Unit Cost Dominance
The economic condition where an AI, with minimal oversight, produces a unit of cognitive work at a far lower marginal cost than a human. The short version is below. The full argument is seven essays away.
It is not about training cost
This is not the cost to train the AI, but the cost to run it for a task. Just as you do not care about Microsoft’s R&D budget for Excel, only your cost to make a spreadsheet.
An AI drafts a 20-page contract in 30 seconds. A senior lawyer, the verifier, spends 10 minutes reviewing it.
The creation work is replaced, for a fraction of its former cost. One elite verifier stands in for a roomful of juniors. This is replacement.
A plumber uses an AI tool to find a leak in seconds. The plumber still does the physical work of fixing the pipe.
The human is a little faster, not removed. The core act stays human. This is augmentation.
Premise 1 · The proof of dominance
To refute the thesis, falsify one of these
The argument is built to be attacked. Break any single condition and the chain does not close. None has broken yet.
Verification Scaling
Prove verification costs scale linearly with AI output, erasing the cost advantage.
Quality Gaps
Identify economically vital tasks where AI quality permanently lags human performance.
Voluntary High-Cost Adoption
Show that markets will systematically choose higher-cost human labour over cheaper AI.
Coordination Mechanisms
Demonstrate a binding, enforceable global mechanism preventing competitive adoption.
The Physical Refuge
Prove physical jobs can absorb tens of millions of displaced knowledge workers at a living wage.
Premise 1 · The math
The cheaper unit wins
Once the AI-plus-verifier produces acceptable quality, its marginal cost sits far below a human’s. In a competitive market the cheaper unit wins. Not by ideology, by arithmetic.
Illustrative of the mechanism, not a measured figure. The actual capability and deployment record (benchmarks, cost trajectories, real cases) is gathered in Appendix III and updated continuously.
Bits. AI performs nearly all of the creation; the human becomes a verifier. One expert stands in for many. Mass displacement.
High-stakes. A surgeon still cuts, a judge still rules. AI removes friction and surfaces precedent. Experts become super-experts.
Atoms. The plumber still fixes the pipe. AI is a diagnostic tool. Incremental efficiency, not replacement.
The standard objection
“But we have always adapted before”
Every displacement argument meets the same reply: past technology destroyed jobs and created more, so it will again. That is an induction. It has held every time, therefore it holds next time.
Bertrand Russell’s turkey reasoned exactly this way. Fed every morning, it grew more confident by the day that mornings mean food. Its evidence was spotless and its record unbroken, right up to the morning before Thanksgiving. The induction did not fail because the turkey was careless. It failed because the structure changed, and the past held no trace of the change coming.
“Jobs always came back” is true of every prior wave and tells you nothing about a wave with a different structure. Earlier automation took muscle and pushed people up into cognitive work; there was always a higher rung to climb to. Unit cost dominance automates the rung itself. The comfort of history runs out at precisely the discontinuity it cannot see.
Premise 2 · Interface collapse
Task wins become workflow wins
Dominance does not stay at the level of a single task. Once the model wins the tasks, the workflow is recomposed around it and the human is no longer absorbed back in. There is no alarm. The work simply stops being given to people. That is the bridge from a cheaper task to a vanished job.
Premise 3 · The trap
Coordination is impossible
Not one prisoner’s dilemma. Billions of them, played daily, by workers, firms, sectors and states.
The Sorites paradox · a gradient into obsolescence
At what point does helpful assistance become total replacement? Like removing single grains from a heap, no step feels decisive, yet the endpoint is absolute. You cannot regulate what you cannot define.
Premise 4 · The collapse, and no way back
The wage-demand circuit breaks
Production without wages means products without customers. This is not a recession. It is a structural break.
Universal Basic Income does not solve this. It formalises it. You are not liberated, you are managed. Without a productive role, the majority become, in the thesis’s blunt phrase, well-kept animals in a comfortable enclosure.
The verification divide
It does not democratise. It stratifies.
Value concentrates at the apex of judgment. And the pyramid is temporary: the machine learns from its auditor, eventually absorbing the verifier too.
Cognitive Obsolescence Syndrome
The anxiety is not a disorder. It is accurate perception.
Society victim-blames and says “just retrain”. Retraining against an exponential curve is futile.
I was told to get a degree. Now my degree is worth less than a £15-a-month AI subscription.
My entire career path was just a temporary lag between two versions of an algorithm.
I played by the rules, but the game ended before I could cash in my chips.
Why can't they just work hard like we did? The system worked for us.
