Nature of Intelligence: The Moves

What high intelligence buys is not more computation but relocation: the power to move to the level where a problem is defined, priced and framed, and to trade on the gap. Call it altitude arbitrage. And like any arbitrage, it is real, it is offense-dominant, and it decays the moment it diffuses — which is why, in an age when machine intelligence is climbing the same ladder from the top down, the durable edge is quietly sliding back to the ground floor everyone is fleeing.

Built on a library of 259 primary documents — peer-reviewed cognitive science, AI-capability measurement, economics of innovation and organisation — compiled by the ENSI Foresight Division, July 2026.

We tell ourselves that the very intelligent simply think more: that they run the same race as everyone else, only faster and further. It is a comforting story because it makes intelligence a scalar — a single dial of horsepower you either have or lack — and scalars are easy to admire, hire for and envy. The story is also, at the top of the range, largely wrong. Watch what a genuinely formidable mind actually does and you rarely see it out-computing the room. You see it changing rooms. It picks the board, sets the win condition, and prices the game before anyone else has looked up from the pieces. Everyone else is solving; it has already decided which problem counted. The gap between the people doing the work and the person who chose the work is not a gap in speed. It is a gap in altitude.

That reframe is the spine of this piece. The extra that high intelligence buys is not more thinking but the capacity to relocate to the level where a problem is defined, priced and framed — and to trade on the difference between that level and the one everyone else is stuck at. The name matters, so let us fix it: this is altitude arbitrage, and we retire the placeholder “extra moves” for it deliberately. An arbitrage is not a gift; it is a position. It exists only because of an asymmetry — you can see a price others cannot — and it carries all of arbitrage’s physics: it is offense-dominant, it is hard to capture, and it is perishable. It pays until it is noticed. The instant the trade is written down and the crowd piles in, the spread collapses. Everything true of altitude is true of arbitrage, which is exactly why the fashionable counsel to “operate one level up” describes a real edge and a wasting asset in the same breath.

This is not an idle observation about clever people. It is a foresight question about capability — the sort of question the ENSI Foresight Division exists to ask — and it has become urgent for one reason: the machines are now climbing the same ladder, and they are climbing it from the top down. Frontier AI’s task-completion horizon has doubled roughly every seven months for six straight years, on METR’s measurements, so the length of work an agent can carry unaided keeps lengthening on a predictable clock. And the rungs falling first are the abstract ones the folk theory prizes most: large models already match or beat humans on analogical-reasoning tests, are rated more novel than expert researchers when generating ideas, and reach adult performance on high-order theory-of-mind. Meanwhile the physical floor — the plumber’s hands, the machinist’s feel — remains stubbornly beyond them. “Meta stays human” is not a law of nature. It is a market prediction, and the market is already trading against it.

Which sets up the twist that turns the whole argument over. Because altitude is the most codifiable asset a mind can own, it is also the most perishable. A reframe fits on a slide; a craft takes a decade of apprenticeship. Frameworks travel frictionlessly and self-commoditise; tacit skill does not travel at all. So as the meta-layer is competed away — by rivals reading the same playbook and by agents that generate the playbook on demand — durable scarcity does not evaporate. It re-concentrates at the bottom, on the tacit ground floor the meta-thinkers spent a generation abandoning. In the United States alone the skilled-trades and manufacturing gap is on track to leave more than two million jobs unfilled and to cost the economy on the order of a trillion dollars by 2030. The people who ran up the ladder left money lying on the floor.

The stakes, then, are not bragging rights but capture. You can be the most intelligent actor in a value chain and bank almost none of the value you create. William Nordhaus put a number on it: innovators appropriate only about 2.2 per cent of the total social surplus their innovations generate; the other 97.8 per cent leaks to consumers, imitators and complementors. Intelligence buys the insight; someone else buys the yacht. Creation and capture are orthogonal, and the meta-layer — pure idea, pure judgement, pure reframe — is precisely the non-excludable layer that markets drive toward zero. For an individual this is a career warning. For a state betting its future on “high-value cognitive work,” it is a strategy warning: the rungs you are racing your workforce toward are the rungs that appropriate least and commoditise soonest.

