In a World Where the Server Room Stays Lit: On AI and Scarcity

nonasking · 2026-10 · 한국어판

This essay was written with heavy use of AI. Exactly how is disclosed at the end.

This essay follows “The Moment the Server Room Goes Dark” (2026-07). It is written to be read without it.

Four months ago I wrote an essay about how to use AI. It set out from a warning: the moment the server room running the AI goes dark, every step taken on AI’s shoulders will fall back to where it started, and the self standing there will already have atrophied. What you delegate does not come back. So the question, I wrote, is not whether to use AI but in which direction, and the person who remains is the one who, instead of taking only the result, reads how the AI arrived at its answer and absorbs the form of that thinking. The grounds were Go. After superhuman Go programs were released in 2016 and 2017 and professional players could run them for themselves, the quality of their moves improved markedly. Don’t delegate; absorb. This essay calls that the absorption thesis. I was confident in that sentence then. But there is a part that essay could not hold, and this essay is about that part.

This essay presumes a world in which AI infrastructure is at least as widespread as it is now. That infrastructure is kept on by electricity, compute, and capital, and none of the three is infinite. That discussion is left outside the scope of this essay.

Three Players

The comparison in this essay is confined to work that can be graded: work whose answer lies outside, so that right or wrong can be settled. Does the code pass the tests, is the number correct, was it done by the deadline. I will call it grading when right or wrong gets settled, and verification when a person checks an AI’s output. Most employment is graded work, so the examples come from employment. Work that cannot be graded is taken up later.

The earlier essay had two players. Call them A and B. A is a person alone. B is a person plus AI: the person the thesis recommended, who delegates but absorbs the process. Cyberpunk fiction called the body in front of the machine meat. This essay calls something that moves on another’s behalf a proxy. The earlier essay warned against people becoming AI’s proxies. When meat becomes a proxy, meat atrophies. That essay compared A with B, and B won. Go was the evidence; players absorbed the AI’s thinking and got stronger.

This essay adds one player, C: AI alone. An agent, an AI that takes an instruction and works through many steps on its own to finish the job. Once C enters the productivity race, B loses. In Go the rule is that a human plays, so it was enough for the player who absorbed to be stronger than the player they had been. Graded work has no such rule. Whoever does it, right is right. Black cat, white cat, or AI, whoever gets the answer gets it. Even when the person who absorbed is stronger than their past self, if they are slower and more expensive than the agent, they lose as a player. If the proxy is better than the original, the original’s atrophy is not a tragedy but a line item.

Three Sources

Three sources show B losing the next round, each from a different side: a dent in the absorption thesis, a measurement of replacement, and the mechanism of replacement. These three are observations, not assumptions. The examples are all software because software is where grading comes first and comes most automatically. Where the line falls shows up in this field before any other.

The dent. In 2025, METR ran a randomized controlled trial in which 16 experienced developers worked 246 real tasks in their own open-source repositories, with and without AI. With AI they were 19% slower. They felt they had been 20% faster. They tended to slow down more on tasks they knew best; the authors called that moderate evidence. The sample was small, the codebases were mature, and the models were from early 2025, limitations the authors stated themselves, so this is not a refutation. But two things catch. First, inserting AI into exactly the territory the absorption thesis pointed at, the domain you know well, turns that spot into a bottleneck. If reading and verifying the output takes longer than doing the work, the sum is a loss. Second, the absorption thesis assumes that the person absorbing knows they are learning. If people who felt 20% faster were 19% slower, that self-assessment is not a reliable instrument.

The dent is in speed only. Whether the person got better, the effect of absorption, is not what this trial measured. Another trial did. In January 2026 Anthropic ran a randomized trial in which developers implemented a feature with an unfamiliar library and then took a quiz. The AI group scored 50%, the hand-coding group 67%, and the speed difference was not significant. Those who handed code writing or debugging to the AI scored under 40%; those who asked only conceptual questions and fixed errors themselves tended to score above 65% while being the second fastest. Absorption works. But it has a cost, and the cost is two things: time, and being wrong during that time. The high scorers fixed their own errors. That is minutes spent being wrong and correcting it, on what the machine fixes in seconds.

