The Easy Problem

The Plant on the Hill

#

Gil is the superintendent of a water treatment plant that serves a town of nine thousand in the Sandhills, and the plant is, in the way of such places, several decades standing in one building. The pumps went in when the plant did, in 1968, and two of the original four are still in service, rebuilt so many times that calling them the original pumps is partly a courtesy. The monitoring layer went in three years ago: a model that watches turbidity, dosing, and draw, and that has twice caught a filter problem overnight that the old alarm logic would have slept through. The model runs on a small box in the office and cost less than a valve. Nobody in town knows it exists.

On the shelf above Gil’s desk is the plant’s commissioning logbook, started the week the plant opened, a green canvas ledger with the first superintendent’s handwriting in it, hard-pencil and slanted. Every superintendent since has kept it going, which makes Gil the fifth hand in the book. He writes in it maybe twice a month, when something happens that the next man will need to know.

The proposal on his desk is the plant’s first physical one: replacing the chemical dosing skid with a new one, closed-loop, driven by the same family of models that already watches the water. It is a good skid. The demonstration was persuasive. And Gil, who adopted the monitoring layer in an afternoon on his own authority, has been sitting with the skid proposal for seven months, because the two decisions are not the same kind of decision, and he can name exactly where they part.

What Holding Means Here

#

The corpus has an altitude it calls the serving unit: the smallest body that can receive a capability and keep it, absorb it into procedure, answer for it, train the next person on it, and maintain it after the enthusiast leaves. Below that altitude, capability arriving is not the same event as capability landing, and the corpus’s first count of such bodies found many fewer of them than the org charts imply.

The count was made with the cognitive case in mind, and the cognitive case is the gentle one. The monitoring model asked almost nothing of Gil’s plant: one competent person, which the plant had, and the permission to change its own procedure, which a superintendent holds. If Gil left tomorrow, the model would idle until someone learned it again, and the water would be no worse than before it came. The floor was low, and the plant cleared it without noticing there was a floor.

The skid asks for a different plant. It needs a technician within reach who knows this skid and not skids in general, and the nearest one the vendor can name is a hundred and forty miles away. It needs a parts shelf, and a plan for the day the vendor’s regional distributor consolidates, as regional distributors do. It needs someone answerable, by name, when the dosing is wrong, because wrong dosing is a public health event with a docket number, and a docket number wants a name. And it needs all of that continuously, for the twenty or thirty years the skid will stand, across the retirements of everyone now in the building.

A physical deployment is not a purchase with a maintenance cost. It is a maintenance obligation with a purchase attached.

That inversion is what raises the floor. The serving-unit altitude in the cognitive regime is set by absorption: can the body take the capability into its procedure. In the physical regime it is set by custody: can the body keep a machine alive and answer for it for decades. Custody is the harder test by far, and the count of bodies that pass it is smaller than the count that pass absorption, in the same places and for the same reasons. The layers of deployment that get cheaper every year, the model, the compute, the connectivity, are exactly the layers Gil’s plant already has. The layers that hold him at seven months, the technician, the parts chain, the named answerable person with a successor, have no deflation curve at all, and in places like his they have been thinning for two generations while everything else got cheap.

So rural exclusion in the physical regime is not a version of the access problem, and no broadband map will find it. The town can afford the skid. What it may not be able to field is the standing institution the skid requires, and no capital budget has a line for an institution. You either inherited one or must manufacture one, and manufacturing one, the last time this country did it at scale, took the better part of a federal program.

The Spread That Is a Population

#

There is a movement inside this subject that the corpus has been slow to write, and it should be written carefully, because for the first time the melting asset is not a position or a franchise. It is a population’s wage.

The arbitrage series established the frame: much of what looks like a business is a position held across a gap, and AI is the technology that prices every gap at once. The largest gap in the world economy has never been informational. It is the difference in the price of an hour of physical work between one place and another, and on that spread stand not firms but whole regional economies, the workshop cities and the assembly corridors, tens of millions of people whose employability is the gap. Cognitive AI barely touched that spread, because the work is done with hands. Physical AI is aimed at it directly. A machine that stacks, sorts, sews, or welds at the same cost everywhere is, among the other things it is, an instrument that closes the labor spread, and when a spread closes, what the position was earning goes away.

Everything the corpus knows about melting assets says the honest questions are the rate and the holder. The rate, here, is the merciful part of the finding, and it comes from this essay’s own regime: the spread closes at the speed of capital, not the speed of capability. Every task the machine takes, it takes by being bought, installed, held by a serving unit, and run to the end of its depreciation, which means the closing runs on the slowest clock in the ensemble and arrives task by task over decades, not model release by model release. That is real time, formation-scale time, the kind a country can do something with.