So what follows is a ledger, not a hymn. Below, the ten moves that intelligence unlocks are laid along a single axis — raw horsepower at the bottom, pure meta at the top — and each is examined through the same six questions: what the move is, why intelligence buys it, where it sits on the ladder, who is driving its price to zero, how it turns into a liability, and how a person or an institution should actually allocate toward it. Read every move as a priced trade. Three findings recur like a refrain, and they are worth holding in mind from the first line: the meta-moves are offense-dominant, they are more IQ-orthogonal than the romance admits, and they are perishable. And keep one eye on the floor throughout — because that is where the last durable scarcity is hiding, and where the ENSI lens, always asking what a country should do, keeps landing.

The argument in brief

• What intelligence buys is altitude arbitrage, not horsepower. The premium is a position — relocation to where a problem is defined and priced — not a faster engine. Real, powerful, and structurally the same as a financial arbitrage.

• Altitude is offense-dominant. The high moves — reading the game, modelling minds, reframing — are weapons that cut toward whoever holds the better model, not shields that rise with your own IQ.

• Altitude is more IQ-orthogonal than we admit. The bias blind spot does not shrink with cognitive ability and may grow; the financially sophisticated are over-represented among fraud victims; trained cognitive skill barely transfers across domains.

• Altitude is perishable. Codifiable meta-moves diffuse frictionlessly and self-commoditise — published market anomalies lose roughly a third of their return once the crowd acts on them — so the premium is the least durable asset you can own.

• Creation and capture are orthogonal. Innovators appropriate ~2 per cent of the surplus they create; the meta-layer is the non-excludable one, so the smartest actor routinely captures the least.

• The machines are climbing top-down. Abstract rungs (analogy, idea generation, theory-of-mind) fall to AI first; the physical floor falls last. “Meta stays human” is a testable prediction, and it is being falsified rung by rung.

• Durable scarcity re-concentrates on the tacit ground floor. As the meta-layer commoditises, the lasting edge slides back to build, craft and deep in-domain judgement — precisely what everyone is fleeing.

• The move for a person: climb for the leverage, but fuse every meta-move to a ground floor that cannot be reframed, retrieved or copied away.

• The move for a state: stop racing the whole workforce to the commoditising top; treat foresight and judgement as an owned capability — run by an agentic engine, anchored by a defended tacit base — not as a personality trait to recruit.

How the ledger is ordered

The ten moves run from BUILD to RECURSE, and the ordering axis is the same one the folk theory uses to flatter the top: horsepower at the bottom, meta at the top. The novelty is what the ledger tracks alongside each move — not merely what it does, but who is competing its price to zero and how fast. Stuart Russell’s hierarchy of rationality is the frame underneath the whole thing: perfect rationality is unavailable to any physical agent, because thinking itself consumes the resource it is trying to spend well; what a real mind can achieve is bounded optimality — the best behaviour given finite information and finite compute. Altitude is what bounded optimality looks like when it is winning. And the governing claim, stated once so the rest can lean on it, is this: moving up the ladder relocates you to where the problem is priced, but every rung up is cheaper to copy, harder to bank and closer to a machine — so altitude buys leverage and fragility in the same motion.

1. Build — the floor is the moat, not the commodity base

BUILD is the instantiation of an idea into atoms: the working engine, the shipped device, the process that actually runs. The romance files it as raw horsepower — the commodity beneath the clever moves — and the romance has it exactly backwards. The scarce ingredient in building is tacit: the process knowledge that lives on the shop floor and never makes it into the blueprint. Michael Polanyi’s formulation remains the whole point — “we can know more than we can tell” — and the industrial record bears it out: the operators who take over new production equipment, Piore found across decades of observation, “understand the work in a different way from the engineers,” because the knowledge is made at the floor, not abstracted down to it. This is why abstraction cannot simply climb down and copy a build.

Where it sits, and who commoditises it, is where the spine turns over. BUILD sits at the horsepower floor — and it is the rung the machines have climbed least. Moravec’s paradox has become a price signal: the sensorimotor skills evolution optimised first are the hardest to automate, and while abstract cognitive tasks are forecast to automate within roughly a decade, dexterous physical work is scored closer to a century out. One exposure index across nineteen thousand tasks finds management and STEM occupations the most exposed and construction and maintenance the least. So the abstract rungs deflate first and the floor reprices up — the trillion-dollar manufacturing gap is what a repricing floor looks like in the labour market.