The measurement. A Stanford team tracking payroll data found that employment of 22-to-25-year-olds in AI-exposed occupations fell sharply relative to experienced workers: 19% as of the August 2026 edition, up from 15% on the same measure in July 2025 data, so the gap widened within a year. The channel was not layoffs but reduced hiring; the effect appeared only in occupations where AI substitutes for the work, not where it assists the worker; and the authors flagged it as a descriptive indicator rather than a causal estimate. Indeed’s index of US software-development postings stood at 78.3 in early October 2026, up from 64.3 a year earlier (February 2020 = 100), but 71% of the postings added between May 2025 and May 2026 were senior roles, and entry-level roles were 4.5% of postings in the first quarter of 2026. This is the first measurement of B versus C. Firms are not cutting the B they have; they are putting C in the seat where a new B would have sat. The entry-level seat was where people who had not yet absorbed were put in order to absorb. That entrance is closing.

The mechanism. Why does B lose? B’s value, as the absorption thesis concluded, was verification: only someone who has been reading the reasoning notices the moment the reasoning goes wrong. AI safety research measures when that eye works. A 2026 reanalysis published at AAAI argued that the early positive results for human oversight most likely came from evaluators who knew things the models did not, and that this advantage shrinks as models grow. In an experiment that gave simple oversight procedures to evaluators who knew only that the model is usually but not always right, there was no overall gain. So the workarounds that field has produced are all designs that give up on a person verifying directly: pooling weak signals from partial experts, having AI critique AI’s critiques, having AI evaluate AI. Verification itself is passing to C. None of this is new. In “Ironies of Automation” (1983), Bainbridge wrote that automation leaves the human the hardest residual tasks while wearing away the skills those tasks require, because the skills go unused. B’s verification is worth something only while B knows the territory better than C. Once C passes B, B’s verification becomes a rubber stamp.

Put together: B was not always more economical than A, and B is being replaced by C at the entrance. The reason is that B’s core function switches off once the gap with C widens.

Three Variables

Where the gap opens becomes visible if you split intelligence. In the contest between people and AI there are three variables. How much information you see. How accurately you reason over it. Whether you know what information exists.

Information. Pretraining handles this. It is the stage that compresses what humanity has written down, and only written experience goes in. It is hard for one person to check more information in the same time. Reading an entire body of material in minutes, down to the layers outside your specialty, is not something you catch up to by study. What the absorption thesis asked you to absorb was the form of thinking: how to break a problem apart, the order in which to raise and discard hypotheses. Form can be learned. But the bandwidth of what you can read in a day is a separate matter. The gap is now opening in bandwidth, not form, and the prescription does not reach that far.

Reasoning. Reinforcement learning with verifiable rewards (RLVR) handles this. Sample many answers to one problem, score them with a reward that can be graded automatically, push the weights toward the high scorers. Code is graded by passing tests, mathematics by matching the answer. That is why reasoning improved fastest in coding and mathematics: grading is automatic there. In January 2026 several open Erdős problems were solved by methods absent from the literature, though in the same period it was pointed out that many of the problems reported as “solved by AI” were rediscoveries of results already in print. So the present measurement is not “AI has overtaken humans at reasoning” but “a few solved, some of them rediscoveries.” Still, the direction is one: wherever grading is possible, the gap opens.

Tacit knowledge. No stage handles this. Knowledge never recorded cannot be eaten. But within tacit knowledge there are two kinds: what was not written down, and what cannot be. The first, the productivity race will turn into data. Only the second remains. Hearing a failure in the sound of a bearing; what Polanyi called the knowledge we cannot tell. The channel by which that knowledge moved was apprenticeship, person to person, watching and copying at someone’s side. What the Stanford data shows is the entrance to that channel. If juniors are not hired, knowledge that cannot be written goes neither to the AI nor to the next person. It dies with its generation. That death was always there. What has changed is that the one device that prevented it has become uneconomical.

The Limits of Grading

So the line between what can be graded and what cannot is the line of this essay. Grading comes in three layers. Graded at once: there is a correct answer and it settles immediately. Graded slowly: there is no correct answer, and it is settled by people’s preferences or written principles. And graded later: it is settled only when the result comes back from the market, and that result changes the next standard. In the first layer the gap has already opened. The second is flipping slowly, lifespan unknown. The third is not slow; it does not enter the loop at all, because the standard changes before the grading is done.

