Part 2 of two. This essay assumes the argument of Part 1, “The Era of Deep-Waste,” which describes the financial structure whose technical premise is examined here. Please read it first; the verdict reached at the close of this essay is the answer to the question Part 1 deliberately left open.
Part 1 ended on a single unresolved premise. The charge of deep-waste, that scarce talent, compute, and capital are being gathered into closed, reputation-priced, unfalsifiable loops, holds only if the labs are wrong about the one thing that would justify the concentration: the claim of an imminent discontinuity, a phase change after which the slow, distributed work of discovery is folded shut and replaced by the systems being built. This essay names the belief that lives beneath that claim, follows it back to the waste, and then asks whether the discontinuity could arrive at all. It argues that it cannot, and that the reasons are not contingent but structural.
1. Defining Intelligence Authoritarianism
There is an old temptation in human affairs, and it returns whenever a society meets a problem larger than any one person can hold. It is the conviction that the slow, distributed, friction-laden machinery by which knowledge and order are actually made can be folded shut, that a sufficiently small number of sufficiently gifted minds can stand in for the whole patient process.
Plato gave the temptation its first clean form in the philosopher-kings, those who have seen the Forms and are set above the noisy deliberation of the many. Carlyle gave it its modern romance in the “great man” theory of history, the idea that the past is at bottom the biography of a few towering figures; and Tolstoy answered him across the second half of War and Peace, insisting that history is the slow sum of countless small actions no general ever commanded. The twentieth century translated it into bureaucracy: the technocratic faith of Saint-Simon, with the engineer as sovereign; Lenin’s vanguard, the few who carry the direction of history on behalf of the many; and, most vividly, the central planning of Gosplan, the belief that a room of brilliant planners holding enough data could allocate an economy more wisely than the dispersed, tacit, local knowledge of everyone living inside it.
Science keeps its own version of the myth, the lone genius who leaps ahead of her field, and it survives in spite of the evidence gathered against it. Robert Merton catalogued the “multiples,” calculus discovered twice, natural selection twice, the telephone in a near dead heat, and showed that a discovery tends to ripen in a field before any one person reaches it, so that the individual becomes the occasion of the finding rather than its sole cause. The true unit of progress is the field held over time, not the mind held apart from it. Even the genuine giants did their work inside the distributed weave and could not have done it anywhere else; the pivotal person is a richly connected node in a living network, never a replacement for the network.
Let us call the modern, technological form of this temptation Intelligence Authoritarianism, or IA: the belief that a few super-heroically intelligent agents, whether human founders or the systems they are building, can carry whole fields forward largely by themselves, so that the proper response to a hard problem is to gather intelligence into one place rather than to spread it. IA forgets the cacophonous, collaborative, world-coupled way that fields in fact evolve. And it has a precise inner structure. Where evidence is unavailable, as it must be for any claim about a capacity not yet shown, IA quietly seats authority, the reputation of the exceptional mind, in the chair that evidence should occupy.
The deepest form of the error is Friedrich Hayek’s knowledge problem, turned inside out and dressed as its opposite. Hayek argued that a market, and by extension any distributed field, is a way of gathering up the dispersed, tacit, locally held knowledge that no central mind can ever contain; the price system computes what no planner can. The cacophonous scientific field is exactly such a gathering mechanism for discovery. Intelligence Authoritarianism is the planner returned, now insisting that a small enough circle of brilliant minds, or one powerful enough model, can do the gathering’s work from a single vantage. The irony sits close to the surface: its most fervent backers are often market fundamentalists who would name the fallacy at once in Gosplan and miss it entirely in a twenty-person lab in Palo Alto. They are funding central planning for discovery.
2. From Belief to Waste
Intelligence Authoritarianism is the layer of belief; deep-waste is what it precipitates in the economy. The link between them is mechanical rather than rhetorical.
IA holds that intelligence is best concentrated. Concentration, laid over an unfalsifiable claim, has only one pricing mechanism left to it, reputation, as Part 1 traced, and reputation priced reflexively becomes a self-funding ecosystem. So IA leads, by the route already drawn, into the belief-based, self-reflexive valuation game. The waste, though, runs deeper than misallocated money, and that deeper layer is what earns the prefix.
The structure selects against the very ground it stands on. A reputation-collateralized, unfalsifiable bet can clear at a thirty-billion-dollar valuation, while falsifiable, publishable, incremental work, the actual substance of the distributed process, cannot raise on those terms, because the honesty about intermediate results that makes such work trustworthy is exactly what an unfalsifiable narrative leaves out. So the two genuinely scarce inputs in the field, frontier-caliber researchers and frontier compute, are drawn out of the shared process and parked inside closed nodes that return almost nothing to the commons: no papers, no released models, no students carried to independence, no methods others can repeat. Capital follows the same slope.