The failure mode is treating BUILD as delegable downward, and the deep-tech graveyard proves the cost: hard-technology ventures die not in the laboratory but in the “valley of death” between proof-of-concept and reliable manufacturing — a company can be technically ready and still “existentially fragile” because its manufacturing readiness is three rungs behind. Google, remember, did not win on a secret idea; the idea “wasn’t secret or even new” — it executed better. The allocation is therefore counter-intuitive and, for a mid-sized industrial economy like the Czech Republic, strategically live: own the boring, tacit, hard-to-copy build layer the meta-thinkers are abandoning, and let the complementary asset — the factory, the process, the installed base — bank the rent. The floor is the least-commoditised rung, not the most.

2. Solve — a cheaper encoding, not a bigger engine

SOLVE is holding a working model of a messy system in the head and pruning the search space instead of brute-forcing it. The mechanism is not a larger engine; it is a cheaper encoding. Nobody beats the branching factor — chess opens onto something like 10^123 positions, and even optimal pruning leaves it astronomically intractable; Deep Blue’s edge was deeper hardware search, not a defeat of the exponential. Genius encodes around the explosion. The chess master’s working memory is ordinary — the roughly four-chunk limit is fixed, and expertise does not widen it — but the master packs far more information into each chunk, seeing configurations where the novice sees thirty-two pieces, drawing on an estimated fifty thousand stored patterns. SOLVE is lossy compression, and its value is only ever as good as the regularity it compresses.

That is also its commoditiser and its ceiling. Where the world is regular and high-validity, machines prune searches no human could attempt — AlphaFold predicted the structure of some two hundred million proteins and closed a fifty-year grand challenge well enough to win a Nobel. But the human premium is claimed precisely in the low-validity domains where compression fails and confidence becomes, in Kahneman and Klein’s exact phrase, an “illusion of validity” — valid intuition requires a regular environment and prolonged feedback, and where either is missing, expert certainty is noise in a lab coat. Worse, the master’s chunks are domain-welded: forensic and radiology experts show no advantage on structurally similar tasks outside their field — a radiologist is no better than you at a “spot-the-object” search. So the allocation is not a portable “problem-solving upgrade” but deep domain comprehension — the substrate that generates good chunks — because a fast-and-frugal heuristic wins only when it matches the structure of its environment.

3. Navigate — read the game, and mind the dark twin

NAVIGATE is reading how a field actually works and finding its leverage points — the first genuinely meta rung and the prototype of the whole trade: relocate to where the game is scored, and win a contest the other players do not know is being played. It is, plainly, a power technology. But two collisions keep it honest. Outcomes ride execution and timing as much as insight — expert entrepreneurs do not out-forecast, they act and adapt — and “navigating the game” is partly a group property: social sensitivity and theory-of-mind predict a team’s collective intelligence more than any individual’s IQ.

Its commoditiser is the market itself, and this is where perishability bites hardest. “Read the game” advice is codifiable, so it diffuses and self-arbitrages: academic study of stock-return predictors finds anomalies decay by about 35 per cent after publication — a quarter of it attributable purely to arbitrageurs acting on the now-public signal — and the momentum factor that paid around 10 per cent a year in the 1990s pays closer to 2 per cent today. The map alters the territory; Soros called it reflexivity, and it is simply arbitrage decay wearing a philosopher’s coat. And the value that is captured accrues to whoever holds the complementary asset, not the navigator.

The dark twin is extraction — navigation that consumes the commons it feeds on. Maximal-extractable-value searchers on blockchains front-run ordinary users purely because they can see a transaction queue others cannot; the market-for-lemons result shows the better-informed party exploiting the worse-informed until the market can collapse entirely. And pure strategy without ground truth fails spectacularly: Olympic host cities have overrun their budgets in 100 per cent of Games, by an average of 156 per cent in real terms. The codifiable half of navigation diffuses to everyone; the durable edge is the tacit, industry-specific read that takes decades to earn — which is why the founders of the fastest-growing firms average forty-five, not twenty-five.

