section 04 of 11

The Last Safe Job

The belief that physical labor is protected from AI rests on an intuition that feels solid: robots are clumsy, expensive, and limited. Anyone who has watched a robotic arm fumble with a doorknob or fail to fold a towel can be forgiven for thinking that the trades, agriculture, construction, caregiving — the work of human hands in unpredictable environments — is safe for decades to come.

This intuition is based on a misunderstanding of where the bottleneck actually is.

Consider a raspberry. Picking one is trivially easy for a human and has been nearly impossible for a machine. Not because we cannot build a manipulator with sufficient dexterity — we have had that capability for decades. The mechanical engineering of a gripper that can pluck a soft berry without crushing it is a solved problem. What we could not solve was everything that happens before the gripper moves. Which berry is ripe? Which is obscured by a leaf? This stem is at an unexpected angle — how should the approach change? The one underneath is overripe and will burst with the slightest pressure — how to reach past it? A cluster is denser than the training data anticipated — now what?

This is not a robotics problem. It is a perception-judgment-adaptation problem. It is, in other words, an AI problem. And it is the part that is now moving — from research demos toward commercial pilots in the field. Broad deployment is still years of unglamorous engineering away, and anyone promising a firm date is guessing. But the bottleneck that mattered has shifted from the mechanical to the cognitive, and the cognitive side is exactly where the fast progress is.

The same pattern holds across virtually every physical job that was considered safe. Plumbing requires diagnosing an unseen problem behind a wall, choosing among imperfect options, and adapting in real time when the situation turns out to be different from what was expected. Elderly care requires reading emotional states, adjusting physical support to an unpredictable body, and making judgment calls about safety and comfort in novel situations. Construction requires interpreting ambiguous plans, compensating for material imperfections, and coordinating with other workers in a constantly changing environment. In every case, the mechanical part — the lifting, gripping, cutting, fastening — is straightforward. The hard part was always the thinking.

The public conversation about robotics is fixated on the wrong image. When people picture a robot replacing a human worker, they picture a humanoid — a walking, bipedal machine shaped like a person. Boston Dynamics videos, Tesla’s Optimus prototype, science fiction androids. The humanoid form is dramatic, legible, and almost entirely beside the point. A purpose-built raspberry picker does not need legs, a torso, or a head. It needs vision, a specialized gripper, and a rail or wheeled platform. A cleaning system does not need to look like a janitor. It needs a fleet of low-profile specialized units with different attachments. The optimal form for almost any physical task is not human-shaped. It is shaped by the task itself.

The humanoid fixation actually slows the correct understanding of what is happening. It lets people think, “well, robots still can’t walk up stairs reliably, so we have time.” Meanwhile, the real deployment path — purpose-built machines that are mechanically simple but now carry the judgment layer that was always missing — is advancing much faster than any humanoid program and attracting far less attention.

There is a deeper structural difference that separates this from every previous wave of automation. Traditional industrial robots required painstaking hand-coded routines for every task. Teach a welding robot to weld one specific joint on one specific chassis, and it could do that single operation beautifully — but any variation required a human engineer to reprogram it from scratch. This made deployment linear: one task, one program, one long integration cycle. Scaling was slow because each new capability was a custom engineering project.

Foundation models loosen this constraint. A system built on a general model of the world does not have to be hand-programmed for every task from scratch; increasingly it can be shown or told a task and work out an approach, the way recent robot-control models trained on broad data generalize to actions they were never explicitly coded for. That bends the deployment curve. Where traditional robotics scaled one custom integration at a time, a capability learned once can spread as a software update rather than a mechanical redesign — a different, faster curve, even if “exponential” is more than anyone can honestly promise. The machine does not need to be rebuilt. It needs to be told.

The implications collapse the last structural argument for human indispensability. Cognitive work is exposed — that was clear by the end of the previous section. Creative work is exposed — AI systems already generate text, images, music, and code that compete with professional human output. And now physical work is exposed — not because robots have gotten dramatically better at moving, but because the judgment layer that held them back is arriving as a software capability that can be deployed to any machine with sensors and actuators.

