The Cyborg Era: What AI means for jobs

"You cannot stack a thousand horses to build a super-horse, but you can stack a thousand humans to build a corporation"

The horse was replaced and never came back. That image is the whole pessimistic argument, and Krier's reply is that it smuggles in an assumption: it treats labor as power output. Horses supplied traction, and when engines supplied it cheaper they had nothing else to offer. Humans are not one input — they organize, and the organizing compounds.

The distinction he draws next carries the rest of the essay. Substitutability is technical: can this input be replaced? Substitution is economic: does the replacement actually happen, given every cost and preference in play? "AI can do anything a human can, cheaply" establishes the first; most arguments treat it as establishing the second.

"As long as the combination of Human + AGI yields even a marginal gain over AGI alone, the human retains a comparative advantage."

Notice how weak the condition is: marginal gain, not parity. That is the bar the pessimistic case must clear — not "AI is better than humans," but "humans plus AI is worse than AI alone," a strange thing to have to demonstrate.

His supporting picture is progress arriving jagged rather than all at once: uneven across capabilities, bottlenecked in the physical world, needing products fitted to local law, and organizations that know what to do with them. Meanwhile humans step up a level of abstraction, becoming "orchestrators of intelligence" the way engineers became architects. He puts a number on it: 5–10 years, not forever.

"I would stop going to the Blythe Hill Tavern if there were robots behind the bar, and they're highly efficient with the Guinness."

The line, from a colleague, does analytic work rather than charm. The standard dismissal treats preference for human involvement as sentimentality markets will erode. Krier's counter: if what the customer wants is the human, the inefficiency is not a cost to optimize away — it is the product. Forty hours of a person's work is a different good from four seconds of a robot's.

The same logic covers positional goods and provenance: when everything copyable is free, demand moves to what stays scarce — the original, not the flawless reproduction. He does not oversell it, noting that self-checkouts are people trading human contact for convenience, and he flags the objection he cannot settle: whether such a market could employ most people, or stays a luxury niche.

"A tiny island in the middle of the ocean with 10 people on it can have the most free market economy ever"

Here the essay turns; read this part twice. A corner solution sits at the edge of the range rather than somewhere sensible in the middle: not a little trade, but zero. It happens when fixed costs — what you pay to set an arrangement up at all, whatever the volume — swallow the gains. The islanders may hold a real comparative advantage; Merck still will not trade with them.

In labor terms, the comfortable half of the essay now develops a hole. If the fixed cost of routing work through a person exceeds what the person adds, no firm needs a perfect substitute to drop them; it redesigns the process so the person was never in it. Employment fails not because AI does the job perfectly, but because the friction of involving a human costs "more than the value humans add."

"their 'low opportunity cost' is irrelevant because they are logistically stranded outside the high-speed value chain"

This is the mechanism behind corner solutions once systems get fast. To use a human, an AI must export the context, wait for a biological brain, and import the result; every hand-off costs time and translation. Past some speed, a system prefers a worse-but-native answer over a superior one that has to leave the machine.

It is the essay's most uncomfortable idea because it does not require AI to be better at anything: cheap labor does not help if connecting it costs too much. Krier names the counter-pressure — AI could get good at reducing the messiness of working with humans — but calls that his assumption, not a finding.

"the error bars and confidence intervals are enormous"

The final movement targets the pessimists' confidence more than their conclusion. Reaching zero human labor needs a stack of conditions holding at once: abundant compute, no physical bottlenecks, no residual demand for human things, fixed costs biting everywhere. Each carries its own uncertainty, and multiplying uncertain claims widens the interval fast. His complaint is that the conclusion arrives with a policy menu attached, as though "AGI within ten years" mechanically implied the rest.

He lands not on "humans will be fine" but on a research question: watch which force wins, comparative advantage or fixed costs, and do not perform policy surgery before the evidence arrives.