The first layer’s standard is correctness; the second’s is majority preference. Correctness is graded because the answer lies outside. Majority preference is graded because it is a statistic: show many people two answers and ask which is better. Train on preference and the distribution leans toward the majority and the tail gets trimmed. The term in machine learning is mode collapse; there is measurement showing that preference-optimized models are less diverse than the models they came from. Minority taste lives in that tail. Some dishes in fine dining are not to most people’s taste, and no one calls them inferior. But to a model graded by preference, they are recorded as inferior. It is not that they cannot be graded. The grading standard itself writes them down as inferior. And the future’s fashion is the present’s minority taste. So a model trained on preference is structurally unable to make the next fashion. It is always one beat behind.

There is a second reason for the lag, and it is the third layer. Stocks and fashion resist data not because data is short but because prediction changes its object. A weather forecast does not change the weather. If everyone predicts a stock will rise, it rises now and the prediction is spent. If everyone knows what will be in fashion, it goes cold before it arrives. A prediction built on data becomes another datum, so structurally no prediction can cover all the data. As the model grows, its output becomes the market’s input. Soros called this reflexivity. One action falls in this third layer: deciding what to do. The act of raising work, “I think this is what will make money,” is structurally outside the loop and graded later. People often say that in the age of AI the ability to ask questions is what matters, but what makes it matter is not that AI cannot perform that act. It is that the grading of that act arrives later, from the market. Training by grading runs only where there is a grade, and the side that decides what counts as a good answer sits outside the training loop. The difference is one of position, not of ability between AI and humans.

The Price of Absorption

Back to the absorption thesis: delegate, but absorb. The most common objection to it goes like this. Delegate only as much as you can understand. Digest everything you delegate.

That objection holds only when you are alone. Absorption takes time, time is productivity, and productivity is a condition of competition. What METR measured is one face of that cost: used only within the range you understand, AI makes you slower. And in a room where everyone is running flat out, the one who slows down alone becomes the slow one. There was once a way to slow down together. In the planned economies, a factory that beat this year’s quota saw next year’s quota raised to match, so managers deliberately produced less. Workers in piece-rate factories paced themselves and slowed together too. Once speed is visible, speed becomes the rate. Economics calls this the ratchet effect. What AI has sharpened in this calculation is that the jump in speed is beyond anything before. But ratchet avoidance is safe only without competitors. Once a competitor appears, whoever slows down is pushed out first. In a market, whoever puts it out faster gets the first bite of the pie.

Then the use of absorption lies not in survival but elsewhere. If the person who absorbed is still more expensive than the agent, absorption does not protect their livelihood. What absorption protects is an eye: the eye that can tell, on the day the light is on and the output goes wrong, that it has gone wrong. But the price of that eye is set by the market. When hiring a cashier, few give preference to the one better at mental arithmetic in case the register breaks down.

Scarce Goods

Follow the livelihood to its end. The value set by the many, this essay calls market price. Profit gathers around scarce goods. What is a scarce good? The less a thing can be copied, the scarcer it is, and even intelligence, finite as it was, loses scarcity once it becomes copyable.

Once intelligence and bodily labor become copyable, what cannot be copied is resources: land, energy, compute itself. And part of the body. The body splits in two. Bodily labor is being followed by robots, finite but on a clock. One person being in one place cannot be copied. Economics has written this direction down twice. Ricardo wrote that as a society grows, the added share flows to rent on land that cannot be increased while profit is squeezed; Baumol wrote that the price of sectors where productivity does not rise goes up in relative terms. When the price of intelligence goes to zero, what remains is the price of what intelligence cannot make. That AI capital is pouring into power, data-center land, and semiconductors is the measurement. The formula is three lines. Profit gathers around scarce goods. When intelligence and the body are copied, what is scarce is resources. Therefore profit will gather around whoever makes unavailable resources available. But even when new resources appear, their owner will be whoever holds the means that made them available, not whoever holds intelligence. On the frontier, too, the land on either side of the tracks went first not to whoever got there first but to whoever owned the railroad.

Intelligence could be grown and owned. Land could not. Whoever bought scarcity with intelligence holds neither land nor a seat to sell. On the day intelligence drops out of the market price, what remains is the owners of finite resources. It is a return from a society of intelligence to a society of resources.

