In the 1830s and 1840s, the United States had railroads galore, but not a railroad system. Investors laid track at whatever gauge suited them, so the country filled up with islands of laid track — lines whose carriages could run only on their own rails and couple only with their own kind. Within each island, travel was miraculous. Between them, it was still 1810: at every break of gauge, passengers and freight were unloaded, carted across town, and reloaded onto a different company’s cars — transfers that mercilessly ate time and betrayed the locomotive’s promise: that through freight and long-distance travel would be forever transformed.

Circa 1850, the railroad had been invented. The railroad revolution had not yet happened. That gap — between a working machine and a transformed economy — is the subject of this essay, and, in a sense, of my forthcoming book, History’s Most Revolutionary Innovation (Cambridge University Press).

We remember industrial revolutions as parades of inventions: the engine, the dynamo, the chip. But invention turns out to be the cheapest part. What separates a gadget from a revolution is everything that happens afterward — commercialization, standardization, complementary investment — and, at every point where those processes stall, a visible hand that unsticks them. Industrial revolutions are made, not born. This essay explains what they are made of.

What an industrial revolution actually is

Strip away the romance and an industrial revolution is the commercialization and diffusion of a general purpose technology, or GPT — a technology that, unlike the zipper or the ballpoint pen, pervades the whole economy, keeps improving for decades, and spawns complementary innovations in every industry it touches. Steam was one; electricity and the internal combustion engine were the second wave; the microprocessor was the third. Each revolution was really a cluster — steam arrived with advanced metallurgy and mechanized spinning, electricity with synthetic chemicals, the microchip with software and telecommunications — but in each case one technology supplied the grammar the others spoke.

And in each case the revolution was as much the birth of supply chains as of machines. The railroad was an elaborate symphony of high-pressure engines, precision machine parts, and improved metallurgy for rail casting, all evolving simultaneously — and to bring those components into profitable relationship with one another, entirely new supply chains had to blossom. What gets commercialized in an industrial revolution is never just the invention; it is the web of intermediate goods, specialized suppliers, and contractual relationships that turns the invention into something you can buy.

For any candidate revolution, this essay asks three questions that appear in my forthcoming book. What is the core GPT, and what complementary technologies had to coevolve around it? What had to standardize — which interfaces, components, protocols, and measurements — before diffusion could accelerate? And what investments in infrastructure, skills, organization, and law were required to convert a working prototype into broad, durable productivity gains? Those three levers — coevolution, standardization, complementary investment — decide whether a technology transforms an economy or remains a frontier demo that struggles to scale.

Why the default is stagnation

The questions matter because the default outcome is stagnation, not revolution. Incumbent technologies fight back: wooden sailing ships and windmills kept improving for decades after steam arrived, staying competitive far longer than the textbooks suggest. And incumbency is often rational. American factories clung to the “group drive” — a single steam engine turning a forest of shafts and belts — for a generation after electric motors were plainly superior, because exploiting the motor’s real advantage meant physically redesigning the factory around it. Workers behave the same way: if most employers still run the old system, it is sensible to train for yesterday’s tools. A superior technology can languish indefinitely when switching costs are high, complements are missing, and everyone is waiting for everyone else to move first.

The organizational half of every diffusion story

Standardization, moreover, is necessary but not sufficient, because the second half of every diffusion story happens inside organizations. A GPT pays off only when firms rebuild themselves around it. Bessemer’s converters could mass-produce steel from the 1860s, yet early adopters kept botching the chemistry — controlling oxygen and impurities demanded expertise most mills lacked — and diffusion required organizational learning: standardized furnace designs, two-stage melting processes that reorganized the entire plant floor, heavy investment in training, and roaming technicians from Bessemer’s own licensing operation to teach clients the metallurgy and calibrate their machinery. Economists call what those mills were accumulating intangible capital — new processes, new skills, new structures — and it obeys a stubborn sequence: firms first use a new technology to substitute for the old, an electric motor bolted down where the steam engine sat, and only later reimagine the workflow around it. The gains live almost entirely in the reimagining.

The productivity paradox is a feature, not a bug

This is also why every GPT so far has produced a productivity paradox on arrival. Early adoption is expensive: firms divert resources from producing to learning, reorganizing, and retraining, so measured productivity stalls or even dips just as the hype peaks. Then, once standards congeal and organizations finish rewiring themselves, the gains arrive in a rush. American labor productivity crawled at about 1.4 percent a year from 1970 to 1994 — computers everywhere except in the statistics — then surged past 2 percent from 1994 to 2004 once firms had rebuilt their operations around the technology they had purchased two decades before. The paradox is not evidence that a technology has failed. It is what the middle of an S-curve feels like from the inside.

Radio households, 1920–1940: the fastest mass-market take-off in history — from zero to 83 percent of American homes in two decades, accelerating through the Depression because one purchase bought a free stream of entertainment. Author’s figure.

Telephones in service, 1880–1940: roughly six decades from Bell’s patent to broad household penetration — note the Depression dip. Networks that need per-node infrastructure — wires, switches, operators, city by city — diffuse on a slower clock than broadcast. Author’s figure.

Four revolutions, one grammar

First: steam

The First Industrial Revolution was about solving mundane problems: draining mines, moving coal, converting heat into motion inside a factory. Newcomen’s atmospheric engine of 1712 pumped water from flooding mineshafts; Watt’s refinements made the engine efficient enough to leave the pithead. What followed was a virtuous cycle: steam engines drained the mines, making coal cheaper; cheaper coal made steam affordable for everyone else. Freed from the geography of rivers, textile mills moved into cities and ran day and night; railroads, born as coal carts on wooden rails, stitched fragmented local markets into national ones. And diffusion accelerated only as designs standardized — the self-acting spinning mule reached small mills once it no longer required a resident mechanic to keep custom machinery alive, and the railroads became a network only as the gauges finally converged.

Second: electricity, engines, and chemistry

The Second Industrial Revolution — roughly 1870 to the 1930s — ran on electricity, the internal combustion engine, and industrial chemistry. Its signature wasn’t any single machine but the reorganization of production around them. Factories abandoned the central line shaft for unit-driven machines plugged into sockets — but only after motors, voltages, and transmission standardized, and after the war between direct and alternating current settled in AC’s favor.

The payoff was Fordism: interchangeable parts on moving assembly lines, churning out goods at prices that fell as fast as volumes rose. Between 1909 and 1924 the Model T barely changed, yet relentless process innovation cut its price from roughly $950 to $290. Managing production at that scale demanded an organizational invention as consequential as the technical ones — the vertically integrated, multidivisional corporation Alfred Chandler chronicled, with General Electric building an entire ecosystem from the dynamo to the toaster and running a corporate R&D lab to feed it. The era’s complementary investment in people was the high school movement, which supplied the literate, numerate workforce the new factories and offices required.

Registered passenger cars in the United States: 19 per 100 people by 1929. Annual production collapsed during the Depression, but the installed stock barely flinched — it is diffusion, not the production line, that transforms a country. Author’s figure.

Third: the microchip

The Third Industrial Revolution began, characteristically, with a bottleneck. Bell Labs invented the transistor in 1947; Texas Instruments put it into commercial silicon in 1954, with the military as its indispensable first customer — and then complex systems hit the “tyranny of numbers”: a missile guidance computer needed thousands of components hand-soldered together, and one bad joint among thousands ruined the machine. The integrated circuit solved the problem by printing transistors and their connections onto a single silicon substrate, and the planar process turned that trick into mass production. Its triumph arrived in 1971 with Intel’s 4004 — a programmable computer on a chip — and Moore’s Law did the rest, doubling transistor counts every couple of years for five decades.

But the industry transformed the economy only after its own standardization phase: the x86 instruction set and the Wintel stack decoupled hardware from software, so a program written once ran on millions of machines; TCP/IP, HTML, and HTTP did the same for networks.

The Third Revolution also produced an institutional innovation to rank with the M-form corporation: the venture-backed startup, seeded by the Fairchild diaspora, which industrialized the financing of long-shot commercialization.

Moore’s Law in economic terms: processor performance per inflation-adjusted dollar, 1971–2020, log scale. Every dollar bought orders of magnitude more computation, decade after decade. From the book’s online appendix, Figure S1.1.