4. Arbitrage across domains — the signature move that commoditises first

ARBITRAGE is spotting that a solved problem in one field is secretly the same shape as an open one in another, and carrying the solution across. “Most big insights are transfers, not inventions” is the romantic heart of the altitude thesis — and it is mostly retrospective. Watched as it actually happens, working scientists’ productive analogies are overwhelmingly local and near-domain; Dunbar’s live studies of molecular-biology labs found the distant cross-field leap to be rare, not routine. Spontaneous transfer is a coin-flip at best: given a semantically distant source, only about 30 per cent of people apply it without a hint — roughly seventy per cent miss the analogy even with the answer in the room. Creativity, too, is domain-specific rather than a general trait you carry between fields. The honest recipe is the one Uzzi found across 17.9 million papers: a large base of conventional combinations pierced by a sliver of the atypical, which roughly doubles the odds of a landmark result.

And this is plausibly the first meta-move the machines commoditise — because the search-and-retrieval half of transfer is exactly what they are built to scale. Transfer has always been a search problem humans do badly; large models already show emergent analogical reasoning that matches or beats people on novel induction tasks. So the premium migrates up the ladder — from spotting the analogy to judging which analogy is load-bearing and building it against a resistant world. The new bottleneck is judgement, not intelligence. (The caveat that keeps the machines honest: reinforcement-learning gains “generalise inconsistently and can vanish” on domains with different reasoning patterns — transfer is fragile for silicon too.) The failure mode is the ingrained metaphor that keeps running after source and target diverge, importing wrong assumptions wholesale; the deeper drag is the “burden of knowledge” — as fields deepen, the age of first invention rises and the lone polymath is priced out (sustaining Moore’s Law now takes some eighteen times the researchers it did in the early 1970s). So the allocation is double-domain depth — real grounding in both the source and the target — plus teams and tools that span fields, not a lone-genius knack for analogy.

5. See further ahead — mostly meta-discipline, hard-capped by physics

SEE FURTHER is simulating second- and third-order consequences over horizons where others lose the thread. The romance imagines a bigger internal forward-model; the evidence says the part that works is discipline, not depth. Superforecasters win on process — reference-class thinking, frequent updating, actively open-minded search — at an average IQ around 115, not 160; and deeper single-model simulation actively hurts in high-branching problems, the documented “lookahead pathology” where more search yields worse decisions. Confident deep simulation is, empirically, the worst-performing style.

Its commoditiser is a physical constant. Atmospheric predictability has an intrinsic wall — about two weeks — that no amount of data or compute can cross; beyond the horizon, extra intelligence buys exactly zero. Inside the wall it buys a real but bounded extension: superforecasters see roughly 300–400 days out about as well as ordinary forecasters see 80–100. And machine horizons are climbing toward the same wall on that seven-month doubling clock, while human experts miss essentially all turning points. The failure mode is that more distance buys more confident error — Tetlock’s twenty-year record of 82,361 forecasts from 284 experts barely beat chance, and expertise correlated with better excuses, not better calibration. The lever that actually pays is a disposition: across three studies, actively open-minded thinking was the only trait that predicted forecast accuracy, and it worked by driving people to gather more information. So the allocation is not a smarter simulator but openness to disconfirmation, externalised as process — foresight as an operating discipline, which is precisely how ENSI has always insisted it be built.

6. Model other minds — a weapon, not a shield, and the smart are easiest to aim at

MODEL MINDS is reading intent, reconstructing another’s reasoning, and detecting when you are being lied to or steered — the capability the folk theory treats as intelligence’s defensive home turf. Invert it: this move is offense-dominant and IQ-orthogonal. Humans detect lies at 54 per cent accuracy across 206 studies and 24,483 judges — a hair above a coin-flip — and catch fewer than half the lies that matter. The bias blind spot does not shrink with cognitive ability; if anything it grows with it. Mind-modelling is not a scalar shield that rises with your IQ. It is a targeting weapon that cuts toward whoever models the other better — and its genuinely prosocial payoff is collective, since theory-of-mind predicts a group’s collective intelligence far better than any member’s brilliance.