This is the precise sense in which the waste is deep. It is not money burned, which would at least be neutral; it is the substrate of distributed progress being quietly mined away to feed a model that denies the substrate exists. Intelligence Authoritarianism starves the very ecology that grows giants while staking everything on the arrival of one, and when a closed node fails, the talent it had gathered does not flow back into the open field but scatters to other closed nodes, so the public ledger never recovers what was taken. The belief and the waste are one thing seen at two depths: a conviction about where intelligence ought to live, and the slow draining of intelligence out of the only soil in which it has ever actually grown.
3. Why the Discontinuity Cannot Arrive
Everything turns on the discontinuity. If the field-paced human process really is about to be folded shut and replaced, then IA is right and the concentration is reasonable. So the claim has to be met on its own ground, and when it is, it does not hold. The reasons come in two families: physical limits on the rate and the very possibility of self-improvement, and epistemic limits on the meaning and the verification of the thing being promised.
A word on rigor before the arguments themselves. Two claims often offered in this debate are better set aside, because a careful reader can use a weak argument to cast doubt on the strong ones beside it. “The laws of physics forbid self-improving systems” is not true; physics bounds the rate and the ceiling of improvement without forbidding it, and evolution and human toolmaking are standing proofs that recursive improvement happens. And “vanishing gradients” names an engineering difficulty the field has already grown past, through rectified activations, residual connections, normalization, gating, and attention; to cite a solved training problem as a permanent cap on intelligence is to hand a skeptic an easy dismissal. What follows leaves both aside.
The binding limit is not intelligence. The strongest physical considerations are energy and the finite-resource economy of labor and consumption, and they are one consideration seen at two scales: intelligence is not the bottleneck. The intelligence-explosion story quietly assumes that cognition is the limiting ingredient and that everything else, the energy, the materials, the time an experiment takes, the capital, the institutions, the demand, will scale along behind it without friction. That assumption does not survive contact with the world. No one can run a wet-lab assay, a fabrication line, or a clinical trial faster than physical and regulatory reality allows, and a mind a thousand times quicker still waits on the centrifuge and the grid. Landauer’s principle sets a thermodynamic floor; the power grid and the supply chain set the ceiling that binds long before that floor comes near. This consideration need not even deny that the systems grow more capable. It denies only that capability is enough.
Self-generated improvement decays without an outside anchor. Model collapse is a real result: train a model recursively on its own output and the distribution degrades, losing its tails. It does not undo self-improvement in general; it closes one particular path, the model that lifts itself by training on what it has produced. Anchor instead to fresh real data, or to reinforcement against a verifiable environment, and the system finds its way around the collapse, which is only to say that the path always returns to some external, real-world signal. The collapse result is a careful statement of what happens when the loop is sealed.
“Superintelligence” is not a possession but a practice. Imre Lakatos drew the line between a progressive research programme, which predicts and then confirms genuinely new facts, and a degenerating one, which only refits what is already known. By that measure superintelligence is not a property a system holds; it is a programme a system carries out, and the only evidence for it is a continuing stream of novel, verified results. Without that stream, “superintelligence” is an attribution rather than an achievement, the same epistemic object as the reputation of Part 1, an unfalsifiable name doing the work that verified output should be doing.
Evolution is our one proof of open-ended improvement, and it is the very opposite of the neo-lab. Evolution is not self-modification. It is selection working at the level of populations, a statistical process over a fitness landscape to which the organism is subject rather than author; there is no improving agent within it, only optimization at the level of the substrate. Once the agency is set aside, look at what the single proof we possess of unbounded capability growth actually asks for: massive parallelism, coupling to a real environment as the oracle of selection rather than a simulation of one, the slowness of deep time, no foresight, and no self-certification at all. That is the photographic negative of the neo-lab picture, which is small, closed, fed on simulated signal, fast, foresight-driven, and self-improving. The one data point we hold quietly contradicts the takeoff architecture along every axis.
Verification is coupling, and a closed loop cannot pay for it. Here the physical and the epistemic become a single constraint. To improve in any open domain beyond the human level, a system needs a signal richer than human judgment can give. That signal comes either straight from the world, slowly, serially, bound by energy and kept on the world’s own clock, or from a model that simulates the world. But the simulating model is offered no free lunch: to give a superhuman signal it must already be more accurate than the system it is teaching, which means it must be checked against reality, and a model checked only against itself simply passes its own errors forward, which is model collapse wearing a second mask. The checking carries the physical bottleneck back in.
Press on the tempting escape, that the coupling bandwidth might after all be small and cheap, and it gives way. Nature is not uniformly irreducible; physics exists because great stretches of it compress, so that one can foretell an eclipse without simulating every atom. Yet a thin channel can carry only the compressions already discovered, and applying a known law is cheap and is not new capability. The discovery of new structure, which is what superintelligence is sold as doing, cannot be cheap, because one cannot know in advance which compression nature will honor. The field of candidate laws is unbounded, and only nature settles among them. To find the right one you have to ask the phenomenon, and asking it is a physical act of contact. So the bandwidth required is not small. It is exactly as wide as the novelty being drawn out, and novelty is by definition the part not yet compressed.