Fourth: artificial intelligence

And the fourth? AI is not one invention but a bundle of complementary capabilities — language, perception, prediction, planning — riding scaling laws that play the role Moore’s Law played for silicon. The training compute behind frontier systems doubled roughly every six months after 2010, and the cost of a frontier training run climbed from an estimated $5 million for GPT-3 to reportedly more than $100 million for GPT-4 — which is why the center of gravity shifted from university laboratories, which could compete when state-of-the-art meant modest compute, to the handful of corporate labs that can field billion-dollar clusters. Training a frontier model is not a software project; it is an industrial undertaking, and it looks like one from the road. Behind it stretches one of the most vertically disintegrated and globally specialized supply chains ever assembled — architecture licensors, fabless designers, a single Dutch lithography firm, fabrication concentrated in Taiwanese foundries, hyperscaler clouds, model labs, and an application layer on top — every interface held together by the patents, licenses, and standards whose origins the next section traces.

In History’s Most Revolutionary Innovation I ask how one would know an industrial revolution was underway without the benefit of hindsight, and the early 2020s pass every test. Intensive capital buildout: just as coal pits, railway hubs, and generating stations swelled across the landscape in every prior installation phase, data centers are doing so now. Exponential adoption: Waymo’s robotaxis took until late 2023 to log their first million cumulative paid trips, quintupled that by the end of 2024, and passed twenty million by the end of 2025 — the classic knee of an S-curve. Exponential performance: on METR’s task-horizon metric — the length of task a frontier agent can complete autonomously half the time — capability has doubled roughly every seven months since 2019, a cadence faster than Moore’s. And multiple simultaneous applications: diagnostic imaging and treatment planning in medicine, predictive maintenance on factory floors, coding copilots, browser agents, discoveries of new protein structures, ten-day weather forecasts generated in under a minute.

Locate all this on the industrial revolution lifecycle, though, and something interesting appears: the fourth one is already deep in its standardization phase, and the first wave has moved with remarkable speed — a story I tell at length in the book. Common model formats began dissolving vendor lock-in at the foundation; shared benchmarks became the era’s consumer-protection standard; the “AI factory floor” of training and deployment pipelines standardized around common orchestration tools. Connecting models to the world, once a matter of bespoke glue code for every database and application, collapsed onto the Model Context Protocol — write once, connect to any compliant system. Even governance began to standardize: NIST’s AI Risk Management Framework and the EU AI Act’s presumption of conformity began turning compliance from a bespoke legal battle into a standardized engineering task. And in late 2025 came what may prove this era’s TCP/IP moment: rival laboratories concluded that open infrastructure beats proprietary control for agents, and OpenAI, Anthropic, and Block co-founded the Agentic AI Foundation — donating MCP, the AGENTS.md format, and an open agent framework to neutral stewardship — because agents that must discover one another’s capabilities, issue task orders, and verify completion need a shared messaging layer, a SWIFT for digital workers, that no single firm can impose.

What remains genuinely fluid is the deeper wave: evaluation regimes rigorous enough for hospitals and banks, liability rules for autonomous action, and the trust conventions of agent-to-agent commerce. Standardization, as this history keeps teaching, is never purely technical; it is always also political.

The binding constraint, then, is the other half of the recipe: organizational adaptation. Most firms remain in what the book calls pilot purgatory — using AI to substitute rather than transform, writing the same email faster instead of redesigning the workflow that generates the email — which is precisely the first wave of factory electrification replayed: motors bolted down where the steam engines sat, the architecture of production untouched, the gains unrealized.

Education is the cleanest example of what the collision looks like. Bloom’s famous “2 Sigma” finding — that one-on-one tutoring lifts students two standard deviations above the conventional classroom — has haunted educational psychology for forty years because no society could afford a tutor per child. AI promises exactly that for the price of electricity, yet rollouts remain stuck in pilots, snagged on incompatible data systems, privacy law, teacher skepticism, and unequal infrastructure: a technology waiting on organizations.

Even the driverless car, the poster child of exponential adoption, took fifteen years of organizational and institutional grind — fleets, insurance, maintenance networks, city-by-city regulatory settlements — to get from the DARPA Grand Challenge to a commercial service. Anyone frustrated that AI hasn’t yet shown up in the productivity statistics is describing the 1980s of every previous revolution.

The state as midwife

Here is the part the invention-parade version of history leaves out. At every stage of every one of these transformations, markets alone under-delivered — predictably, for reasons economists understand well — and the state stepped in, not as a planner, but as a midwife to a birth that was otherwise going badly.

Why markets under-deliver

Start with the deepest failure. Ideas are public goods: non-rival and hard to fence. William Nordhaus estimates that innovators capture only about 2.2 percent of the total surplus their innovations create and estimates of the social return to R&D run to 60 percent or more — several times the private return. Faced with such a meager slice, rational firms systematically underinvest in exactly the research that launches GPTs, leaving trillion-dollar bills on the sidewalk because no one can privatize the pickup. Standardization poses a second failure: a coordination trap. Everyone gains if the industry converges on one design, but no firm dares invest in a standard that might strand it — so markets can sit in fragmented, inferior equilibria indefinitely. The gauge wars were exactly that trap, and no railroad could escape it alone.

The underprovision of ideas: the wedge between private and social marginal cost is the spillover firms cannot capture, and the hatched triangle is the deadweight loss a well-designed subsidy eliminates. From the book’s online appendix, Figure S1.2.

Against these failures, governments have historically deployed a five-part toolkit.

Fund the science

First, they fund the science. Britain chartered the Royal Society in 1660 and built the culture of useful knowledge that preceded the first revolution; the American century ran the model at scale. The Army bankrolled ENIAC; ARPA funded the packet-switching research that became the internet; and federal science agencies — NSF, ONR, NIH — together with defense procurement built the ecosystem around solid-state physics, semiconductors, and computing, in which private laboratories like Bell Labs did much of the inventing while government supplied both the underlying science and the first markets.

Patents are supply-chain infrastructure

Second, they build property rights for ideas — and here it matters to see what patents actually do, because the textbook story sells them short. A patent converts a public good into something like a club good — excludable by law, still non-rival in fact — and in doing so creates a market for ideas: inventions that can be licensed, divided, collateralized, and sold.

But the deeper function is that patents are the legal infrastructure of new supply chains. Exclusion and disclosure work together to let strangers find each other: because the invention is published and protected, financiers, assemblers, distributors, and marketers can seek out the inventor and contract with her at arm’s length, each specializing in its comparative advantage — the inventor in inventing, everyone else in what they do cheapest. The licensing contract, underpinned by a legally enforceable patent, is what helps transfer the tacit know-how a technology needs to travel — the blueprints, the machinery, the onsite training — and, as licensees improve what they license, the knowledge flows back upstream.

The Cohen-Boyer patent on recombinant DNA shows the mechanism in miniature. Stanford and the University of California licensed the foundational gene-splicing technique broadly and non-exclusively; hundreds of startups and pharmaceutical firms built the modern biotechnology industry on it — and that horizontal diffusion called into existence a vertical one: a whole new market of specialized suppliers of enzymes, reagents, and laboratory equipment. One patent, licensed well, summoned a supply chain.

Crucially, the marketplace where such deals happen is itself grafted atop public infrastructure: examiners, registries, and specialized tribunals — above all the Court of Appeals for the Federal Circuit, created in 1982, which helped make patent rights predictable enough to contract on. Therefore, the government does not merely protect ideas; it bankrolls the market in which they trade. And the payoff came due in the late Third Revolution, when strong, tradable IP catalyzed the vertical disintegration of the electronics industry — fabless designers like Qualcomm and ARM specializing in high-value chip design while foundries an ocean away did the fabrication, an arrangement thinkable only because a design is an enforceable, licensable asset.

How intellectual property rights convert non-rival knowledge from a public good into something like a club good — excludable by law, non-rival in fact. From the book’s online appendix, Figure S1.3.

Finance the gap, break the deadlock

The state provides other support structures too.

It sometimes directly finances the valley of death between prototype and product. Parliament’s Board of Longitude paid John Harrison more than £23,000 across decades to perfect the marine chronometer that private investors dismissed; two and a half centuries later, the SBIR program wrote early checks to a struggling San Diego startup called Qualcomm.