The scapegoat cycle
The political system cannot address an invisible force, so it invents a visible enemy. Justified anxiety is harvested and redirected.
The knowledge trap
Knowing you are in a trap does not hand you the tools to escape it. Our political tools were built for the 20th-century fight of Labour versus Capital. Player 3, the machine, has entered the game, and it answers only to unit cost.
The dissolution of an era
Four pillars of postwar capitalism, removed
End of productive participation
The system was defined by mass employment, not just mass consumption. UBI and dividends are system replacement, not survival: welfare for a population that no longer has an economic function.
End of democratic agency
Political power for the majority was a downstream effect of economic leverage. When most people depend on transfers from a tiny AI-owning elite, their votes become performative.
End of upward mobility
The cognitive ladder is gone. Earlier revolutions automated muscle and pushed humans into mind-work. AI automates mind-work. There is no next rung for the majority to climb.
End of competitive wages
The market for mass human cognitive labour ceases to exist. AI's marginal cost becomes the new wage floor, a floor no human can live on. Wage-setting, the engine of the labour market, dies.
The machine’s concession
The constraint wearing four costumes
Assume the mechanism holds. Something has to replace the wage-demand circuit: not redistribution as a slogan, but a real fiscal architecture that keeps a polity fed, stable and legitimate without mass wage labour. An AI model was asked to design one, several times. It failed every time, and the failures had a shape.
Four constraints, each killing whole design families
The Sorites line
You cannot tax a line nobody can draw. Is spellcheck replacement? A drafted email? A contract another model wrote? Every use-tax dies on definition, before enforcement is even discussed.
The mobile base
Tax mobile profit or income and capital relocates; the base evaporates. The global treaty that would stop it needs the same undrawable line, and could not be enforced if written.
The thinning rent
One design survives the filters: tax what cannot move or be redefined (land, energy, water, compute footprint) and trade permits for citizen equity. Cheap energy and jurisdiction-shopped compute slowly drain even that.
The steering node
Every patch routes through one question: who controls systems that outgrow the need to obey. A constitution cannot bind a party it can no longer enforce against. That is not an economic constraint.
These were never four constraints. They are one, the control problem, wearing four costumes. No fiscal instrument reaches it. This is where the economics ends and alignment begins.
Why the order of operations is everything
Solved alignment is amber. It fossilises whatever ownership exists at the moment steering becomes reliable. Aligned superhuman systems do not die, do not err against their owners, and would suppress rebellion competently. So the next twenty years do not design the final system. They decide what gets locked into it.
Five no-regret moves, while the window is open
Stop drawing the line
Legislate around cognition, not about it. Anything premised on separating assistance from replacement fails on definition before it reaches enforcement.
Move the base while it is boring
Shift off payroll and income, onto what cannot move or be redefined: land, energy, water, compute footprint. The political price only rises with delay.
Trade permits for equity, not promises
AI capital still needs power, water, land and connection rights from territory. Spend that leverage on permanent citizen equity, entrenched as property. Cheap to grant now, costly to claw back later.
Watch the right dials
Headline unemployment lags by design. Track collapsing junior-to-senior ratios, graduate non-absorption, and the wage-financed share of consumption. Build the triggers on those.
Know what the window is for
You are not designing the final system; no one can. You are setting the initial ownership conditions of whatever steering later locks in. A narrower job, and a more urgent one.
Every one is defensible under perfectly ordinary economics: land value taxation, sovereign wealth funds, equity for infrastructure. If the thesis is wrong, you have modernised your fiscal base and built a sovereign fund. If it is right, you have made the only preparation that was ever possible, inside the only window in which it was possible.
The Oracle
Ask the Oracle of Obsolescence
A forensic read on the argument, on your profession, or on your own situation. Ask anything, or leave it blank and let it deliver its verdict.
Runs on MiniMax through a server-side proxy. No key is exposed in this page.
Don’t trust us. Ask another model.
Red-team the argument
The coordination problem is a prisoner’s dilemma with extreme payoffs; the Sorites paradox means the game can barely be defined. Hand the whole thing to a model on pure-logic mode and see if it survives.
“CGP Grey was the prophet. The Discontinuity Thesis is the autopsy. Horses were not exterminated. They were simply no longer worth feeding.”
The comprehensive version
The TL;DR above is the shape of it. The full case is seven essays, each with its own falsifiability spine, plus three appendices.
Unit Cost Dominance
Interface Collapse
The Multiplayer Prisoner’s Dilemma
The Sorites Collapse Principle
Categorical Recursion
The Successor System
Drag Is Not Rescue
Listen & watch
