Wisdom Engine

Source: blog + substack (v3 unified) — 42,162 records total.

> The city lights blurred outside the taxi window, a muted current against the deeper, more unsettling hum of Marc’s internal landscape. Justine sat beside him, still as statuary, her gaze fixed somewhere beyond the glass. The silence between them was not empty. It had density. It had shape. It pressed against their chests like unseen hands, woven from the weight of all that had been seen and could not be unseen.Marc remembered her dreams now—the ones she had never spoken aloud. The locked doors. The empty cribs. He felt them with a strange, pulsing ache in the centre of his chest, like phantom pain from an amputation that had never happened to him, but now belonged to both. Her private terrors, once her own, now stirred within him like sediment in disturbed water. And with that ache came a new guilt—not his own, but hers, bleeding into his system like a transfusion of shame.Justine tasted coffee. But not her own—bitter, clean, bracing. This was burnt. Metallic. The taste of his childhood kitchen. She felt it with nausea, remembered fear that was not her memory, yet lingered in the back of her throat like spoiled air. She could feel the shape of his father’s silence, the disciplinary chill that had coated his formative years. It wasn’t her trauma, but it marked her now. The boundary of their separate selves had ruptured.The merge had promised clarity. What it had delivered was a cracked lens—too clear, too harsh. No romance in shared understanding, only dissection. A landscape of jagged edges. A brutal map of betrayal, so subtle, so mundane, that it could only have been etched across a lifetime. In the clinical room, Justine had almost laughed. The urge had surged in her chest—wild, hysterical, evolutionary. A cry for absurdity. How had they believed that technology could mend what years of studied repression had left in ruin?She reached for her bag. Her fingers brushed his. A flicker. And in that flicker, a cascade: she felt the resistance in his shoulder, the ghost of a step backward that he didn’t take. It was his shame. Not fresh. Not recent. But old and weary, stained with the effort of hiding. A shame that no longer needed confession—because she had felt its source. She had become its archive.He sighed. But before the air moved through his throat, she felt it. It was internal. The collapse of hope reshaped into survival. It was her feeling now too. She could no longer separate his weariness from her own. He wanted to leave. Not her. Just everything. And she knew this, not by inference, but by the dull heaviness in the pit of her own stomach—his despair rewired into her anatomy.He remembered her laughter. How she had trained herself to keep it soft so as not to seem cruel. She remembered his suspicion of praise, the way it clung too closely to fear. They were full of each other now—complicated, corrupted, irrevocably merged.The car pulled up to the house. The porch light was off. The curtains drawn. From the outside, they looked like any other couple returning from an ordinary appointment. But inside—deep inside—the coordinates had shifted. What had once been distinct was now fused, noisily and without symmetry.They stepped out into the quiet, moving not together, but in approximation. A lag between their rhythms. Ghosts of reflex. Echoes of thought. They would enter their home. They would go through the motions. But beneath every movement, every wordless glance, the signal remained—a persistent, whispering static of entangled cognition.They were not destroyed.They were layered.And every thought thereafter would be held not in solitude, but in resonance.

> The interface was not designed for containment. It was designed for passage—fluid, reciprocal, bounded by calibration. But calibration had failed. Emotional load exceeded threshold. Cognitive partitioning collapsed. The system began to burn.Not physically. Not wires or sparks. But saturation.Affective resonance peaked. The boundary between Marc and Justine, already frayed, dissolved completely. They weren’t linked now—they were overlaid. Not communication. Superimposition. Two minds occupying the same affective bandwidth, broadcasting fullvolume, unfiltered.The system’s fail-safes triggered in sequence. Pulse lag. Neural echo delay. Temperature spike. The algorithm initiated a soft exit, but exit required separation, and there was no separation left to initiate.Marc blinked, but it wasn’t his eye.Justine tried to speak, but the thought passed through his vocal centre and stalled. Feedback loop. Mutual paralysis.Emotion no longer belonged to anyone. Pain. Regret. Guilt. Yearning. They spun, a gyroscope of unresolved affect, so fast they lost shape. He couldn’t remember if he’d loved her first, or if he was just remembering her memory of loving him. She couldn’t recall if the child had looked more like her or more like the version of him she had mourned before he was gone.They felt the same memory from opposing angles, vibrating against each other in contradiction.The system attempted shutdown.Too late.A flash: Justine’s mother’s hands. A memory Marc never met. A sound: Marc’s first fistfight at age nine. She wasn’t there. Now she was.Cognition blurred into sensation. Time unraveled. Syntax disintegrated. Meaning degraded.In the observation room, the technician stared at the vitals—sharp spikes, echo collapse, data noise. The AI’s internal voice labeled the session: “UNSTABLE—MERGE IRREVERSIBLE.”Inside, Marc whispered a name and Justine heard it in her own mouth.She tried to remember her last thought before they connected.But there were no separate thoughts anymore.Only combustion.System burn.