It may also help break coordination deadlocks — sometimes as the customer of first resort (the military’s bulk transistor purchases and the Minuteman contracts gave the young chip industry guaranteed demand and forced a design to converge), sometimes by seeding an installed base (the Defense Department’s 1983 TCP/IP cutover and the NSF’s backbone gave the protocol running code and users that later mandates for rival OSI architectures could not dislodge), sometimes as a convener and measurement authority (NIST and the legal scaffolding around private standards bodies like IEEE and 3GPP, which produced Wi-Fi and 4G under FRAND licensing commitments), and sometimes by consortium — SEMATECH pooled fourteen rival chipmakers with federal money and, alongside trade policy and corporate restructuring, helped reverse a decade of market-share losses to Japan.

Build the complements

It also goes without saying that the state helps build complements to the technologies that power industrial revolutions in the form of basic infrastructure and educating people. Canals and harbors in the nineteenth century, the TVA’s rural electrification that crowded in decades of private industrial investment, the interstate highways, the internet backbone. And, above all, education: the Morrill land-grant colleges trained the engineers of the industrial age, the GI Bill and the Higher Education Act stocked the third revolution with graduates, and the research universities they built anchored every innovation cluster from Route 128 to Silicon Valley. Goldin and Katz called American history a race between education and technology; the state was how education kept pace.

Referee the split

Finally, there is one more role, subtler than the rest. Even when everyone agrees a common standard is worth having, the parties fight over which standard — because the choice decides who captures the surplus. Game theorists call this a Battle of the Sexes, and it is endemic to technology: upstream patent holders and downstream implementers both want interoperable smartphones, but they want opposite royalty structures. Left alone, the strong impose rules that favor themselves, or cooperation collapses. The job here — shared among public law, antitrust policy, courts, and the private standards bodies that operate in their shadow — is not to pick the winner but to design the game: revenue-sharing formulas, supermajority voting rules, dispute forums, terms like FRAND, so that even the weaker party can commit to cooperating. Done well, this converts zero-sum brawls into positive-sum settlements. It is the least glamorous thing a government does for innovation, and among the most important.

Not a brief for dirigisme

None of this condones dirigisme. The record I’ve sketched is not one of states picking winners or planning economies; where they tried, they mostly failed. It is a record of states solving specific, well-defined market failures — public goods, coordination traps, missing credit, distributional standoffs — that private actors demonstrably could not solve alone. That is the version of industrial policy the American postwar system institutionalized, and the version my book argues built the legal and financial scaffolding on which the AI economy now stands.

The fourth revolution’s to-do list

Run the checklist against the still in progress and still in doubt AI Revolution and a familiar agenda, albeit with new characteristics, emerges.

The science: reliability, reasoning, alignment, and evaluation are open problems whose returns no single firm can fully capture — exactly the class of research the public purse exists to fund.

The standards: the first wave — model formats, benchmarks, MCP, the Agentic AI Foundation — congealed fast, but the deeper wave of evaluation regimes, liability rules, data rights, and agent-commerce trust is today’s ungauged track, and whether it congeals through open consensus bodies or proprietary lock-in will shape the industry’s structure for a generation.

The infrastructure: the buildout this time is measured in gigawatts, and permitting an interconnection queue is the modern equivalent of granting a railroad right-of-way.

The organizations: firms escape pilot purgatory only by accumulating intangible capital — redesigned workflows, reskilled people, new structures — and that accumulation, not model capability, now sets the pace of the revolution.

The people: every previous revolution paid off only after a generation-scale investment in complementary skills, and there is no reason to believe the fourth has repealed that law.

The machines, as ever, are the easy part. In the 1840s the locomotives worked beautifully; what America lacked was a gauge. The AI economy is laying its track right now — its protocols, its standards, its institutions, its people. Whether this becomes the fourth industrial revolution or a long archipelago of impressive, incompatible islands depends far less on the next model release than on whether, this time too, we do the unglamorous work of making the gauges match.

A strong version of the unitary executive theory — an idea that took hold among young Justice Department lawyers in the early 1980s — is therefore now the law of the land: Humphrey’s Executor, the 1935 precedent that let Congress insulate these agencies from presidential whim, has been swept aside — “if anything more is left of Humphrey’s, we overrule it,” wrote Chief Justice Roberts for the 6–3 majority, over a dissent Justice Sotomayor read aloud from the bench. The Federal Reserve, on idiosyncratic historical grounds, appears to remain outside the rule. For now, at least — the boundaries of that exception are anything but clear…

What does this decision do to the machinery that actually produces American technological dynamism?

Innovation is a long-horizon bet. Firms sink enormous, irreversible investments — fabrication plants, data centers, R&D pipelines, spectrum licenses, clinical trials — on the assumption that the regulatory treatment of those bets won’t swing wildly with the political winds. The whole purpose of insulated, professionalized agencies was to make regulation legible and stable enough to invest against.

In my forthcoming book, History’s Most Revolutionary Innovation (Cambridge University Press), I argue that the engine of American innovation over the last half-century was not “neoliberalism,” and not deregulation for deregulation’s sake — paradigms that, I contend, never really governed the United States. Instead, it was an exceptionally American experiment in policymaking that I christened the Creative Destruction Paradigm: an evidence-based, cost–benefit regulatory order, assembled incrementally from Carter through Obama, that lowered transaction costs, solved coordination failures, and managed distributional conflict in the service of technological change.

Neither laissez-faire nor state-directed industrial policy, the Creative Destruction Paradigm was a rule-bound device for underwriting American innovation. It depended on the shared expectation that the rules of the road are durable, evidence-driven, and applied predictably rather than opportunistically.

Within this context, Humphrey’s Executor mattered not because independence is sacred, but because insulating agency leadership helped preserve space for professional staff to run long-range studies, gather data, and engage in fact-finding without every technical conclusion being immediately subordinated to presidential preference. What mattered instead was neutral competence.

On top of that insulation sat procedural discipline: the Administrative Procedure Act’s notice-and-comment regime and its “arbitrary and capricious” standard; the requirement that agencies justify decisions with a reasoned explanation grounded in an evidentiary record; and, for executive agencies, the centralized cost–benefit review system that ran from Reagan’s Executive Order 12291 through Clinton’s 12866 at the Office of Information and Regulatory Affairs. Independent agencies were never fully inside that OIRA system, but they operated within the same broader technocratic culture of record-building, expertise, and judicial review.

For example, after the Telecommunications Act of 1996, it was the FCC — through a methodical series of market-based spectrum auctions — that allocated wireless frequencies to the carriers best positioned to use them, seeding the mobile networks on which the smartphone economy, and then the data economy, would run. That same predictability let a fabless upstart called Qualcomm bet its future on Code Division Multiple Access shows the same logic from another angle: years before the 1996 Act, the fabless upstart staked its future on a technically risky spread-spectrum standard that would become foundational to modern wireless — a long-horizon wager on spectrum policy, standards, and patents that only a predictable rule-bound ecosystem could reward. And once antitrust had settled into an evidence-based consumer-welfare standard that tolerated scale when the data showed consumers gained — and once Section 230 shielded platforms from ruinous liability for what their users posted — companies like Google, Facebook, and Amazon could build the multi-sided markets that turned billions of users into gushers of behavioral data. That flywheel proved a godsend for the AI developers who drew on it to train large language models like ChatGPT.

Figure 1. Auctioning the airwaves. Winning bids in six major FCC spectrum auctions, from AWS-1 in 2006 ($13.7 billion) to the record C-band auction of 2020–21 ($81.1 billion). Since Congress authorized competitive bidding in 1993, the auction program has replaced allocation by administrative fiat with allocation by market — assigning spectrum to the carriers that valued it most and, in the process, capitalizing the networks on which the mobile and data economies run. It is a quiet, technocratic program of exactly the kind that agency insulation made credible. Source: Federal Communications Commission auction results as reported at auction close (AWS-1: Auction 66; 700 MHz: Auction 73; AWS-3: Auction 97; 600 MHz broadcast incentive: Auctions 1001/1002; CBRS: Auction 105; C-band: Auction 107).

Or consider intellectual property, where the machinery of American innovation runs through a lattice of expert institutions — the USPTO inside the Department of Commerce, and independent or quasi-independent bodies like the FTC and the International Trade Commission. The IP revolution of the 1980s and 1990s was not an ideological project; it was built claim by claim on evidentiary records. The Bayh–Dole Act of 1980 passed because its sponsors marshaled data showing that fewer than five percent of some 28,000 government-held patents had ever been licensed, and argued that promising federally funded inventions — including biomedical discoveries — were left undeveloped because no firm would finance costly clinical trials without reliable exclusive rights. Giving universities clear title to commercialize federally funded research created a repeatable path from lab bench to market — the path that carried the Cohen–Boyer recombinant-DNA breakthrough out of Stanford into broad commercial licensing, helping seed the biotechnology industry, from firms like Genentech to products like recombinant insulin and new cancer therapies.