Which lets the whole matter rest on a single line: verification is coupling. To verify a new claim about the world simply is to touch the world. The Lakatosian demand that superintelligence be verifiable and progressive is therefore not a second condition standing beside the thermodynamic one; it is the coupling requirement said again in the language of confirmation. They were never two demands. And so the closed, self-improving loop is incoherent in the most economical way available: it claims to make verified, progressive novelty about the world from inside the loop, without paying for coupling, and yet verified novelty about the world is coupling, by definition. The claim is coupling without coupling, which is not so much unlikely as self-cancelling. A map cannot be cheaper than a territory it sits inside and must contain; a part cannot out-compute the whole that holds it.
One honest toehold remains for the other side. A system can genuinely speed the use of the coupling that humanity has already paid for, metabolizing centuries of accumulated, verified compressions faster than any single person could, and from inside one field this can even resemble a discontinuity. That is real, and it is valuable. But it is inheritance rather than self-improvement; it draws down a stock of coupling already bought, and it does not mint new coupling for free. The moment it spends the inheritance and reaches for the genuinely new, it rejoins the rest of us at the centrifuge and the telescope, on nature’s clock, inside the cacophony. The giants’ shoulders were never a courtesy. They are the only ladder, because every rung had to be checked against the world by someone willing to touch it.
The discontinuity, then, is not a hard problem awaiting enough compute. It is a category error in the shape of a hard problem, and the reflexive financial structure of Part 1 is secured against it.
4. If It Fails: Two Couplings, and the Road Back
Suppose Intelligence Authoritarianism fails in the near term, unable, as the previous section argues it must be, to reach its discontinuity before the conditions that hold it up give way. There are two ways the holding-up can fail, and they answer to two couplings the closed loop tried to avoid and cannot. Call them economic coupling and societal coupling.
Economic coupling. Sooner or later the economic reality is paid. A valuation built on future production is a loan drawn against that future, and one can borrow against the future only so far and only for so long. If value is not produced quickly enough, the cost of the borrowing rises, because the lender, which is finally the real economy, begins to doubt repayment. And there is a harder turn that the standard story misses. If the value the system does produce is itself corrosive to the society carrying it, eroding the commons, hollowing the institutions, displacing the trust on which exchange depends, then the cost of the borrowing does not merely rise with delay; it compounds, because the borrower is now drawing down the very capacity that was meant to repay the loan. The interest is charged in friction, and the friction grows more raucous: the withdrawal of social license, the arrival of regulation, the refusal of talent, the flight of patient capital. The early signs of this are legible now in outline. Follow-on raises that fail or reprice downward, of which the faltering of a reported fifty-billion-dollar round is plausibly a first tremor; talent dispersing from the closed nodes; secondary and tokenized markets marking down as the reflexive loop that swelled on rising valuations runs in reverse; a quiet pivot from the straight shot toward shipping a product, which is itself an admission that the unfalsifiable bet now needs a falsifiable revenue line to live; and, most diagnostic of all, gains that keep arriving in the cheap-oracle domains while conspicuously failing to cross into the expensive-oracle ones the system was funded to transform. The engine of the rise becomes the engine of the unwinding.
Societal coupling. Beneath the economic settlement lies a slower one, the long tug between concentration and diffusion that runs through every human society. The two are opposed, and yet, paradoxically, bound to one another. A highly diffuse state, the cacophony in its pure form, holds almost no collective power; it cannot act in concert, and that very powerlessness calls concentration into being, since power must coalesce somewhere if anything is to be done. But concentration carries the seed of its own weakening. As a cluster grows large it grows brittle, more homogeneous, more centrally arranged, and it loses the dispersed and tacit knowledge that made the diffuse state fertile in the first place. So neither pole holds still: pure diffusion has no power, and grown concentration turns fragile, and the society moves between them, each forever implying the other.
Intelligence Authoritarianism is an extreme wager on the concentrating pole, and by this logic the extremity is itself the danger, for it is precisely the brittle, homogenized condition that surrenders the generative cacophony it was meant to surpass. For a discontinuity to escape either failure, the economic reckoning or the societal brittleness, the system would have to solve the coupling problem rather than evade it: it would have to produce, or discover, designs that can actually sustain the society on which it is hoisted. That means far more than lifting our civilization to its next energy level, a fusion economy or the rung above it, though that may well be necessary. It means arriving at an economic and social value-system that holds a stable equilibrium, a way of balancing concentration and diffusion that collapses into neither impotence nor fragility. And this is the hardest part, because such a coupling would require the system to work against the very primitive traits of the humans it depends upon and must keep alive, the hoarding, the status-seeking, the tribalism, the hunger for the near reward, which are at once the substrate it runs on and the forces that pull every concentration toward brittleness. One cannot simply design away the ground one is standing on.