> He was too young. That was her first thought.Mid-thirties, maybe. Hair neat but unimaginative. His badge was laminated and perfectly aligned on his chest. His eyes never quite met theirs, scanning instead the walls, the equipment, the time. He had the posture of someone who had learned bedside manner from modules. He called them by first names with the performative warmth of someone taught not to overstep.“Justine. Marc. I’ll be running your interface today.” He said it like he was checking a boiler.Marc nodded. Justine stared. The technician didn’t falter.“You’ll be fitted with independent cortical mesh units, L-model series. One for each of you. They don’t store data. They don’t transmit. No external systems. Closed loop. Your experiences won’t be recorded—only translated, momentarily and only inward.”He smiled again. It was practised. Not grotesque. Just off by a few degrees of humanity.Justine tilted her head. “Have you ever done it?” The smile paused. Returned.“Technicians are not eligible for synchronisation. Interface familiarity risks entanglement bias. We stay outside the mesh.”He gestured toward two chairs separated by half a metre and a dividing console.“You’ll sit here and here. The link is initiated by simultaneous biometric readiness and consent signal. That means if either of you wants to stop at any point, the system will disengage.”Marc blinked. “Do people ever try to fake it?”The smile again. Same symmetry. Same nothing.“You’d be surprised how few actually want the truth.”The technician turned and keyed in the session sequence. On-screen, their profiles appeared side by side—bare bones clinical: age, weight, emotional disposition index, neural plasticity estimate. No names. Just numbers and shapes and plots.“Breathe normally,” he said, sliding gloves on with bureaucratic ease. “This will only feel like everything.”

Assemble the mechanisms. An ageing population is, by the compositional and biographical-availability logic, less capable of disruptive collective action, so the demographic fuel of the mass mobilisation that historically forced elites to concede is thinning. An ageing electorate is, by the stakes-and-turnout mechanism and independent of any change in opinion, more oriented toward the defence of accumulated assets and the existing order, so the pivotal voter’s interest increasingly runs with the grain of concentration rather than against it. Ageing lowers g and, through the r − g channel, tends to entrench inherited and accumulated wealth — the housing wealth of the old above all — deepening the specific hereditary form of inequality that open societies exist to prevent, even as a serious contrary thesis holds that labour scarcity may push the other way. And immigration, the one remedy for the economics of ageing, generates a cultural backlash concentrated in the ageing native population — a backlash that populist entrepreneurs can and do direct toward cultural and national targets while leaving the distribution of wealth intact, so that the discontent of the demographically anxious is channelled away from concentration rather than against it. Each of these lowers, in its own way, the resistance that a drift toward plutocracy would otherwise meet. This is what it means to call demography the current beneath the politics: it does not command the outcome, but it sets the odds, and it is setting them against the countervailing power on which the earlier essays’ hopes rested.

It was precisely this care that Aaron Clauset, Cosma Shalizi, and Mark Newman brought to the question in their 2009 paper in SIAM Review, the most important corrective in the modern literature and one that ought to be read by everyone tempted to wave a log-log plot. Their argument was twofold and devastating. First, the method by which power laws are usually “found” — fitting a straight line to log-log data by least squares — is statistically unsound: it produces biased estimates of the exponent and, worse, gives no indication whatever of whether the data obey a power law at all, since it never asks the question. It will return a slope for data that are not remotely power-law distributed, and the slope will look just as authoritative as a real one. Second, when one does the work properly — estimating the exponent by maximum likelihood, measuring goodness of fit with a proper statistic, and, crucially, testing the power law against serious alternatives such as the log-normal and the stretched exponential — a great many of the power laws confidently claimed in the scientific literature simply evaporate. For dataset after dataset, the data are fit at least as well, and often better, by an alternative that no one had bothered to rule out. The lesson is not that power laws do not exist. It is that almost no one who claims one has earned the claim, because almost no one has done the test that could have refuted it.

Cohort associations are the most visible body of evidence because large birth cohorts can link early dietary reports to later cognitive testing. These studies do not prove causation, but they do establish that diet quality in early life is not cognitively inert. In the Avon Longitudinal Study of Parents and Children, Northstone and colleagues identified dietary patterns at multiple early ages and related these to IQ at 8.5 years measured using the Wechsler Intelligence Scale for Children. After adjustment for confounders, an early “processed” pattern at age 3 years was negatively associated with later IQ, and the effect sizes were small but measurable in population terms (Northstone et al., 2012). Smithers and colleagues extended the same general approach by examining dietary trajectories in infancy and toddlerhood and relating them to IQ at 8 and 15 years; again, fully adjusted associations were weak but present for some trajectories, with the key methodological point being that early patterning can carry detectable association signals into later cognitive outcomes (Smithers et al., 2013). A related analysis in The Journal of Nutrition reported that infant diet quality—framed as adherence to complementary feeding guidance—was associated with later cognitive outcomes at 7–8 years in UK data, again with small effects that survive some adjustment (Golley et al., 2013).

And corrigibility — the capacity to be corrected — is the precondition, because it is the only mark of real rightness available to you from the inside. You cannot perceive the truth-relation directly. But you can do the one thing that stands in for it: you can keep the channel open, expose your beliefs to the things that could refute them, seek the test you might fail, listen to the person who disagrees not as an enemy of your certainty but as your only possible instrument for detecting an error you are constitutionally blind to. The man who is “so sure” that he will not be moved has not achieved a higher grade of rightness. He has amputated the only organ that can detect wrongness. His immovable certainty feels, to him, like the strength of his position; it is in fact the removal of the one thing that could ever tell him his position was weak. Being right, insofar as you can pursue it at all, is not a state you occupy but a relationship you maintain — a standing exposure to correction, a willingness to be shown, a refusal to let the feeling of certainty close the door that the truth might yet need to come through.