There is no category of human labor left with an obvious structural moat. Cognitive, creative, physical — all three are now within reach of systems that keep improving, and faster than the last automation wave moved. The retreat from “AI can’t do creative work” to “AI can’t do reasoning” to “at least physical jobs are safe” has been a retreat across a shrinking island, and the water is now at every shore.

Exposure is not yet displacement. The standard economic reply is comparative advantage: even a machine that does everything better cannot do everything at once, so humans keep a role in whatever they are least bad at. That is true, and it guarantees a role, not a living. The wage for any task is capped by the cheapest other way to get it done. Offshoring moved work to the cheapest human who could do it. This moves it below the cheapest human, to the price of compute, and that price falls every year. Horses kept a comparative advantage too. Horses and mules on American farms still fell from about twenty-six million to under three and a half million in less than forty years, because the work they could still do was no longer worth their feed.

The natural response to this is to zoom out. Fine — individual jobs are vulnerable. But people lose jobs every year and find others. There are benefits to bridge the gap and programs to retrain. Losing a job is painful, but it is an unemployment problem, and unemployment is something modern societies know how to manage. Don’t they?

going deeper · optional

Why this holds the mechanism, and where it's soft

The section’s claim: physical labor’s protection was never mechanical difficulty — it was the absence of judgment — and that’s an AI problem now being chipped at. What holds, and what’s a pace bet:

Moravec’s paradox is why this looked safe, and the section splits it in two. It’s a well-known observation in robotics and AI: the things humans find hard (chess, arithmetic, formal logic) turned out easy to automate, while the things humans find trivial (perceiving a cluttered scene, walking over uneven ground, manipulating a soft object) turned out extraordinarily hard. The section’s claim is about which half of “manipulating a soft object” was the hard one. The dexterity to pluck a ripe raspberry without crushing it is a solved mechanical problem; deciding which berry, obscured how, at what angle, past which overripe neighbor, was the wall. That wall is perception and judgment — an AI problem, not a gripper problem.

The humanoid form is largely a distraction, and the section is right about it. A raspberry harvester needs vision, a specialized end-effector, and a way to move down a row — not legs, a torso, and a head. Purpose-built machines carrying a judgment layer are the realistic deployment path; the walking-android image is dramatic, attracts attention and capital, and misleads people into thinking “robots still can’t climb stairs, so we have time.”

Foundation models are the mechanism that changes the curve. Traditional industrial robots needed bespoke hand-coding per task — one joint, one chassis, one long integration. Recent robot-control models trained on broad data (the vision-language-action line of work) can generalize to instructions and objects they weren’t explicitly programmed for. That means a capability can propagate as a software update rather than a mechanical redesign, which is a genuinely different and faster scaling dynamic than classical robotics.

Exposure becomes displacement through price, not capability. The strongest economist’s objection is comparative advantage: even if machines are better at everything, their capacity is finite, so humans keep a role in whatever carries the lowest opportunity cost. That is correct and it guarantees trade, not a wage above subsistence. A task’s wage is capped by the cheapest alternative way to get it done; offshoring pushed that alternative down to the cheapest human, and compute pushes it below. Wassily Leontief made the horse comparison: horses and mules on US farms fell from almost 26 million around 1920 to under 3.4 million by 1958 (USDA), not because they lost every comparative advantage but because what remained was worth less than their upkeep. The checkable form of the claim is a ratio: the cost of a model doing a fixed task at quality parity, divided by the cheapest human rate for it, tracked over time.

The honest limit. “Faster curve” is defensible; “exponential” and any firm timeline are not. Physical-world autonomy has a long record of humbling confident forecasters — see the thinking-aloud note below. The section’s durable claim is directional: the bottleneck moved from the mechanical to the cognitive, and the cognitive side is where progress is quickest. How fast that reaches your specific trade is exactly what nobody can honestly promise.