The holder is the unmerciful part. A firm holding a melting position can read its own ledger, and the arbitrage series has watched firms do it well and badly. A population holds its position without holding it anywhere: no ledger, no single decision-maker, no sell order available. The wage spread melts out from under people who never knew they were holding a position at all, and the proceeds accrue, as the second essay of this set found, to whichever account the closing is billed to. What a population can do that a firm cannot is act through the one body that answers for it, which returns this movement, by a road nobody planned, to the same place the skid decision lives: the serving institutions, the formation systems, the floor. The corpus’s account of what states still hold was written about a different question. It turns out to be the answer to this one too.

I wonder, and the wondering is uncomfortable, whether anyone responsible for such a population is reading its position at all, or whether the first full reading will be done afterward, by people like the corpus’s own archaeologists, working from what the record kept.

The Swap

#

Now the turn, and it has been waiting underneath both halves of this set.

The corpus’s account of the cognitive regime is, at bottom, pessimistic about a cheap thing: deployment costs almost nothing and stalls anyway, for years, on identity, on the professions, on the objection that changes shape, because the thing in the way is people, and nothing about people is cheap to change. Its account of the physical regime now has to be, at bottom, optimistic about an expensive thing. The cliff is real, the floor is high, the obligations run for decades, and none of that will stop it, because every difficulty in the physical ledger is a difficulty institutions were built to eat. Depreciation has a schedule. A maintenance obligation has an actuarial table. A parts chain has a contract, a safety case has an engineer’s stamp, a capital program has a lender who has financed ten thousand of them. The machinery of finance and administration has digested exactly this kind of problem since the railroads, and it does not need anyone to have an identity crisis first. It needs the numbers to clear, and one replacement cycle at a time, the numbers will clear.

The physical transition is slower and surer; the cognitive one is faster and stuck. The hard problem is us, and we were never the expensive part.

That inversion is checkable, and the check enters the register as this set’s second dated claim. Through September 2033, physical AI adoption inside existing facilities will track replacement cycles, not capability releases: a new physical capability, however striking, will produce no visible discontinuity in adoption, because the binding constraint is the capital and custody structure this essay has described, and that structure does not read announcements. If instead some release produces a step change in working facilities ahead of their replacement timing, then the cliff argument is weaker than three essays have made it, and the miss gets recorded where the claim is recorded.

Gil approves the skid in October, contingent on one clause he writes himself: the vendor trains two people, not one, and the training repeats at year five, on the vendor’s calendar, not the plant’s memory. It is the clause of a man who has read a certain green logbook all the way through. The first superintendent’s entries are mostly numbers, flows and doses in a hard slanted hand, but every few pages there is a sentence meant for whoever came next: where the intake fouls in a wet spring, which valve lies about being closed. Fifty-eight years of the plant knowing itself, five hands, one book. The skid will outlast Gil, the way the pumps outlasted the man who commissioned them. What he is deciding, this October, is whether the book will have a sixth hand.

References:

Institutions, Custody, and Infrastructure:

Hughes, Thomas P. Networks of Power: Electrification in Western Society, 1880-1930. Johns Hopkins University Press, 1983.

Brown, D. Clayton. Electricity for Rural America: The Fight for the REA. Greenwood Press, 1980.

Ostrom, Elinor. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990.

Labor, Trade, and the Spread:

Autor, David H., David Dorn, and Gordon H. Hanson. “The China Syndrome: Local Labor Market Effects of Import Competition in the United States.” American Economic Review, vol. 103, no. 6, 2013, pp. 2121-2168.

Rodrik, Dani. “Premature Deindustrialization.” Journal of Economic Growth, vol. 21, no. 1, 2016, pp. 1-33.

Baldwin, Richard. The Globotics Upheaval: Globalization, Robotics, and the Future of Work. Oxford University Press, 2019.

Series Anchors:

The Approximate Mind, TAM-LEV.05 (The Serving Units) and TAM-LEV.07 (The Floor): the altitude this essay raises, and the bottom the void requires.

The Approximate Mind, TAM-CMN.02 (The Tap and the Bottle): the rural cooperative, the one documented program that manufactured serving units on purpose.

The Approximate Mind, TAM-ARB.13 (Melting Ice) and TAM-ARB.14 (Owning the Instrument): the melting asset and its holder, the frame this essay extends to a population.

The Approximate Mind, TAM-095 (The Many Clocks): depreciation time and formation time, the two clocks this essay’s mercy and its worry both run on.