What makes this so difficult to detect — and far more difficult to confess — is that motivated reasoning is not the absence of logic. It is the abuse of it. The reasoner is not being stupid; very often he is being brilliant. The machinery of inference is fully intact, every gear meshing, every step formally respectable. Only the direction is reversed: instead of running forward from evidence to conclusion, the apparatus is set running backward from conclusion to a tailored selection of evidence, and then the tailored selection is presented as though it had come first. This is why intelligence is no protection against the disease, and frequently makes it worse. A clever man is simply a better advocate for whatever he already wished to believe. He builds smoother roads to the same false country. The capacity for reasoning that ought to be his defence against error becomes, in the grip of a motive, the engine of his self-deception, because he can always think of one more reason, one more distinction, one more way of reading the data, and he never runs out of them before he runs out of patience.

The US–Iran negotiations are compellence wearing the mask of diplomacy: a blockade maintained “in full force” while the talks proceed, an ultimatum of “unconditional surrender,” a counterpart who announces that he regards the negotiation as part of the war. Reassurance is structurally impossible here for the same reasons it was impossible for Antiochus. The commitment problem bites: even a sincere Iranian promise on enrichment cannot bind a future Tehran, and Washington, suspecting that today’s reassurance is merely tomorrow’s tactical pause, has no reason to trust it — while Tehran, watching the blockade stay in place regardless, has no reason to believe that compliance will be rewarded. And the information environment is poisoned in every direction: each Chinese move (defying sanctions, supplying dual-use parts) is read in Washington as proof of the axis; each American move (the blockade, the Taiwan arms, the sanctions on Chinese refineries) is read in Beijing as encirclement; each Iranian move confirms the worst to everyone. This is Jervis’s world and it is also Antiochus’s: once the frame is hostile, the reading machinery serves the frame, and more channels only multiply the signals available to be misread.

Mises held that the state is the social apparatus of compulsion and coercion, that it holds the monopoly of violent action, and that it is indispensable — not a necessary evil to be abolished as soon as men are enlightened, but a permanent requirement of social cooperation. His refutation of the anarchist hope is characteristically blunt: the anarchist, he wrote, overlooks the undeniable fact that some people are either too narrow-minded or too weak to adjust themselves spontaneously to the conditions of social life, and that even granting every sane adult the capacity to grasp the benefits of cooperation, there remain the infants, the aged, and the insane. A society that wishes to preserve peace, on this view, must be prepared to crush those who break it, and the institution that does so is the state. This is not a grudging concession wrung from Mises by his critics. It is his settled position, stated in Human Action and elsewhere, and it is flatly incompatible with the claim that Misesian economics culminates in the rejection of the state and of state-recognised rights.

Thesis. The standard observation is that AI capability is uneven: systems that solve competition mathematics fail elementary perception, hold inconsistent beliefs, and come apart over long sequences of actions. The observation is correct and the usual gloss on it — that the frontier is “jagged” — treats the unevenness as mysterious. It is not. Capability tracks the density and checkability of the training signal, not the intrinsic difficulty of the task. Where a domain supplies abundant examples and a cheap correctness check, performance is extraordinary. Where it supplies neither, performance collapses, and it collapses on problems that are formally identical to ones the same model solves. The ten unsolved areas are not ten independent research programmes; they are ten places where the signal was never there, and they stand in a dependency order. Four of them are foundations — knowing what the network computes, what it is uncertain about, what causes what, and why any of it works — and the remaining six are waiting on those four whether or not their researchers say so.

The empirical literature is now too strong to ignore. Anthropic reports multi-agent systems consuming about fifteen times the tokens of chat interactions in its data and describes early agents wildly overinvesting effort. LangGraph documents looping evaluator patterns and the need for recursion limits. Huang et al. and Kamoi et al. show why intrinsic self-correction cannot simply be assumed to add reliable information. Zhang et al. demonstrate that substantial portions of multi-agent communication can be pruned while retaining comparable results. Y. Li et al. show that repeated sampling can often be stopped much earlier. Tran and Kiela show that multi-agent advantages can disappear when reasoning-token budgets are matched. Y. Kim et al. provide controlled 2026 evidence of multi-agent coordination consuming up to 6.2 times the reasoning turns of a single-agent baseline in observed configurations, with hybrid coordination reaching 515% overhead and much poorer token efficiency (Hadfield et al., 2025; Huang et al., 2024; Kamoi et al., 2024; Y. Kim et al., 2026; Y. Li et al., 2024; Tran & Kiela, 2026; Zhang et al., 2025).

The economists have a colder name for one engine of this: the information cascade. Suppose each person has a little private information and can also see what others do. The first few people act on their own information. But once enough have acted the same way, the next person reasons — rationally — that the crowd’s collective information must outweigh his own small scrap, and so he ignores his own signal and follows. So does the next, and the next, each rationally discarding what he knows in deference to what everyone is doing. And now the terrible feature: once people stop acting on their private information and start merely copying, the crowd stops gathering information at all. It is no longer aggregating a million scraps; it is echoing one early accident, amplified. A cascade can be enormous, confident, unanimous — and completely wrong, because after the first few movers, no one added any knowledge to it. Everyone was looking at everyone else. The mirror had swallowed the world.