— thinking aloud —

The thing I can least defend here is pace, and the cleanest way to admit it is to point at the graveyard of confident predictions I’m standing next to.

Self-driving cars were “a couple of years away” for well over a decade. Serious people, serious money, repeatedly wrong — not because the idea was foolish but because judgment in the open physical world turned out savagely harder than the demos suggested. Every edge case is a new edge case; reality does not run out of them. If I’m honest, that’s the single best reason to be skeptical of anything I say about physical automation timelines, and I’m not going to pretend it away.

So why do I still think the direction holds, even if the clock doesn’t? Two reasons. First, self-driving is close to the hardest possible case — open world, high speed, safety-critical, one mistake is fatal and public. A berry picker in a polytunnel, a machine sorting parcels, a rig cleaning a fixed floor plan overnight: these are bounded, slow, and forgiving in a way a car at 70mph on a wet motorway is not. The hard case running late doesn’t tell you the easy cases won’t arrive. Second, the bottleneck genuinely did move — the mechanical side was never the problem, and the cognitive side is where the fast progress is real.

But notice what I just did: I defended the direction and conceded the timeline, and a lot of the section’s emotional force comes from the timeline feeling near. Strip the urgency and you get “physical work loses its structural protection eventually, at a pace nobody can name.” That’s a weaker, truer sentence, and it’s the one I’d actually stand behind. If you want to fight me, fight me on whether “eventually” is close enough to matter — that’s the real argument, not whether it happens.

Objections 5 the strongest counters, answered
Robots still can't reliably fold laundry or handle a cluttered kitchen. This is hype.
Manipulation in clutter is genuinely hard — that's Moravec's paradox. The section's claim is about which part is hard: perceiving the scene and judging what to grab and how, not the actuator doing the grabbing. And the claim isn't 'robots are good at this now.' It's that the wall was always judgment and perception, and judgment is the thing improving fastest. How soon that clears your particular example is a pace question, and pace is the section's soft spot — see 'thinking aloud' under Why this holds. 'Not yet' isn't 'not coming.'
Self-driving cars were 'five years away' for fifteen years. Physical AI is perpetually overpromised.
This is the best objection, and it's correct as a caution. It's exactly why the section shouldn't be read as a timeline. Two things soften it without dismissing it: self-driving is close to the hardest case (open world, high speed, safety-critical), and many physical jobs are far more bounded and forgiving. The direction — bottleneck moved from mechanical to cognitive — survives the timeline skepticism. Bet against the clock, not the arrow.
The trades are protected by licensing, liability, insurance, and the need for a trusted human on site. Regulation is a moat.
Those are real frictions and they do slow deployment — genuinely, for years, especially where safety and liability are highest. But they slow rather than reverse, and they're strongest precisely where the stakes make automation cautious, not where they make it impossible. A moat that buys time is not the same as a moat that holds. The section's claim is about the structural protection, not next quarter.
Humanoid robots are an expensive gimmick. Boston Dynamics videos aren't an economy.
Agreed — and the section says exactly this. The humanoid form is mostly a distraction; the real path is purpose-built machines carrying the judgment layer. If you're dismissing the androids, you're dismissing the thing the section also dismisses, not the argument it's actually making.
Comparative advantage. Even if AI is better at everything, humans will still do whatever they're relatively best at. Economics 101 says there's always work.
Comparative advantage guarantees a role, not a living. It says trade still pays; it says nothing about the price. The wage for a task can't exceed the cost of the cheapest other way to get it done, and when that alternative is compute, the ceiling falls every year. Horses had a comparative advantage too; horses and mules on American farms still fell from about twenty-six million to under three and a half million in less than forty years, because their remaining work wasn't worth their feed. Humans aren't horses — they vote, own capital, and can be redistributed to — but those are political protections, not economic ones, and comparative advantage isn't among them.
Argue it yourself ↓ a scoped prompt for any AI — bring your own disagreement

Broken yet?

The irony is noted. You may continue either way. Disagreeing opens a short form that asks why; skip it if you like (privacy).