Figure 2. Unlocking the ivory tower. Annual U.S. patents issued to American universities, before and after Bayh–Dole. Universities received just 264 patents in 1979; patenting then roughly doubled between 1979 and 1984, doubled again by 1989, and doubled yet again by 1997, reaching about 3,380 by 1999 — the takeoff that turned research campuses into commercialization engines. Values after 1979 are approximate, reflecting the successive doublings documented in the source. Source: David C. Mowery, Richard R. Nelson, Bhaven N. Sampat & Arvids A. Ziedonis, “The Growth of Patenting and Licensing by U.S. Universities,” Research Policy 30(1), 2001; see also Mowery et al., Ivory Tower and Industrial Innovation (Stanford University Press, 2004).

The USPTO worked the same way. When software firms pressed for patent protection in the early 1990s, the agency did not simply take a side. It held public hearings in 1994, solicited formal comments in 1995, and weighed evidence presented in that record that software development had become a capital-intensive industrial process — with average project budgets surging past $10 million — before updating its examination guidelines in 1996.

The result was a patent system that could underwrite the software and semiconductor booms: total applications at the USPTO nearly tripled between 1980 and 2000, from roughly 112,000 to over 315,000.

Meanwhile, the nonpartisan ITC supplied the technical analyses — on tariffs, on import competition, on the economics of trade liberalization — that let successive administrations of both parties open markets for American technology on the strength of quantified evidence rather than lobbying muscle. Its studies informed major trade-policy debates from the Tokyo Round of GATT to NAFTA.

Figure 3. The evidence-based IP revolution. Total applications filed at the U.S. Patent and Trademark Office each year, 1963–2020. Applications were essentially flat for two decades — about 108,000 in 1979 — then began a forty-year climb after the reforms of the early CDP era (shaded): the Bayh–Dole Act (1980), the creation of the Court of Appeals for the Federal Circuit (1982), and, following public hearings and formal comment, the USPTO’s 1996 software examination guidelines. Applications nearly tripled between 1980 (112,379) and 2000 (315,015) and kept climbing, exceeding 600,000 a year by the 2010s. Source: U.S. Patent and Trademark Office, Patent Technology Monitoring Team, U.S. Patent Statistics Chart, Calendar Years 1963–2020.

We already know what happens when that machinery gets politicized, because we have run the experiment on a small scale. In the narrow but economically vital domain of standard-essential patents — the technologies at the heart of 4G, 5G, and Wi-Fi — policy has lurched with each administration: pro-patentholder guidance in 2019, withdrawal of that guidance by the Biden administration in 2022, and then a renewed move toward stronger patentholder remedies under the second Trump administration, which deployed the USPTO to file statements of interest in private litigation. Each swing shifted substantial licensing leverage between patent holders like Qualcomm and implementers like Apple. Trump v. Slaughter threatens to scale it to the entire regulatory state.

Notice what the lurching does. Distributional fights are inevitable in a modern economy — patent holders and implementers will always battle over how to split the surplus a technology standard creates, and each side will always lobby to weaken the other's rights. The Creative Destruction Paradigm's quiet achievement was to keep those fights inside a stable, evidence-based framework, so that arguing over shares of the pie never stopped the pie from growing. Once the rules themselves swing with each election, the calculus changes: firms rationally redirect resources from engineering to lobbying, from R&D to litigation, from contributing technologies to open standards toward hedging against the next reversal. The conflict over who gets what bleeds into how much there is to get at all. That is the deeper cost of politicized regulation — it converts positive-sum games into negative-sum scrambles.

To be sure, serious people defend the ruling as a restoration of democratic accountability, and the old administrative state did carry real pathologies — capture, overreach, the sheltering of incumbents — failures I document at length in my book. But accountability to the president is not the same thing as accountability to evidence.

The danger now is not that we get less regulation, or more. It is that we get regulation bent to political ends — an FCC leaned on to punish a broadcaster the president dislikes; an FTC or NLRB deployed to reward friendly firms and harass disfavored ones; the technical judgments of the NRC or CPSC overridden for reasons that have nothing to do with the underlying evidence. Trading rule-bound expertise for presidential discretion doesn’t lighten the regulatory hand so much as politicize it.

None of this is destiny. Litigation under existing statutes remains a potent accountability channel; states continue to fill federal vacuums; private ordering — standards bodies, audits, insurance markets — can backstop where public capacity falters. But the Creative Destruction Paradigm’s institutional foundations just got conspicuously thinner, and the burden now shifts to whoever holds the presidency to wield the newly concentrated power with the restraint that evidence-based governance used to enforce automatically. This seems unlikely.

Perhaps this phenomenon is not all that remarkable, though. Consider that in the dying days of the USSR, polls found a comparable share of Soviet citizens still preferred some form of socialism to American-style capitalism. That young Americans are now about as keen on socialism as the people who watched it collapse may say less about a coming revolution than about the enduring appeal of the idea.

But “socialism” — “democratic socialism,” even — is a slippery word. What most American politicians who claim the label mean by it bears little resemblance to how the system actually works in the countries like Sweden they hold up as models.

Don’t tell these self-styled democratic socialists and their supporters this dirty little secret though, for they are too busy celebrating the death of the post–Cold War center-left. Their jeremiad is that the coalition Bill Clinton assembled and Tony Blair exported — the “neoliberal” Third Way of free trade, balanced budgets, welfare reform, and deference to markets — has collapsed under the weight of its own broken promises and socialism is rushing into the vacuum. To hear America’s rising “socialists” tell it, the era of triangulation and coddling the rich is over and is making way for true progressivism only they can deliver.

This post explains how American democratic socialism circa 2026 is neither democratic nor socialist; the version of social democracy that Sanders and Ocasio-Cortez and Mamdani and Wilson claim they’re delivering is actually something much older and more distinctly American than Swedish. It’s the populist-statist regime that ran from the Progressives through Richard Nixon, the one stagflation discredited in the 1970s.

This post also takes issue with another claim that’s become ascendant, the idea that the Sanderistas are burying a neoliberal order. It turns out that this never really existed. What actually animated economic policymaking after the Cold War, as I argue in History’s Most Revolutionary Innovation, was a technocratic, evidence-based, pie-growing consensus that solved market failures, disciplined itself with cost-benefit analysis, and built the digital economy and the runway for AI. And we dismiss the real, lasting achievements of the moderate center — to which the center-left was pivotal — at our peril.

I address each of these myths—the neoliberal order that never was, the socialism that isn’t Scandinavian, and the democracy that isn’t very democratic—in reverse order.

Part I: There was never any neoliberalism to kill

Let’s start with the corpse that supposedly nobody is mourning. If “neoliberalism” means what its critics say it means — the government as a minimalist referee that protects property, enforces contracts, and otherwise steps back and trusts markets to allocate everything — then it never commanded a consensus in Washington, and it was never actually implemented. What ran the United States from Carter to the end of the Obama administration was not minimalist-referee neoliberalism. It was a technocratic, evidence-based, bipartisan paradigm I call the Creative Destruction Paradigm, and its whole purpose was to grow the pie by solving the market failures that markets, left alone, can’t solve themselves.

To be sure, on the surface, the post-1970s record reads like a neoliberal victory lap. Carter deregulated airlines, rail, and trucking. Reagan cut top marginal tax rates and loosened financial regulation. Washington privatized where it could, preached fiscal restraint at home, pushed austerity abroad, signed NAFTA, and built the WTO.

Figure 1. The deregulation reforms of the Carter–Reagan era cut real prices for consumers and slashed the cost of moving goods through the economy — gains that compounded for decades.

Take Figure 1, which reveals that after deregulation, inflation-adjusted rail freight rates fell by about 40 percent and airline fares collapsed too, albeit by less. The logistical cost of moving goods through the American economy dropped from around 16 percent of GDP toward 10 — the hidden plumbing that made just-in-time supply chains, e-commerce, and the data economy possible. A clear win for neo-liberalism.

But strip away the free-market vocabulary and look at the reforms that built the digital economy that has helped make Americans the richest inhabitants of the liberal democratic world, and you find a government creatively solving the market failures that were keeping private capital on the sidelines. The federal government spoke out of both sides of its mouth: it lionized fiscal restraint and deregulation while quietly spending money, time, and attention to act as the strategic conductor of an orchestra of universities, firms, agencies, and investors to build the digital economy together.