The empirical structure of the Bitcoin network does not even satisfy the egalitarian-mesh picture that the folk inference quietly assumes. Lischke and Fabian (2016), Tao, Dai, Wu, Ho, Zheng and Cheang (2022), and Javarone and Wright (2018) document heavy-tailed degree distributions and small-world structure. Park, Im, Seol and Paek (2019) and Essaid, Lee and Ju (2023) confirm persistent role asymmetry between high-connectivity miner nodes and peripheral non-mining nodes. The Bianconi and Barabási (2001) fitness-driven preferential attachment model accounts for the condensation: mining nodes possess high fitness (uptime, bandwidth, latency advantage) and form a dense, low-diameter core; non-mining nodes occupy low-k-core shells with high churn. Eigenvector centrality concentrates more than 97 per cent of its mass on under 5 per cent of nodes (Wright, 2026b). Decker and Wattenhofer (2013) showed in the earliest measurement study that inter-miner propagation completes on the order of 100 milliseconds; Fischer and Meiklejohn (2020) confirmed that propagation bottlenecks reside at the boundary between the high-bandwidth relay tier and the general population, not within either.

A parameter is a dial: a number in a model that you are free to choose. Give a model one dial and it can shift up and down. Give it two and it can shift and tilt. Give it enough and there is almost no shape it cannot be bent into. A straight line has two parameters and can pass exactly through any two points you like. A parabola has three and can pass exactly through any three. A polynomial of degree n has n plus one, and can be driven exactly through n plus one arbitrary points — through the closing prices, if you wish, of n plus one consecutive days, no matter how those prices jump and stagger. The fit will be perfect. It will also be worthless, because a curve threaded exactly through every past point has learned the noise along with the signal, and will lurch off into nonsense the moment it is asked about a point it has not already seen.

AI aggregates across every dimension of your life. A single surveillance tool — a license plate reader, a facial recognition camera, a phone location ping — might reveal relatively little about you in isolation. But AI systems don’t operate in isolation. They integrate data from dozens, sometimes hundreds of sources, assembling a composite picture of your life that is far more than the sum of its parts. Research has demonstrated that metadata alone — the records of who you communicated with and when, stripped of any content — can predict personality traits, mental health conditions, and substance use patterns with startling accuracy. Purchasing data can reveal your religion, your health conditions, and your political beliefs. When these data streams are combined and analyzed by AI systems, the result is a form of surveillance so comprehensive that the concept of “practical obscurity” — the idea that your privacy is protected in part by the sheer difficulty of assembling scattered information about you — ceases to exist.

This is the engine of every spurious correlation paraded as proof, every prophecy that turns out to have been written after the event it foretold, every “pattern” that emerges only because someone went looking through enough noise to find a shape they liked. Give a man a large enough body of data and the freedom to define his target afterward, and he will always find a bullseye, because with the target undrawn, everything is a near miss waiting to be circled. For the motivated reasoner the appeal is obvious: he can take the raw mess of what actually happened, comb through it until he locates the fragment that supports the conclusion he wanted, draw his target snugly around that fragment, and present the result as though the universe had obligingly confirmed a prediction he never actually made. The shot came first. The target was painted to fit it. And the marksmanship is announced as though the order had been the other way around.

The labour market has now supplied a rather brutal natural experiment on how much information the certificate actually carries. Between 2017 and 2019, employers began stripping degree requirements from postings at scale — some 46% of middle-skill and 31% of high-skill occupations experienced material degree resets, and most of those resets were structural rather than pandemic-driven (Fuller et al., 2022). If the degree had been chiefly a measure of skill, dropped requirements should have produced a wave of non-degreed hiring. They did not. Analysing 11,300 roles at firms that had changed their stated requirements over a decade, Sigelman et al. (2024) found that fewer than one in 700 hires in the most recent year reflected the change; 45% of firms had altered the posting and nothing else. Employers say the degree is not necessary and then hire the degree anyway, which is roughly what you would expect if the degree were functioning as a signal whose informational content nobody has ever independently audited.

So, cultivate and practice it: the micro-eternity between inhale and exhale. Do not attempt to fill it with superficial distractions or the clamor of the external world, for its true value lies in its emptiness, its capacity to hold. Do not flee from its profound stillness or its challenging introspection, for it is there that true self-knowledge resides, unclouded by external noise. Inhabit it. That sacred space is where your will becomes sharp, honed by deliberate contemplation, a blade forged in silence. That is where your name is forged, not by external acclamation or fleeting fame, but by an act of profound self-authorship, a declaration of internal sovereignty, a testament to your unique being. The world shouts. Let it. You—you will act. But only once it is truly yours to do so, once your action is an emanation of your deepest principles, a deliberate choice born from the stillness of your own sovereign will, a testament to your mastery over self.

The implication for the “open, anonymous, equally accessible channels” requirement is correspondingly decisive. The requirement is not merely costly to satisfy—that point is developed in the section that follows—it is structurally unnecessary. The user does not require access to every miner; the user requires access to any miner. A single open, anonymous, equally accessible channel to any single miner—the GorillaPool mAPI endpoint that requires no API key, the Slipstream public terminal, any one ARC instance on the BSV network—is functionally equivalent, for purposes of confirmability, to N such channels to N miners. The miner subgraph completes the equivalence at no cost to the user and through no action of the user. The correspondent’s two-tier framing, considered against the actual topology, collapses to the observation that some channels are more convenient than others—which is true of every market that has ever existed and which is not a property of censorship resistance under any defensible definition of the term.