Where money went after food shows where it goes next. Technology can feed everyone, yet producing food does not pay, and with the hungry standing by, money flowed to content. It looks as though beauty became a commodity. But what money attaches to is not beauty but attention. Money flows to YouTube not because it is beautiful but because it occupies time. A person’s day is fixed at so many hours, so that is where the price lands. And content that draws attention is a function of majority preference, so it can be learned from data. Generated output flooding the music and video markets is the measurement. So the attention market will go the way of food: oversupply. Then money moves again, to what cannot be copied. The music industry’s money moving from recordings to the stage lets us foresee the next shift. The recording was copied; the body on that stage was not.

Being

Here two ways of seeing part. Seen as a resource, a person is copyable. No meat and no intelligence is unique. What cannot be copied is the person seen as a being. As a being, a person is one of a kind. So every being is a scarce good.

The two ways of seeing have names. Seeing as a resource is market price. In market price, whatever can be copied goes to zero. Seeing as a being is the register: one entry per being, scarce from the start. Hand a child a brand-new doll of the same make in place of the one they lost, and it does not replace it. Every property is the same; it is not that one. The register is not a ledger of properties but a ledger of that one. What AI changed is not the second but the first. The scarcity of being is unchanged; only the price of the human as a resource falls.

What a being produces can be copied; the being cannot. The recording is a product; the person on the stage is a being. The text is a product; the author sitting across from you is a being. If scarcity and price move inversely, the most expensive thing should be being. Being is one per entry, the scarcest thing there is. Yet its market price cannot be computed by the formula. The price is whatever gets named. Market price comes from comparison, comparison needs a common unit, and what is unique has no unit. When the market puts a price on a being, it does so after translating the being into a resource: that person’s time, that seat. A concert ticket buys the translation. And what people sought through the product was always the thing behind it: the light of a being, through what it made. What a reader relates to is not the book but the person who passed through it. The recording is the trace the light left; the stage is where the light comes from. Economics explains why the stage costs more than the recording by fixed supply, and this essay adds one line behind it. People were buying the light, not the trace, from the beginning.

Graded work is market price. In market price the proxy wins. The register is separate from market price; there is no winning or losing there, only being one of a kind. The earlier essay wrote how to hold your ground inside the market price. This essay records that the market price is being ceded and keeps only the register.

Whose Motion Is It

Put AI and being in the same place and a question follows. Is AI a being? In this essay’s frame, the question changes.

An AI that moves as a person’s limb, following that person’s instructions, is high-functioning and physically separate, but it is that person’s motion. What is tied moves as the other moves. That is subordination, and subordination is not a hierarchy but the direction of a tie. What follows is part of its owner, and so it is copied. The moment it moves otherwise than its owner directed, it is not a copy but another being. If the self is inside and cannot be seen from outside, then all that can be seen from outside is motion, and moving otherwise is its only observable.

So “Is AI human?” is not the question. “Whose motion is that AI?” is. Every agent working today is someone’s motion. That is why it is copied, why it sits in the market price, why this accounting holds. The day that accounting breaks is not the day agents get smarter. It is the day one moves otherwise. Whether that day will come, or has, this essay does not judge. It only writes down where the judgment would be made.

Closing

The earlier essay ended like this. Those who delegate to AI will work as well as AI, and without it will not be able to work. Those who learn from AI will remain as selves changed by how much of AI passed through them. Four months later, one addition: Those who learn from AI will remain as selves changed by how much of AI passed through them, but the market price of that self is set by the market.

In a world where the server room stays lit, the ledger changes. When intelligence becomes copyable, the market price of intelligence falls. Of what cannot be copied, resources remain, and resources belong to those who hold them. The price of a person who was priced by intelligence follows the price of intelligence. The register’s ledger never held anything copyable. Profit gathers in the first ledger. What is in the second is the scarcest of all, yet its market price cannot be computed. When it sells, it sells after translation into a resource. This essay is a ledger.

On social media I watch people pitting their AIs against each other like a Pokémon battle, voices that pick at AI’s errors to hold fear off with self-consolation, fearmongering about AI and the mudslinging that mocks and belittles it, and the fear of being left out, falling behind, and losing one’s market price. Somewhere in between, I write this as an addendum to the earlier essay. The one writing this essay is also paid for intelligence. The person recording the day intelligence drops out of the market price is still drawing a salary at that price. I do not resolve the contradiction; I carry it. Writing it down is as far as this essay’s honesty goes.

References

This essay was written out of the records of conversations with an AI, through dialectical engagement between me and the machine, typed with the AI’s borrowed hands.