Consider that the first electronic computers were built to break enemy codes during World War II and, soon after, to run the calculations for the hydrogen bomb. Radar, the transistor, the laser, and the integrated circuit came out of a public-private system — Bell Labs, Fairchild, and university laboratories did the inventing, but federal research budgets, defense procurement, and guaranteed military demand underwrote the risk. Cold War air-defense systems pioneered the real-time computing and networking the internet would later inherit; the race to the moon pulled whole supply chains of microelectronics into being; and DARPA, the Pentagon’s blue-sky shop, funded the packet-switched network that became the internet.

Basic research is the textbook good a market underprovides — the payoff is enormous, diffuse, and impossible for any single firm to capture — so for half a century Washington supplied it, and the entire edifice of Silicon Valley was raised on that publicly funded foundation.

Having paid for the science, the government then built the machinery that let private firms own it. Property rights to ideas are not a fact of nature; they are constructed and enforced by the state, and Washington supplied the whole apparatus — the patent system, the examiners, and eventually a specialized appeals court built to make those patents stick — so that a discovery could be owned, licensed, and borrowed against rather than copied the moment it appeared.

Similarly, federal telecommunications policy built a market where none had previously existed: working with auction theorists, the FCC ran “simultaneous multiple-round” spectrum auctions that turned slices of the electromagnetic spectrum into clear, tradeable property rights, moving the “beachfront” airwaves from television broadcasters to the carriers who would build 4G. There was no market until the state designed one. And when Verizon won the 700 MHz “C-Block” in 2008, the FCC wrote openness into the deal as a condition of sale — the winner had to let any device and any app run on its network. That single auction rule is much of why your phone runs whatever software you want instead of a carrier-approved walled garden.

The same hand was at work all the way down. A “shot clock” kept local zoning boards from strangling cell-tower construction with endless delays. The Act forced incumbents to lease their legacy copper to competitors, then deliberately exempted new fiber — a credible promise not to make firms share the upside of their riskiest bets, which is exactly what freed the capital to build them. Around 150 new carriers entered within three years, connection prices fell by a fifth, the E-rate program wired 98 percent of American public schools to the internet by 2000, and the whole reform pulled something like $1.4 trillion of private investment into the network the data economy now runs on.

Technical standards are another example of the federal government making a difference. Every smartphone is an assembly of parts from rival firms, arrayed up and down a supply chain, that interlock only because they were built to common standards — IEEE for Wi-Fi, 3GPP for the cellular air interface, the USB consortium for the port. But writing those standards means sitting fierce competitors down at the same table to agree on a shared design, which is very close to the one thing antitrust law exists to forbid, because it looks like collusion under the Sherman Act. Left to the market, the process would have seized up: firms afraid of being sued for cooperating, afraid of backing the standard that loses out, afraid that a rival would conceal an essential patent and spring it once everyone was locked in.

So, Washington, D.C. built the table and policed it. Congress carved out the antitrust safe harbors — the National Cooperative Research Act in 1984, extended to standards bodies two decades later — so that sitting down with your competitors to write a standard wouldn’t land you in court. NIST, the federal standards agency, played neutral broker, supplying the metrology, the performance benchmarks, and the testbeds where rival firms could prove out their technology in a pre-competitive zone, defusing the paralysis that comes from fear of betting on a dead end. And when a firm tried to cheat the process — as Dell did, certifying it held no relevant patents during a standards fight, then trying to enforce a hidden one after the industry had committed — the FTC stepped in and turned the standards bodies’ private disclosure rules into enforceable public obligations. The reward was a smartphone that stitches together dozens of standards on fair and reasonable terms — and a Fourth Industrial Revolution built on top of it. None of that is refereeing. It’s engineering.

Part II: The Scandinavian mirage

Mamdani and Wilson and the broader Democratic Socialists of America point, as Bernie Sanders pointed before them, to Scandinavia. We just want what Denmark has. We want what Sweden has. It’s a seductive picture: a humane, prosperous, low-inequality society that proves you can tax the rich, fund a lavish welfare state, and still thrive.

The picture is almost exactly backwards. Sweden isn’t a refutation of capitalism. It’s one of the most capitalist countries on earth, with a big welfare state bolted onto the back end.

Look at what Sweden actually does. It has one of the highest rates of billionaires per capita in the world — Forbes counted 43 in 2024, roughly four per million people, double the American rate. It allocates goods and services through prices, protects strong property rights, and lets its global champions — IKEA, Spotify, Ericsson — earn fat profits. It’s not a command economy: private enterprise dominates, though the state still owns a few large firms outright, like the miner LKAB and the utility Vattenfall. When its banks collapsed in the early 1990s, Sweden did rescue them — but it did it the hard way, guaranteeing the system while wiping out shareholders and forcing the losses onto the owners. Sweden made shareholders eat the losses as the price of rescue — closer to the opposite of the 2008 American reflex, which propped up institutions and largely spared their owners. It let Saab Automobile go bankrupt, and Volvo Cars ended up Chinese-owned, bought by Geely. And for decades it welcomed foreign capital with almost no strings, screening sensitive deals only since 2023.

The tax system is where the left’s image of Sweden really falls apart. Sweden has no wealth tax — a Social Democratic-led government abolished the inheritance and gift taxes in 2004–05, and a center-right government abolished the wealth tax in 2007, after watching its rich, including IKEA’s founder, decamp for lower-tax countries. Its corporate rate is 20.6 percent, cut from 28 over fifteen years by governments of the left and the right. And it pays for its welfare state not by squeezing the rich but by taxing everyone — a 25 percent value-added tax and heavy levies on ordinary wages.

The numbers are the opposite of what Americans assume. The total tax burden on an average worker — the “tax wedge” — is higher in every Scandinavian country than in the United States, and in Sweden it’s dramatically higher.

Figure 2. The Nordic welfare state is financed by broad taxes on ordinary wages and consumption, not by levies on a narrow top. The average Swedish worker hands over 41.5% of the cost of their labor to the state, versus 30.1% for the average American. The U.S. is also the only OECD member without a VAT.

So if the Nordics tax their middle class harder than we do, where does their famous equality come from? Two places, and neither is the one American socialists imagine. The first is that their market incomes — what people earn before a cent of tax or transfer — are far more equal than ours to begin with. The second is that taxes and transfers then redistribute somewhat more on top. Mogstad, Salvanes and Torsvik put numbers on it: in the Nordics, taxes and transfers cut the Gini coefficient by about 12 points, versus 8 to 10 in the United States. But in their decomposition, that back-end redistribution explains only about a third of the gap between us. The bigger share is that the Nordic market distribution is already about as equal as American income is after all our taxes and transfers.

Here is where the American left thinks it has found its mandate — and misreads it completely. Yes, that more-equal market distribution comes from compressing wages before taxes, through coordinated, economy-wide bargaining and powerful unions. The socialists hear “predistribution” and “strong unions” and assume Scandinavia is vindicating their program. But look at the constraint the Nordic system accepts. Wages are set by sectoral bargaining in which the export sector leads — the wage norm is whatever firms like Volvo and Ericsson can pay and still win business against Germany and South Korea. The unions hold the line on raises that would price their own employers out of world markets, because they know a wage the traded sector can’t bear is a wage that destroys their members’ jobs. The compression is real, but it is tethered to productivity and to global competitiveness. It grows the pie even as it shares it more evenly.

And the Nordic worker accepts that discipline because the bargain comes with generous social insurance: a world-class system of education, training, health care, and income support — all of it financed by the thriving capitalist economy that the wage restraint keeps competitive. That is the deal. Wage moderation and openness to creative destruction on one side; security and opportunity on the other; growth underwriting both.

The American socialists want the equality without the discipline that produces it — pay floors set by political demand rather than by what the traded sector can bear, rent freezes and price caps that override the market by fiat, none of it tethered to whether the pie is still growing.

When the Swedish left pushed hardest in exactly this direction — the 1980s wage-earner funds meant to hand corporate ownership to union-controlled funds (the very scheme Bernie Sanders would propose importing to America in 2020), make-work public jobs, public spending swelling toward 70 percent of GDP — it triggered capital flight and a furious business backlash and when a financial crisis hit in the early 1990s and ushered in stagflation that immiserated rich and poor Swedes alike, Stockholm dismantled this socialist experiment and turned decisively toward markets. The lesson Sweden learned the hard way is the one the Democratic Socialists of America (DSA) hasn’t: redistribution works only when capitalism produces a big enough pie to spread around more equally.