Abstract. I take the nine properties a trustworthy reasoning system needs — claim-level rather than token-level uncertainty; separation of epistemic from aleatoric uncertainty; source provenance; contradiction detection; reliable abstention; active information seeking; explicit tracking of assumptions; verification proportional to consequence; and calibration that survives tool use and multi-step reasoning — and show which of them have a formal answer in the architecture I have set out, and which do not. Claim-level commitment, provenance, contradiction handling, abstention, staleness decay, consequence-weighted verification and confidence propagation along inference chains all have specified mechanisms. Two do not: the epistemic/aleatoric split is absent, and the whole structure rests on the availability of an external truth signal that is thinnest in exactly the domains where the architecture matters most. I state both plainly rather than papering over them, and I identify the independence assumption in the evidence-aggregation rule as a third defect requiring repair. Two figures. Both are schematic; neither reports measurement.

The Lightning Network, in its mature form, is a network of payment hubs. The hubs hold liquidity. The hubs route payments. The hubs can refuse to route. The hubs know who is paying whom, at least probabilistically, because routing leaks information about path and amount. The hubs are, in every economically meaningful sense, banks — institutions that hold funds, provide payment services, and exercise discretion over which payments to facilitate. The fact that the hubs are non-custodial in a narrow cryptographic sense — they cannot, in normal operation, simply steal the channel funds — does not change the institutional fact that they are intermediaries, and the network they form is an intermediated network. The user who routes a payment through a Lightning hub is depending on the hub’s behaviour in a manner functionally indistinguishable from the way a payer depends on a payment processor. The cryptography rearranges the failure modes; it does not abolish them.

> Antonio Damasio posits that feelings are not afterthoughts to cognition, but the foundational scaffolding upon which consciousness is built. In his model, emotion precedes reason, and the self arises from the integration of bodily feeling with internal narrative.Applied to neural merging, Damasio’s thesis provides the neurological basis for what the subjects undergo: they are not exchanging data, but integrating affect. The merge does not show what the other thought, but what they felt—and by inducing that feeling within the self, forces reorganisation of one’s own narrative identity. Consciousness, then, becomes contaminated—deliberately and irrevocably—by foreign affect.This reframes the ethical implications of the device: it is not merely a tool for empathy or reconciliation, but an invasive reconstitution of selfhood. In becoming host to another’s sorrow, one no longer owns their original emotional topology. You emerge altered—still yourself, but also something else. And this something else cannot be separated from the feeling that made you.

This is the exact inverse of the precedent the bulls invoke. The reassuring historical analogy is information technology: alarmists once forecast that IT would consume half of US electricity by 2010, and instead efficiency gains plateaued the load near two per cent (Huber and Mill’s forecast, cited in Gill et al., 2026, p. 1). But that decoupling happened precisely because the output of a data centre is the computation it performs, so making each computation cheaper let the same output run on less power. In proof-of-work the output is the cost, and the difficulty adjustment exists specifically to keep the cost high: if efficiency improves and the network’s aggregate hashing becomes cheaper, difficulty rises to restore the ten-minute block interval, and the savings are immediately reinvested in more hashing. You cannot make a cost signal cheaper. You can only make it larger for the same money. The IT analogy runs backwards.

The evidence for this distinction is now substantial. Liu et al. (2023) showed that models could perform much worse when relevant information appeared in the middle of a long context rather than near the beginning or end. LongBench found that models struggled as contexts became longer even on tasks explicitly designed to measure long-text understanding (Bai et al., 2024). RULER then demonstrated that near-perfect performance on simple needle-in-a-haystack retrieval could coexist with severe degradation once tasks required more needles, multi-hop tracing, aggregation, or other forms of structured use (Hsieh et al., 2024). NoLiMa made the problem harder by removing easy lexical overlap between the query and the relevant evidence; at 32K tokens, 11 of 13 evaluated models fell below half of their strong short-context baseline, and even a leading model suffered a large decline (Modarressi et al., 2025). The common pattern is that context admission is easier than context utilisation.

The second feature is X-inactivation mosaicism, and it is conceptually separate from heterozygosity even though they interact. In XX individuals, one X chromosome is largely inactivated in each cell early in development, producing a mosaic pattern in which some cells express one X and other cells express the other X. This mosaicism is not a superficial footnote; it changes the way X-linked alleles manifest across tissues, including brain tissue, because it means that an X-linked variant is not necessarily expressed uniformly across all relevant cells (Migeon, 2007). Instead, its expression can be distributed across cell populations, with some cells expressing the allele from one X and some expressing the allele from the other X. That distribution can attenuate the net functional impact of a variant at the organism level because the phenotype is produced by networks of cells rather than by a single locus acting uniformly in every cell.

Consider, third, the credibility trap and the poisoning of signals. After Hannibal, Rome could no longer give Antiochus the benefit of the doubt; the symbol had foreclosed the charitable reading. In any rivalry that has tipped toward suspicion, the same inversion takes hold: ambiguous moves that might once have been read as routine get read as hostile by default, and each side’s attempts at reassurance are decoded as maneuvers. Both Washington and Beijing now operate in an information environment where the other’s gestures are routinely read in the darkest available light — where a freedom-of-navigation passage and an island fortification, a tariff and a technology restriction, a military exercise and an arms sale, are each slotted into a story the other side has already decided is true. This is Jervis’s world, and it is also Antiochus’s: once the frame is hostile, the reading machinery serves the frame.