Its commoditiser is machine persuasion, and the numbers are sobering. A personalised GPT-4 had 81 per cent higher odds of shifting someone’s position than a human opponent — using only basic demographic data — and large models out-persuade incentivised human persuaders in both truthful and deceptive conditions. Higher-order theory-of-mind emerged spontaneously with scale, with GPT-4 exceeding adult humans on sixth-order “A thinks that B believes that C wants” inferences. The decisive edge comes from modelling the target, not from being abstractly smarter — and that lever is now cheap to point at everyone, which is why the “hypernudging” and “instrumentarian power” the surveillance-capitalism literature warned of is a governance question, not a sci-fi one. The failure mode is that your own self-model of competence becomes the attack surface: the financially sophisticated are over-represented among fraud victims precisely because they trust their model-building and skip the verification a naive person would perform. Manipulative skill is Machiavellian strategy, not raw g — the “dark triad” barely correlates with intelligence — so the edge is a learnable technique, hence governable. The allocation is not horsepower but external verification and dispositional rationality, and the discipline of treating “operate one level up on yourself” as the exact lever a superior modeller will pull on you.

7. Reframe or dissolve the problem — meta-selection, until everyone reframes

REFRAME is redefining a problem so it collapses, or choosing which problem is even worth attacking — meta-selection over raw solving, and the flagship altitude trade: relocation to where the problem is defined and priced. But split it in two. The interpretive half — reading a situation as admitting several framings — resists crowding, because it does not reduce to a rule. The mechanical half — the teachable “ask better questions” checklist — decays like any factor; and the celebrated scientific-method edge turns out to pay through disciplined termination of bad ideas, not clever moves. Reframing is not free either: meta-reasoning is charged against the same finite budget as object-level thought — the value of a computation minus its cost.

Its commoditiser is Goodhart’s law plus the crowd. Once “go meta / choose the problem” becomes the universal instruction, the signals of meta-ness get gamed — when a measure becomes a target it ceases to be a good measure — and the crowded judgement trade competes its own premium away. Codified reframes diffuse frictionlessly; pure-idea appropriability collapses; management concepts even follow a measurable bell-shaped fashion cycle. The very ubiquity of “reframe / operate one level up” advice is the tell that its premium is being competed away. The failure modes are two: narrative capture — reframing as technocratic enclosure of the epistemic commons — and the fallacy of composition, because if everyone climbs to meta, allocation breaks. Production is complementary, not substitutable: the O-ring theory shows a single weak link degrading the whole multiplicative output, and comparative advantage says not everyone should specialise upward. A brilliant reframe times a botched build is still a botched product. So the durable move fuses meta-selection to a tacit ground floor that cannot be reframed away — and, for a society, means resisting the temptation to send everyone to the top.

8. Coordinate complexity — instrument the organisation, do not enlarge the head

COORDINATE is holding a large system — an organisation, an argument, a machine — in coherent relation while others watch it fragment. The source note frames it as one enormous working memory. That is the scaling anti-pattern. Working memory is a hard four-chunk channel that expertise cannot widen; the coordination premium at scale comes from externalised cognition — structure, process, and the “transactive memory” of a group that is more capable than any of its members. Ben Horowitz, who could plausibly have held a company in his head, says the scaling move is to stop and learn “the black art of scaling a human organisation.”

This is a rare rung where the human moat maps onto a lawful, closing boundary. AI still breaks precisely here — state-of-the-art agents perform strongly on short tasks but “break down” on long-horizon, interdependent sequences. The winning human design conserves scarce cognition by decoupling: mirroring organisational structure to problem structure (across 142 studies), decentralising into a “team of teams,” and — in software — treating loosely coupled architecture as a top predictor of delivery performance. Team dynamics beat team composition: Google’s study of 180 teams found psychological safety, not star density, was what mattered, and the collective-intelligence factor is not strongly correlated with members’ average or maximum IQ. The failure mode is the founder bottleneck — a single point of failure, a decision architecture in which strategy lives implicitly in one head; premature scaling on a hero-coordinator kills roughly three-quarters of high-growth startups. Deming’s estimate governs the fix: 94 per cent of trouble belongs to the system, not the individual. So the allocation is to instrument the workflow — build the system a competent-but-ordinary team can run — and, tellingly, this is the exact shape of the agentic engine ENSI advocates for public institutions: externalise coordination into auditable systems rather than betting the state on a handful of irreplaceable minds.