Indeed, consider that the Nordics are as productive per hour worked as the United States. Denmark and Norway produce more GDP per hour than America — Norway’s number is flattered by petroleum — and Sweden is essentially tied. The reason American GDP per capita runs higher: Americans work more hours, over 200 more a year, not that we’re more productive.

In short, the Scandinavian model doesn’t trade prosperity for equality. It gets both, on the back of competitive, innovative, productive markets. The Nordics have no statutory minimum wage — wage floors come from bargaining, not legislation. Denmark’s celebrated “flexicurity” model gives employers unusual freedom to hire and dismiss by European standards, paired with generous income support and retraining. And they invest in innovation at a level few countries on earth match.

Figure 3. Sweden out-invests the United States in R&D as a share of GDP, and Finland and Denmark sit above the OECD average too. Among OECD economies with available data, only Israel and South Korea spend more than Sweden. The “socialist” utopias are innovation machines.

So, the American democratic socialists aren’t asking for Sweden. They’re asking for Nordic outcomes through anti-Nordic means. They want the safety net without the growth engine that pays for it, the equality without the competitive markets and coordinated wage-setting that produce most of it, and the generous transfers without the broad, regressive tax base that finances them. Hand the DSA Sweden’s actual policy toolkit — abolish the wealth tax, scrap the estate tax, cut the corporate rate, install a 25 percent national sales tax, ditch the minimum wage, let iconic firms fail and be sold to foreign buyers — and they’d revolt against nearly every line of it!

Part III: The thing that’s actually rising

If it isn’t Scandinavia, what is it that America’s ascendant social democrats want? It’s a populist-statist paradigm — the regime the Creative Destruction Paradigm replaced, the one that ran from the Progressives through Richard Nixon and that stagflation discredited in the 1970s.

Populist statism runs on a recognizable toolbox, and it isn’t, despite the rhetoric, mainly about correcting market failures. It’s about cementing political coalitions through rent creation and redistribution — trading long-run growth for immediate, targeted, excludable gains to favored constituencies. The devices recur across a century: price-distorting controls on food, fuel, housing, and utilities; credit and financial repression that steers cheap capital to the connected; public hiring as patronage; make-work and white-elephant spending; and trade protection for politically useful incumbents.

The signature move is rhetorical — attack “elites” and “the corrupt establishment” on behalf of “the little guy” — even as the policies disadvantage diffuse entrepreneurs, workers, and consumers in favor of concentrated, organized interests.

To be sure, these episodes differed enormously in purpose and context — Progressives, New Dealers, wartime mobilizers, Great Society liberals, and Nixonian price controllers were not one coalition or one ideology. But they reached again and again for the same tools. The Progressives gave us railroad rate controls and occupational-licensing regimes that fenced out competitors in the name of protecting consumers. Wilson imposed wartime price controls. FDR’s New Deal entrenched the whole approach: the National Industrial Recovery Act’s “codes of fair competition,” the Agricultural Adjustment Act’s price floors, the Wagner Act, Regulation Q’s caps on deposit rates, and the WPA and CCC’s public employment on an unprecedented scale. Then a second wave: LBJ’s minimum-wage expansion and managed farm pricing, and Nixon’s literal 90-day freeze on all wages and prices — before the model collapsed into the stagflation that discredited it.

Now set Mamdani’s and Wilson’s platforms beside that list. Mamdani’s rent freeze: a price control on housing. Government-run grocery stores — floated in both New York and Seattle: public provision aimed at capping consumer prices. Social housing financed by new corporate and business taxes, and pledges to raise the tax burden on the wealthiest residents: redistribution through the tax side and the public balance sheet. And the framing — the working class against a rigged system, affordable cities against the billionaires — is the populist moral binary almost verbatim. It’s the FDR-to-Nixon program, rebranded for a generation that has never seen it fail.

Ironically, and because perhaps history indeed always rhymes but does not perfectly repeat itself, the same attempt to return to past failures is running on the right. When Donald Trump floats a cap on credit-card interest rates and most-favored-nation drug prices, imposes sweeping tariffs, asks for the federal government to take equity stakes in everything from steel to semiconductors, and pulls the independent agencies under tighter presidential and OIRA control, that’s the same instinct from the opposite flank: direct political control over prices, trade, capital, and administrative discretion.

Part IV: The means defeat the ends

None of this makes today’s anger on the left, or right, for that matter, fake or the pain imaginary. The affordability crunch that elected these candidates is real — housing, healthcare, childcare, and energy is indeed more expensive than just a few years ago and these costs are hitting the poor harder than the rich. A movement doesn’t win the closest Seattle race in 119 years on a fantasy.

But a real grievance can have a wrong diagnosis, and the populist toolbox doesn’t just fail to deliver — it tends to produce the opposite of what it promises. The cleanest evidence comes from the left’s flagship policy, rent control. When economists Rebecca Diamond, Tim McQuade and Franklin Qian studied San Francisco’s 1994 expansion of rent control, they found exactly the boomerang basic economics predicts.

Landlords newly subject to rent control cut their rental housing supply by 15 percent — converting units to condos, redeveloping, or selling to owner-occupants. That lost supply drove a 5.1 percent rise in rents city-wide. Rent control protected the specific tenants who already held controlled leases, but it raised rents for everyone else and shrank the housing stock — “ultimately undermining the goals of the law,” in the authors’ words.

The pattern generalizes. Price controls often produce shortages and black markets. Minimum-wage floors can price out the most vulnerable workers when set too high — though the historical record is genuinely mixed, and recent work on the 1960s expansions finds large wage gains, especially for Black workers, without big job losses. Tariffs typically raise the consumer prices they claim to tame. Make-work and patronage crowd out the private investment and R&D that future prosperity runs on.

What the populist program reliably does is entrench incumbents, scramble the signals that allocate capital and labor, and choke the dynamic competition that grows the pie — which is why, in the end, it tends to betray the very “little guy” it was sold to protect.

Part V: Conclusion

We’ve run this experiment before, and we know how it ends. The American version became a byword for inflation, shortages, and stagnation, and it was discredited in the stagflation of the 1970s. The Swedish version blew up in the crisis of the early 1990s — and Sweden answered not by doubling down on “socialism” but with fiscal, monetary, and market reforms, then kept turning toward markets for decades after: an independent central bank, then the abolished wealth tax, then the lower corporate rate. They rebuilt the pie-growing machine that allowed them to continue to finance a generous safety net and productivity enhancing social insurance.

So, the choice in front of us isn’t “neoliberalism versus socialism.” That framing is a category error built on a regime that never existed. The real choice is between a pie-growing consensus — the technocratic, evidence-based, innovation-first order that actually delivered, and that genuine Scandinavian social democracy embodies — and a pie-fighting populism rising on both the left and the right, dressed in new clothes, reaching for old and discredited tools.

The democratic socialists believe they’re the future. On the economics, they’re the past. And the affordability crisis they’re right to be angry about deserves a better answer than the one that already failed.

While vivid and definitely memable, the K-shaped narrative us ultimately a claim about trajectories — about who is pulling ahead and who is falling behind over time — not about the familiar fact that the rich have more than the poor at any given moment.

And it turns out that separating fact from meme is difficult when you go beyond simple snapshots of inequality that tell you nothing about direction. The New York Fed finds a K-shaped pattern in wealth since 2023. The Atlanta Fed, using a payments survey, finds a “bifurcated recovery” in which both high- and low-income spending grew, just at different rates. The Minneapolis Fed, reviewing the available series, concludes the data are messier than the headlines suggest.

What explains why these sources are reaching different verdicts? Who’s right and why?

Each picks a starting point and faithfully reports an accurate answer — but it turns out the answer depends almost entirely on the starting point. The New York Fed anchors its window at the first quarter of 2023. Moody’s runs 2020 to 2025. The Atlanta Fed, 2021 to 2025. Figure 1 shows that if you choose an earlier starting point, one before Covid, it makes a world of difference.