The verification is equally benchmark-like. A quantum run gives single samples, not a full distribution, so the experiment must be repeated many times to build a statistical picture. That picture is then checked with cross-entropy benchmarking, which is a way of asking, “Do the samples look like they came from the ideal quantum distribution rather than from uniform noise?” (Boixo et al., 2018). In practice, the researchers can only compute the ideal probabilities for smaller versions of the circuit, because large ones are too hard to simulate classically. They calibrate on those smaller circuits, infer fidelity, and then extrapolate to the larger regime. This is methodologically sound for a physics benchmark. It is also unavoidably indirect: you are not verifying a single right answer, you are validating that the machine is sampling from the right sort of statistical landscape (Arute et al., 2019; NIST, 2023).

AI generates knowledge that doesn’t exist in the raw data. Traditional surveillance collects information — it records what you said, photographs where you went, logs who you called. AI surveillance does something fundamentally different: it creates new information through inference. Machine learning algorithms can analyze your purchasing patterns and predict your pregnancy before you’ve told anyone. They can analyze your social media activity and infer your political affiliation, your sexual orientation, your mental health status. They can analyze your movements and predict where you’ll be tomorrow. This inferential capacity has no real precedent in the history of surveillance. The government isn’t just learning what you’ve done; it’s generating predictions about who you are and what you’ll do — predictions that you never disclosed to anyone, because they existed only as latent patterns in data you didn’t even know you were producing.

The power of the system emerges in the elegance of its minimalism: a single hash—just a few bytes of output—can anchor an entire universe of information into an immutable chronological structure. This is the triumph of hash-based anchoring. It is not the content itself that needs to be stored, nor the file, nor the document, nor the agreement; it is the cryptographic fingerprint, the mathematically unique digest that binds the underlying data to a specific moment in time. That fingerprint becomes the anchor—small enough to fit seamlessly inside a transaction, strong enough to serve as an incorruptible witness. The ledger does not need to contain the world; it only needs to contain the proofs that the world existed as claimed. In this, the system achieves a universality no physical registry, no central authority, no isolated archive could ever hope to match.

My disagreement begins one step earlier. The fact that information has become cheap does not establish that education has become obsolete. It establishes that information delivery has become cheap. The distinction matters because universities have spent decades confusing the delivery of information, the certification of competence, the signalling of status and the education of a person. Once those functions are separated, the apparent contradiction disappears. The internet is devastating to institutions whose principal product is content delivery. It is highly disruptive to institutions whose principal product is routine training. It is potentially devastating to institutions whose principal product is a credential. But none of those propositions demonstrates that the formation of judgement, intellectual independence and the capacity to reason well has become unnecessary. If anything, a world in which anyone can obtain an answer instantly makes those capacities more important.

Social media has made this catastrophically visible. The digital public square operates on exactly the principle that Madison identified as fatal to republican governance: it amplifies passion and suppresses reason. It rewards the inflamed will and punishes the considered will. It selects for the factional impulse — the common passion, the tribal signal, the identity-confirming assertion — and selects against the deliberative faculty that Madison’s system requires. And a population that has been trained rather than educated has no internal resistance to this selection pressure, because the capacity to resist the pull of factional passion is not a skill. It is a habit of character, formed by education, and specifically by the kind of education that teaches a person to hold his own convictions at arm’s length, to subject them to scrutiny, to entertain the possibility that he is wrong.

The problem becomes more perverse when users feel faster while objectively becoming slower. Becker et al. (2025) conducted a randomised controlled trial with experienced open-source developers working on real issues in repositories they knew well. Developers using early-2025 AI tools took 19% longer to complete tasks than those working without AI assistance. Before the study, participants expected AI to speed them up by 24%; after experiencing the workflow, they still believed it had sped them up by 20% (Becker et al., 2025). METR later warned that its subsequent 2026 data were affected by severe selection problems and provided an unreliable signal of the current productivity effect, so the 19% figure should not be generalised to every modern AI coding workflow (Becker et al., 2026). The important point is narrower and robust: subjective impressions of AI productivity can be badly miscalibrated.

The hydroponic greens engine is the highest-frequency production unit in the system, and it is treated as such: a compact footprint designed to turn water, nutrients, and labour into a reliable stream of edible greens with short cycle times. The core module occupies a 3 m × 3 m area, supplied from well water, and planted on standard lettuce and rocket rhythms. In year-one terms, the decisive attribute is not the footprint itself but the throughput it supports. Under the operating practice described, the module produces around 1,200 units of mixed greens per cycle. The word “unit” is used deliberately as a harvest-count measure at this preliminary stage, because the present aim is to map system capacity and cadence before moving to a weighed-output dataset; later measurement will translate these units into grams, rejected fractions, and household portions.