9. Generate new concepts — real power, downstream of an adoption lottery

GENERATE is inventing the vocabulary, frameworks and aesthetics that others then think with — the categories that become invisible infrastructure. The power is real: adopted categories literally rewire perception, down to the pre-attentive level where the language you speak measurably changes an early visual brain response to colour. But which concept wins is not decided by correctness. Kuhn’s five values for theory choice are, in his own word, “imprecise” and underdetermine the decision; the winner is selected by community adoption and memetic fitness. And discovery is frequently inevitable — the history of science catalogues around 148 major simultaneous, independent discoveries — so concept-creation is a spark followed by a lottery, not pure horsepower.

Its commoditiser is the collapse of creation cost. As Martin Casado puts it, the microchip drove the marginal cost of compute to zero, the internet drove distribution to zero, and “these large models bring the marginal cost of creation to zero” — and LLM-generated research ideas are already rated more novel than expert humans’. So when creation is nearly free, the scarce, decisive step migrates to distribution and adoption: the firm that invents a category can create enormous value and capture almost none — Stability AI open-sourced Stable Diffusion and commoditised itself, and across the generative-AI stack it is the infrastructure vendors that bank the dollars while the model-makers who created the market struggle to reach scale. The failure modes are twin: intelligence is uniquely fluent at persuasive-but-empty vocabulary — Sokal’s hoax passed because it “sounded good and flattered the editors” — and every winning concept escapes its author into a cage, with credit tracking fame through the Matthew effect rather than contribution. So the allocation is not to coin the concept but to distribute and build it, and to anchor it to a tacit substrate — because ideas are cheap and execution is the moat.

10. Recurse — turn intelligence on itself, the crown with the worst individual record

RECURSE is turning intelligence on itself — metacognition, better tools for thought, compressing experience into structure so the next problem costs less. It is the apex move, and for the individual human mind it has the worst empirical record on the ladder. Cognitive “brain training” produces a grand-mean effect near zero, with far transfer “null when placebo effects are controlled,” and the better the study’s controls, the closer transfer shrinks to nothing. Turning intelligence inward can even make it self-sealing: the bias blind spot grows with cognitive ability, and clever people are better at constructing arguments for the conclusions they already hold — the “intelligence trap.” The compounding that actually reshaped human intelligence came from external notation: cities that adopted the printing press grew 20–35 percentage points faster over the following century — dissemination and standardisation, not smarter private minds.

And here is the double-contrarian twist. The move that fails for humans is exactly where machine progress is fastest — because machine self-improvement is externalised in inherited code. The Darwin-Gödel Machine rewrites its own codebase and more than doubled its coding-benchmark score; AlphaEvolve mutated algorithms in an evolutionary loop and found the first improvement on Strassen’s 1969 matrix-multiplication result in its setting; and the whole capability compounds on the seven-month doubling clock. The apex of the human ladder is being automated from underneath, precisely because machine recursion compounds when written down and inherited rather than kept in one skull. The failure mode is a self-improvement market running ahead of its evidence — Lumosity paid the US regulator two million dollars for unfounded claims — and a partly-fixed ceiling that practice does not erase. The brain already runs near thermodynamic perfection, at about twenty watts. So the allocation is not a “thinking upgrade” but external, shared instruments — retrieval practice, spaced practice, tools, code — the cheap procedures anyone can copy, and instrumenting the organisation rather than the thinker.

The agentic engine: what this means for a state, and how agents run it

Read the ledger from a national vantage point and a single conclusion assembles itself: the rungs a state instinctively races its workforce toward — analysis, arbitrage, foresight, reframing, concept-creation — are the rungs that appropriate the least value and that machines are commoditising first. The instinct to “move everyone up the value chain” is, on this evidence, a partly self-defeating strategy: it crowds the perishable trades and starves the tacit floor where durable scarcity and unfilled demand actually sit. A mid-sized European economy — the Czech Republic is the useful home example — should therefore treat cognitive altitude not as a personality trait to recruit but as a capability to be built, owned and defended, with two halves: an agentic engine that runs the commoditising meta-work at scale, and a deliberately protected tacit base that the engine cannot hollow out.