Figure 1 — “It depends when you start the clock.” Top-1%-minus-bottom-50% real net-worth growth, by the start date of the measurement window, all ending 2026 Q1. Start before the pandemic and the bottom 50% beat the top 1% by 50 points (anti-K). Restart anywhere after 2022 and the top pulls ahead (weak K). The sources I drew on to make these bar graphs are Federal Reserve, Distributional Financial Accounts (net‑worth levels); specifically, the Net Worth Held by the Top 1% — the DFA series (WFRBLT01026, via FRED) for the top arm of the gap; Net Worth Held by the Bottom 50% — the DFA series (WFRBLB50107, via FRED) for the bottom arm; and CPI‑U (BLS) — the price index used to convert nominal net worth to real before computing growth rates.

The pandemic transfers of 2020 and 2021 are the reason for this reversal, and, in the rest of this post, I argue that the so-called K-shaped economy is largely an artifact of where that episode falls relative to your start date — a pattern that gets clearer still once we look past wealth to income and consumption.

First and foremost, consider that a real, durable K-shaped trajectory should tend to push income, wealth, and spending in the same direction and keep them there. If that’s what folks mean by a K-shaped economy, then that is broadly what the data should show if they’re right.

Conversely, consider that a transient transfer episode — as happened during the pandemic, when the federal government flooded households with stimulus checks, expanded unemployment benefits, and an expanded child tax credit — should instead hit “the economy” in three different ways. It will show up loud and temporary in after-tax income, because transfers are income. It will also show up as a one-time step in wealth, because the transfers that get saved, and the assets they’re parked in, leave a lasting mark on a balance sheet. And it will show up barely at all in consumption, because the money was largely saved rather than spent.

To adjudicate between these competing views — the K-shaped structural break in the economy versus the one-off hump caused by a transient transfer — I look across income, wealth, and spending to see whether groups that were already ahead grew faster, in real terms, than groups that were behind. A strong K means the top rises while the bottom falls. A weak K means the top rises faster while the bottom still rises. Anything else — the bottom rising faster, or no clear gradient from bottom to top — counts against the claim. Moreover, when independent measures of “the economy” point the same way, that is strong evidence of something real and durable; when they point different ways, that is the signature of something transient and local. A real K should show up across wealth, income, and spending, and it should persist.

What about spending and inflation?

Don’t people usually mean something specific by “K-shaped economy” — and isn’t it spending, not net worth? They probably mean a split in spending: the affluent still dining out and traveling while everyone else trades down to discount stores — the “bifurcated consumer” that turns up on retail earnings calls. Bundled with it is a claim about inflation: that the bottom faces a steeper cost of living, with rent, groceries, and insurance swallowing whatever raises they get. I should be straight about how much of that bundle this post settles, because I don’t want it to read as a bait-and-switch.

While measuring real income that we that I do already nets out the average bite of inflation, what a common price index such as the one I use doesn’t capture is whether the bottom faced a higher inflation rate than the top, and the evidence elsewhere — the New York Fed among others — suggests it did, modestly, since late 2022. If the bottom’s true cost of living rose faster, then deflating its income by an average index overstates its real gains, making it easier to reject the hypothesis that we’re living in a K-shaped economy. Therefore, using the BLS's estimate that the lowest-income households' inflation runs about 0.4 percentage points a year above the highest-income households', I reran the analyses that follow by deflating the bottom’s income with that higher rate (and, as an upper bound, with a deliberately generous full point a year).

While that choice indeed makes the bottom look worse and a K-shaped economy easier to find, the contrarian finding continues to strongly hold: Over the full window the bottom 50 percent still vastly outgrows the top 1 percent in real net worth; the pre-pandemic "anti-K" gap of about fifty points would take some eight points of extra bottom-specific inflation per year — roughly twenty times the BLS estimate — to disappear. The income hump still round-trips to its 2019 level by 2022, and real spending still peaks in the middle of the distribution, not the top. Tilting the deflator toward the K sharpens only the modest, already-conceded post-2022 wealth divergence; it does not manufacture the strong, economy-wide K the headlines describe.

In what follows below, I report only the results obtained when using the average rate of inflation because it is the more reliable and transparent measure. The all-items CPI rests on a far larger price sample and a published, replicable methodology, whereas group-specific inflation rates have to be imputed from smaller samples and assumptions about each income group's spending basket — they are estimates built on top of estimates. The headline numbers should rest on the firmer measure; I use the group-specific rates only to stress-test the conclusion, which they leave standing.

On consumption, what I find again complicates the popular story that we’re living in a K-shaped economy. Real spending by income group, which I get to below, doesn’t sort into a clean K at all — and the available consumption measures themselves disagree, ranging from a steep K to no K-shape. The bottom line is that the bottom’s real expenditures didn’t collapse, and the fastest growth sits in the middle and uppermiddle of the income distribution, not the top.

This post is silent about the composition story: the idea that consumption has bifurcated into rich consumers splurging on luxury goods and European vacations while poorer consumers trade down, category by category, to afford the basics within a roughly flat budget. But “trading down” is a claim about how a given dollar gets spent; “the economy is K-shaped” is a claim about whose dollars are growing and whose are shrinking. The trading-down story only adds up to a K-shaped economy if those underlying trajectories are actually diverging — and that is exactly what turns out to be fragile.

Figure 2 — The after-tax hump. Cumulative real growth in average income after transfers and taxes, by income group, 2019–2022 (2019 = 0). Source: CBO.

Income: the hump

Let's start with income, where the transfer story is clearest: the bottom quintile's after-tax income shot up in 2020–21, only to give almost all of it back by 2022. The poorest fifth of households' real income after transfers and taxes jumped about 17 percent above its 2019 level by 2021, then dropped to just +1.6 percent in 2022 as the pandemic's recovery rebates, expanded unemployment benefits, and one-year expanded child tax credit expired. In dollars, the bottom fifth went from about $44,000 in 2019, up to $52,000 in 2021, and back to $45,000 in 2022. The middle three quintiles traced the same arch, only shallower — up to about 10 percent at the peak and back to roughly 2 percent by 2022. The only group still clearly elevated in 2022 was the top fifth of the income distribution, at +8.2 percent — the receding tail of a 2021 capital-gains spike that hadn't yet fully normalized.

Figure 3 — What the transfers added. The lowest quintile’s after-tax income growth, actual versus CBO’s counterfactual that strips out the recovery rebates, expanded unemployment compensation, and expanded child tax credit. The shaded gap is the transfers. Source: CBO.

As Figure 3 suggests, if you strip the pandemic policies out, then the lowest income quintile doesn’t surge at all — it falls in the 2020 recession, to about −4 percent, exactly as you’d expect when a downturn hits the people with the least financial cushion, then climbs slowly back. By 2022, the two paths converge.

If we pan the camera further out, beginning the data series in 1979, the point is starker still. Figure 4 shows that, against four decades of slow, grinding gains for the bottom fifth, the 2020–21 spike is a single jag that snaps right back to trend.

Figure 4 — Blip, not break. The lowest quintile’s after-tax income, 1979–2022 (2019 = 0). Decades of gradual climb, a sharp pandemic spike, and a snap back to trend by 2022. Source: CBO.

Survey data point the same way. In the Census Bureau's money-income series — the measure that most undercounts top incomes, and so most flatters the bottom's relative position — the lowest fifth's real income was essentially flat from 2019 to 2023, the top fifth's was flat too, and standard inequality measures barely moved. The most bottom-friendly lens available still doesn't show a widening K.

Wealth: the level shift

Wealth is where the anti-K case looks strongest — and where it’s easiest to overread. Over the full post-pandemic window, the bottom 50 percent’s real net worth rose about 80 percent, against just 30 percent for the top 1 percent. Taken at face value, that’s not a K at all; it’s the opposite. But the number is a trap, and three things show why.

First, it’s a step, not a trend. In nominal dollars, nearly three-quarters of the entire 2019–2025 gain — roughly $1.8 trillion of $2.4 trillion — was banked by the first quarter of 2022, and in real terms the bottom half has added almost nothing since. This isn’t a hump that reverts like income; it’s a one-time jump that sticks and then plateaus. The bottom half isn’t still gaining ground.

Second, the K story imagines two separate economies — the rich riding rising markets, the bottom sinking under housing costs. The balance sheets say otherwise: one asset-price rebound off the 2020 trough, lifting both, just scaled differently. The bottom half got richer the same way the top did — through financial-asset appreciation, the same lever the New York Fed finds doing the work at the top, where it's markets and not housing. Of the roughly $1.8 trillion banked by early 2022, financial assets account for the largest slice, about $0.78 trillion: deposits where saved stimulus landed, plus retirement and pension balances rising with the market. Home equity accounts for slightly less, about $0.77 trillion, and the rest is durable goods net of other debt. Alas, the bottom half wasn’t deleveraging through it all — mortgage debt rose by close to $0.8 trillion; home equity grew only because prices outran the new debt.