Capital is not a metaphysical substance. It is saved effort looking for a future task. In healthy economies it migrates towards undertakings that enlarge productive capacity: better tools, new techniques, wider markets, deeper knowledge. In a diseased economy it migrates towards whatever yields the easiest paper gain. If the highest return comes from hoarding, hoarding becomes the national vocation. Savings that should have financed the difficult, slow miracles of civilisation—factories that take years to plan, research that fails ten times before it succeeds once, infrastructure that yields quietly over generations—are poured instead into a ledger that produces nothing except the hope that someone else will pay more later. The crowding-out is not mere theory. It is the plain geometry of finite resources: every pound locked in sterile scarcity is a pound not at work in fertile enterprise.

The advantages of size in the production of coercion compound in a way they do not for ordinary goods. A larger protection-insurance firm commands better intelligence, more armed personnel, deeper databases, superior surveillance, stronger legal drafting, larger capital reserves against correlated shocks, and greater bargaining leverage over everyone it deals with. Crucially, it can threaten a smaller agency far more credibly than the smaller agency can threaten it, and it can impose its terms on the arbitration networks that every agency needs, refusing to deal with firms that will not accept its rules. Each of these advantages raises the return to being larger, and the firm that is larger is better placed to become larger still. The competitive equilibrium the thesis assumes is not stable, because the product itself rewards the accumulation of the capacity to coerce.

First, the location. You might reasonably expect the trouble to show up in the part of the physics that fluid dynamicists care about most: the shear, the sloshing and sliding of the fluid that carries viscosity and drives the spectacular flow patterns seen in collider data. It doesn’t. The shear part stays perfectly well-behaved. In fact, it settles down to a tidy, exact number — the fraction 90/7 — and just sits there, calm, no matter how light the particles get. The infinity hides somewhere else entirely: in the compression part of the physics, the bulk squeezing of the gas. All the discarded information, all the divergence, all the failure of the guess, is confined to that one channel. The part everyone watches is fine; the part fewer people watch is where the description quietly breaks.

To reach a regime capable of breaking RSA-2048, a machine would need an estimated 20 million physical qubits arranged into logical qubits with error rates below 10⁻¹⁵ per gate. Each logical qubit would be encoded using thousands of physical qubits through complex surface codes that maintain coherence and detect fault propagation. The required coherence times, gate fidelities, and readout precision are not incremental improvements upon what we possess; they are many orders of magnitude beyond it. Current systems boast coherence times measured in microseconds, not minutes; error rates of 10⁻² to 10⁻³ per gate, not 10⁻¹⁵; and thermal noise that overwhelms the delicate interference patterns on which quantum logic depends. The scale of the gap cannot be overstated—it is not a step from a prototype to a product, but from an idea to an impossibility.

This is why the defence of markets is not a lullaby about rich men being generous. It is a hard statement about information and limits. Prices are not decorations; they are signals. They compress scattered knowledge—preferences, scarcities, risks, alternatives—into something that allows millions of strangers to coordinate without a single mind pretending to be omniscient. A planner cannot know which factory should expand, which crop should be planted, which technology will matter, which style will suddenly take off, which medicine will fail, or which novelty will quietly become a necessity. He cannot know it because nobody can know it ahead of time. The market does not know it either. The market discovers it through error, experiment, and correction, and it does so without requiring everyone to be marched in lockstep behind a single dream.

What I keep coming back to is that the question I started with — who controls a blockchain? — does not have a single answer, and the framework’s main payoff is to explain why it doesn’t. Control is distributed across five layers, the producer holds only one of them, and the system’s behaviour under stress depends on how separable the other four are from each other and from the producer’s. When separability holds, the existing models are roughly right and the chain behaves as the textbook says. When it fails — when the people setting the price are the people producing the blocks — the feedback loops that drive lock-in get a direct channel, and the chain can tip toward a concentrated outcome that no individual chose and that history, not fundamentals, selected.

This is where “biological fragility” must be stated carefully. A well-planned vegetarian or vegan diet can be adequate. The fragility is not that it cannot work; the fragility is that it is compliance-dependent. The diet’s adequacy depends on consistent execution across time. In real households, compliance is not evenly distributed. It varies with parental education, conscientiousness, mental bandwidth, financial resources, and the practical realities of childcare. It varies across school environments, social events, and caregiver changes. It varies during illness periods when appetite collapses and dietary diversity narrows. A diet that is “adequate if engineered” therefore has a predictable distribution of outcomes: some households execute it very well; some do not; and the children in the lower tail of execution are the children the thesis predicts will show growth and cognitive penalties.

The fundamental problem is technical. Modern AI systems are trained on datasets that inevitably contain personal information — names, faces, biographical details, health data, financial records — scraped from the open internet or obtained through data partnerships. And “deleting” someone’s data from a trained neural network isn’t like deleting a row from a database. The information isn’t stored in discrete, retrievable units. It’s encoded in the statistical weights of the model, distributed across billions of parameters. Research has shown that large language models can memorise and regurgitate verbatim passages from their training data, including personal information. Meaningful erasure might require retraining the entire model from scratch — a process costing millions of dollars and weeks of computation. The gap between what privacy law demands and what AI technology can deliver is enormous.