The agentic engine is ENSI’s signature layer, and each move on the ladder names the agents that now run it. Scanning and early-warning agents execute the codifiable half of navigation and foresight — continuously reading fifty jurisdictions, a thousand trials and the live signal-stream, so the human is handed the reference class rather than asked to conjure it. Arbitrage and analogy agents do the search-and-retrieval half of cross-domain transfer, surfacing the solved problem in another field and leaving the human to judge whether the deep structure holds. Simulation and red-team agents extend lookahead to the physical wall and no further, war-gaming second-order effects and adversarially attacking a plan before it ships. Persuasion-defence agents meet the offense-dominance of mind-modelling head-on, flagging manipulation and fabricated evidence — the governance response to a world where a model out-persuades a human from basic data. Coordination and workflow agents externalise the eighth move directly, turning a hero-coordinator’s implicit knowledge into an auditable system a competent team can run. In every case the pattern is the ENSI pattern: the agent does the is — the scan, the retrieval, the simulation, the draft — and an accountable human owns the ought, the judgement, and the answer at the next election or board meeting.

The strategic payoff is option value. The state that builds this engine buys itself more available moves precisely as the cost of the meta-layer collapses — it fields foresight and coordination at machine scale while its rivals are still hiring for altitude as though it were scarce. And the state that pairs the engine with a defended tacit floor — the manufacturing capability, the skilled trades, the deep in-domain judgement that neither rivals nor agents can copy from a slide — captures the value the pure meta-players will keep leaking away. That is the whole thesis, restated as policy: altitude is worth climbing for, but only if it is fused to something that cannot be arbitraged, and only if the climbing is done by an owned engine rather than rented from whoever reaches the top first.

What to do first

Retire the flattering picture of intelligence as more thinking, and adopt the accurate one: intelligence buys altitude arbitrage, a real edge that is offense-dominant, IQ-orthogonal more often than we admit, and perishable the instant it diffuses. For an individual, the move is to climb for the leverage but fuse every meta-move to a ground floor that cannot be reframed, retrieved or copied away — to be the person who can both choose the problem and build the answer. For an institution, the move is to stop mistaking a bandwidth problem for a talent problem: build the agentic engine that runs the commoditising meta-work at scale, defend the tacit base the engine cannot hollow out, and measure the whole thing by outcomes — time-to-evidence before a decision, the share of choices carrying a simulation and an ex-post evaluation, the durable capability retained rather than the cleverness displayed. Intelligence’s real dividend was never a bigger engine. It was a better address — and the rent comes due the moment the neighbourhood fills up. The task is to own the building before it does.

Sources — a 259-document library compiled by the ENSI Foresight Division (July 2026), downloaded to sources/downloads/ with per-perspective manifests in sources/. Load-bearing evidence includes: METR on AI task-horizon doubling; Epoch AI and ONET automation-exposure work on Moravec’s paradox; Polanyi and Piore on tacit knowledge; the deep-tech “valley of death” (TRL/MRL) literature; Cowan on working-memory limits and Chase–Simon on chess chunking; Kahneman & Klein on conditions for valid intuition; Webb et al. and Kosinski on emergent analogical reasoning and theory-of-mind in LLMs; the AlphaFold Nobel; Dunbar and Kubricht on analogical transfer; Uzzi et al. (17.9M papers) on atypical combinations; Jones on the burden of knowledge; Tetlock’s Expert Political Judgment and the Good Judgment Project; the lookahead-pathology and atmospheric-predictability literatures; Bond & DePaulo on lie detection; Salvi et al. on GPT-4 persuasion; West & Stanovich on the bias blind spot; McLean & Pontiff and “not all factors crowd equally” on anomaly decay; Nordhaus on innovator surplus capture and Teece on complementary assets; Horowitz, the mirroring hypothesis, transactive-memory and Project Aristotle on coordination; Deming’s system principle; Kuhn, Boroditsky, Sokal and Casado on concept-creation and the marginal cost of creation; the brain-training/far-transfer null results; the printing-press growth study; AlphaEvolve and the Darwin-Gödel Machine on machine self-improvement. Full provenance and fetch status in the per-perspective manifests.*