radically cheaper and faster. The cost of durable goods — refrigerators, washing machines, radios — plummeted, freeing enormous amounts of time once spent on manual chores. That freed time and money flowed somewhere. It always does.
The automobile. US car production jumped from under 200,000 vehicles in 1910 to nearly 4 million by 1925. Employment in the sector exploded from ~50,000 workers to more than 400,000 in the same span, while the real price of a car fell. The auto mechanic — a profession that barely existed in 1900 — became one of the most in-demand jobs of the century.
And my favorite: the spreadsheet. When electronic spreadsheets took off in the 1980s, everyone assumed accountants were finished. The number of clerks manually entering numbers did fall. But the number of accountants, auditors, and analysts soared. After Microsoft added VBA to Excel in the mid-90s, US accounting employment climbed from under 700,000 jobs to over a million.
The spreadsheet didn't kill the accountant. It killed the boring part — manual data entry — and promoted the professional to higher-value work. The tool didn't replace the person. It moved the person up a level.
That last one is the closest mirror to what AI is doing right now.
But let's be honest — the transition is real, and it has losers
This is where I have to be straight with you, because "it'll all work out" is its own kind of lie.
Every example above came with a painful transition. "Agriculture went from 70% to 5%" is a clean sentence that hides decades of displaced families and towns that never recovered. The aggregate worked out. Individual people did not always work out. Aggregate optimism is cold comfort if you're the one being displaced this quarter.
This wave may be faster than the others. Electrification took decades to diffuse. AI tooling diffuses in months. A faster transition means less time to retrain, less time for new industries to absorb the displaced. Speed is the genuine risk here — not the direction.
"New jobs will appear" is a historical pattern, not a physical law. It has held for 200 years. That's strong evidence, not a guarantee. Anyone who tells you the future is certain — in either direction — is selling something.
So I hold both things at once: the apocalypse framing is wrong, and the disruption is real for real people. Calm is not the same as complacent.
So what does AI actually change?
The honest version, without the doom and without the hype:
AI won't replace you. But it will replace the parts of your work that could be automated. That distinction is everything.
Writing code, drafting copy, generating a first image, reconciling a spreadsheet — the execution layer is becoming cheap and fast. What used to be a differentiator (knowing how to do the thing) is becoming a baseline available to anyone.
What rises in value is everything the model can't do alone: figuring out which problem is worth solving, exercising judgment, understanding messy human context, deciding what to build and for whom, and orchestrating these tools toward something real. That's the part that doesn't become a commodity.
A Goldman Sachs analysis makes the pattern concrete. The jobs most exposed to automation — telemarketers, payroll clerks, certain paralegal tasks — are narrow and rule-based. The roles most likely to be augmented rather than replaced — engineers, managers, doctors, lawyers, executives — are the ones built on judgment and responsibility.
And here's a fact that cuts directly against the panic: demand for software engineers, supposedly the most "doomed" profession, has been climbing again through 2025. The people closest to the technology are being pulled in by it, not pushed out.
Why machines won't "take over"
One more thing, because it gets lost in the science-fiction framing.
Machines don't have will. No desire, no agenda of their own. Humans choose and act; machines execute. A model is a tool that exists to help us meet our needs — and our needs are infinite. If some are now met more cheaply, that frees time and resources for needs we didn't even know we had.
The goal of human work was never to keep people busy. It was to satisfy human needs. There's an old Milton Friedman story about this: visiting a canal site where workers moved earth with shovels, he asked why they weren't using machines. Told it was a jobs program, he replied — "Then why not give them spoons instead of shovels?"
If the point is to build the canal, you use the best tool available. Refusing better tools to protect "work" for its own sake is just choosing spoons over shovels. You won't do that voluntarily — and neither will anyone competing with you.
What to actually do — depending on where you are
If you're a developer
The goal isn't to stop coding. It's to close the gap between "I can build this" and "I understand why this is worth building."
- Use AI tools seriously, today. You can't develop judgment about a tool you've never touched.
- Find your 40% — the judgment, the context, the problem definition that isn't automatable — and get deliberately better at it.
- Learn to evaluate AI output critically: is it correct? Secure? Actually what was needed?
If you're not technical
This is the moment the gap between you and "the builders" narrows the most.
- Build something small end-to-end with AI tools. Feel the ceiling firsthand — what agents can and can't do.
- Lean into your edge: you already understand a real problem and a real user. That's the scarce half.
If you're early in your career
You have an advantage: no habits to unlearn.
- Don't just learn to code — learn to build. Solve a real problem end-to-end.
- Develop taste and clear writing. Prompting an AI well and writing a clear spec are the same underlying skill.
The question worth sitting with
So no — it's not the apocalypse. The lump-of-labor fear is, historically and economically, unfounded. But that is not permission to sit back and watch.
The people who do well in this shift won't be the ones who learned the most frameworks or fought the technology hardest. They'll be the ones who started using it early, and who figured out where their own judgment is worth more than any model's output.
The machine does the work. You decide what's worth doing.
Jobs will change over the coming years — the tasks, the titles, the day-to-day — exactly as they always have. That part is certain. Which side of the transition you'll be on is not.
That part has always been a choice.