Network-friction loss. A pool depends on infrastructure it does not own: domain naming, internet routing, interconnection, propagation. Root-zone and identifier coordination remain externally governed, and the security literature — Apostolaki and colleagues on routing attacks against Bitcoin, and the routing-integrity guidance from the National Institute of Standards and Technology — shows that interference at this layer can isolate mining capacity, slow propagation, and raise the rate at which a pool’s blocks are stranded. The paper is careful here: it makes no claim that any single state can simply switch off another country’s miners. The narrower, more defensible point is that a publicly identified, dishonest pool is operationally entangled with infrastructure it cannot unilaterally control, and that entanglement is one more margin on which an attack can be made to bleed.

And consider, finally, the thinness of reassurance. The ancients had no way to make a promise stick and no shared language for good faith, and so their negotiations clarified incompatibility instead of dissolving it. We have far better machinery — and yet the deepest problem survives the upgrade. The commitment problem is not solved by having more channels. A rising power cannot credibly bind its future self any more than Antiochus could, because both sides know that promises last only as long as the balance that backs them. When the established power suspects that today’s reassurance is merely tomorrow’s tactical pause, the reassurance loses its power to reassure. More bandwidth does not fix that. It can even make it worse, by multiplying the signals available to be misread.

PIAAC also gives the hard proportions that let the grade-level claim breathe. In earlier rounds, about 19 per cent of U.S. adults sat at Level 1 or below; by the 2023 round that share had risen to roughly 28 per cent, while the high-performing share fell. (National Center for Education Statistics) The 2019 NCES data point counts tens of millions at Level 1 and millions below Level 1, with the below-Level-1 group effectively functionally illiterate—unable to determine the meaning of sentences in ordinary prose, locate a single piece of information in short texts, or complete simple forms. (National Center for Education Statistics) If you prefer plain speech: that is not “a bit weak at reading.” That is living in a printed world with the tools of a tourist.

Britain’s 19th-century dominance rested on a maritime system tuned to sail, coal, and steam (ports, coaling stations, shipbuilding, doctrine). Churchill’s 1913 remarks encapsulate the lesson: oil freed ships from coal but made them dependent on foreign fields and tankers—“safety and certainty in oil lie in variety” of sources and routes (Churchill, 1913). The U.S./China AI competition has a similar logic. In place of “fuel logistics” are compute and data logistics: how to field massive compute clusters, data pipelines, and software for autonomy. China’s leaders recognize this. PLA Strategic Support Force doctrine and national AI plans emphasize building an “intelligent” military using big data, cloud, and AI, and treating data and chips as strategic resources (China’s State Council, 2017; Department of Defense, 2025; National Institute for Defense Studies, 2022).

A second confounding issue is more subtle and more important for the thesis: “vegetarian” is not a biological exposure. It is a label that contains heterogeneous nutritional realities. In high-income settings, a well-planned vegetarian pattern can be associated with high diet quality, high fruit and vegetable intake, lower processed-food intake, and strong supplementation adherence. That constellation can improve health and can correlate with higher cognitive outcomes through multiple channels. This is why vegetarian children can look good in observational data. But a superficial positive association does not refute the thesis, because the thesis is not that vegetarian identity causes harm; the thesis is that developmental constraint emerges when adequacy fails, and adequacy failure is more likely when the diet requires precision and that precision is not maintained.

Finally, the essay maps this history onto the emerging AI/autonomy era. If coal–steam power was conditioned on coaling stations and shipping capacity, then AI-enabled military power is conditioned on compute, data, software ecosystems, resilient command-and-control (C2), and manufacturing scale for low-cost autonomous systems. Evidence from Ukraine, research from RAND and CSIS, and even large-scale civilian drone displays underscore the logic of mass, attritable, networked systems: quantities that can saturate defenses and compress decision cycles. The implication is not that aircraft carriers become irrelevant tomorrow, but that drone swarms and autonomy may change the cost-exchange ratios and the operational art—forcing states to re-architect procurement, doctrine, and defenses around “cheap–smart mass” (Bondar, 2025; CGTN, 2026; Franke, 2025; Gerstein & Leidy, 2023; Guinness World Records, 2025; Slusher, 2025).

The contrast to the BTC ritual is not subtle. The ideology says that assurance equals everyone processing every byte of every block forever. That is not a control system; it is a pageant. Redundant re-validation of the same data by a million hobbyist machines does not multiply information; it multiplies heat, bandwidth, and latency. The statistic that matters—uncertainty about the population—falls with correctly designed n, not with parallel repetition of the same check against the same object. In statistical terms, overlapping, identically constructed samples do not generate independent evidence; they generate correlation without new information. In systems terms, gossip floods and universal replay generate hot links and buffer thrash; they do not generate additional confidence beyond what properly configured sampling and proof-based verification already provide.

U.S. posture and doctrine. The U.S. Navy’s force structure is officially anchored in carriers. A Congressional Research Service report notes the Navy has 11 operational nuclear carriers and by law must maintain at least 11. The 2020 tri-service Advantage at Sea maritime strategy affirms that carriers remain central but calls for “new platforms…new thinking…and new technologies” to enable distributed naval operations against adversary surveillance (Congressional Research Service, 2025; Royal Navy et al., 2020). The 2022 National Defense Strategy explicitly labels China as the “pacing challenge,” mandating urgent modernization but not prescribing precise force trade-offs (Department of Defense, 2022). In practice, U.S. naval doctrine is evolving—Distributed Maritime Operations, JADC2 connectivity, unmanned experimentation—but large platforms are still heavily funded and not scheduled for replacement anytime soon.