Figure 5 — What the bottom half’s wealth gain is actually made of. Change in the bottom 50%’s net worth (nominal dollars), split into home equity, financial assets (deposits, pensions, market balances), and durables net of other debt, for the pandemic window and since. Source: author’s calculations, Fed Distributional Financial Accounts.

Third, what gain there is sits on a depressed, leveraged base. The bottom half’s net worth in 2019 was still climbing out of the wreckage of 2008, and it carries a 60-to-69 percent loan-to-value ratio on its real estate — so small house-price moves swing its thin slice of equity around, and percentage gains off that small base look enormous. An 80 percent rise on almost nothing is still almost nothing; it doesn’t mean the bottom half has caught up to anyone.

Taken together, the wealth data cut against the K in both directions — the bottom half gained ground rather than losing it, but it isn't "winning" either; it banked a one-time windfall. The only thing here that looks K-shaped is the slow drift after 2022, and that shows up only if you start the clock in 2023, which is where the New York Fed starts theirs. Again, where you start the clock decides whether you see a K at all — and start in 2019 and there's no K to squint at: the bottom half outgrew the top, and the shape inverts.

Consumption: No K in sight!

When it comes to spending, the case for a K-shape collapses entirely. Real expenditures by income group from 2019 to 2024 barely move, and they don’t line up into a K at all. A clean K should climb step by step from bottom to top. This doesn’t: the lowest quintile is slightly down; the top is slightly up; but the fastest real growth is in the middle and upper-middle of the distribution, not at the top. Even the Atlanta Fed’s payments survey, which does find higher earners pulling ahead, has both income groups’ spending growing — a bifurcated recovery, not a collapse at the bottom.

The consumption story is the mirror image of the financial-asset build-up in wealth I just discussed: much of the transfer money went into deposits and debt repair, not spending. If a transfer windfall had been spent, you’d expect the bottom’s consumption to spike alongside its income in 2020–21. But surveys of how the 2020–21 checks were used found households spent only about 40 percent of them, saving the rest or paying down debt — so most of the windfall never reached the checkout line.

The Transfer Hypothesis is Vindicated, not the K-economy shibboleth

Put the three pieces of evidence together and consider the larger pattern that emerges — an idiosyncratic pandemic compression episode that’s fading. Income: a loud, temporary hump. Wealth: a one-time step that holds. Consumption: near silence. Whereas a real K economy should show up across all three outcomes and persist, this one shows up in one outcome, wealth, in one window, after 2022, in its weakest form.

This is also why, while pedigreed sources disagree, none of them is wrong per se. The New York Fed, anchored in 2023, sees the weak post-transfer K in wealth and spending. The Atlanta Fed sees a bifurcated recovery in spending where both groups grew. The Minneapolis Fed says the data are messier than the headlines. CBO documents the income hump and its reversal. Each of them saw one panel of a three-panel picture.

The only thing I’ve done is hold the three panels side by side and notice that, together, they rule out the structural-break reading — and that the confident phrase “the economy is K-shaped” quietly generalizes the single narrowest, most recent, most asset-specific result to the whole economy and the whole period.

The strongest case for the K-shaped economy

Since 2022, if we confine the analysis to wealth, the top of the distribution has grown somewhat faster than the bottom half. That’s a weak K — asset-driven, real, but modest, and it does not involve the bottom falling in real terms. Moreover, the “anti-K” signals that make the bottom look like it’s winning are the temporary 2020–21 compression, now fading. So, yes, after the pandemic transfers faded, you can squint and you’ll kind of see a K-shaped economy. But squinting is bad for your eyesight!

Why it matters

A few things follow. The first is a caution about reading the economy through a vivid letter. “K-shaped” can be confirmed or debunked at will by choosing a start date, which means it’s carrying more rhetorical weight than the data support at the level of generality people use it. And it cuts against both confident camps at once. Against “the economy is K-shaped and the bottom is being crushed”: it’s a weak, recent, asset-side pattern, and the bottom’s real income and net worth both sit above where they were in 2019. And against “the K is a myth, the bottom is actually winning”: that was the transfers, and they’re gone.

The second is about policy, where the clearest lesson is in the income hump. The years 2020–21 are a natural experiment showing that transfer policy can compress the after-tax distribution substantially, and in real time, and that the post-2022 “re-widening” is in large part the mechanical expiration of that deliberate compression — not an autonomous market force pulling the economy into a K. That reframes the question from “how do we stop the K” to something more honest: the tools that flattened the distribution demonstrably worked, and they were allowed to lapse by design.

The outstanding question is whether to make any of them permanent. That’s above my paygrade — or at least beyond the scope of this post.

Beyond that, since the only durable divergence between the rich and poor since 2019 is in the trajectory of their wealth appreciation and is essentially asset-driven, the policy lever that matches the diagnosis is twofold: housing supply, and the rent and interest-rate exposure of the bottom half’s leveraged, real-estate-heavy balance sheets — not broad income or consumption programs aimed at a K that mostly isn’t there in those domains.

None of this makes the affordability pressure that ordinary folks — meaning, the non-rich — feel imaginary. It says the pressure is better described domain by domain — with housing and asset access at the top of the list — than by a single letter that turns out to mean different things depending on when you start looking.

Sources and notes: Wealth figures are my own calculations from the Federal Reserve’s Distributional Financial Accounts, deflated by CPI-U. Spending figures are from the BLS Consumer Expenditure Survey. Income figures are from the Congressional Budget Office, “The Distribution of Household Income, 2022” (January 2026), Figures 3, 17, and 18, deflated by the PCE price index. The K-shaped commentary at the top draws on Moody’s Analytics, the Bank of America Institute, U.S. Bank, EY-Parthenon, and Navy Federal Credit Union; the skeptical and decompositional work draws on the New York Fed’s Liberty Street Economics — tracking and explaining the divide — the Federal Reserve Bank of Atlanta, and the Federal Reserve Bank of Minneapolis.

Introduction

Artificial intelligence: no technology has ever been asked to carry so many fears — not the splitting of the atom, the internal combustion engine, or even social media, which up till now was the undisputed champion of manufactured anxiety. We’ve been told that AI will steal our jobs and reintroduce feudalism. That it has an insatiable appetite for electricity and water and spells environmental calamity. That its mushrooming harms — the chatbot that deepens a teenager’s psychotic spiral; the rising flood of deepfakes; the dissolving line between what a real person wrote and what a machine produced — are just a taste of worse things to come. That on a not-too-distant day AI may pursue a banal command like “manufacture some more paperclips” with such maniacal fervor that it tiles the entire globe in paperclips. That AI is Frankenstein’s monster and soon after it awakens to the sound of its own “I am alive!” it will find us excruciatingly boring, at best ignoring us and, at worst, turning us into mulch to ring its datacenters.

These fears summon a first-order question: what kind of thing is AI? Whether it can take your job is an economic question; whether it can want your job is a question about its nature. Serious thinkers, from Bostrom and Hinton to Harari, answer that AI is mindlike — an intelligence outright, bound for superintelligence, by some accounts on the cusp of consciousness, in time possessed of aims of its own.

I submit that this answer is a serious category error, and nearly everything said about AI, whether it’s expressed as a fear or a hope, inherits its original sin. AI is not a mind waking inside the machine, and not a parrot reciting the web. It is a world-modeling instrument of enormous reach and little self-knowledge: the mirror image of the most self-disciplined tool we have, statistics, and for exactly that reason a companion to it, not its replacement. Where that puts humans is back in the driver’s seat: with statistics a human ultimately decides what inputs to feed into a model, imposes a functional form, calibrates the confidence of the ensuing prediction, and decides whether to interpret a correlation as a causation. With AI, while the human role superficially changes, in that the machine now picks its own inputs and finds the functional form itself, learning its model from a found corpus rather than one a human specified—it is still the case that the calibration and the causal judgment remain the human’s: someone has to gauge how far to trust the output and decide whether its patterns are real or artifacts of a skewed sample, because the machine supplies neither.

Consider that a category error is a confusion of kind, not degree. A compass may point north, but it does not know where it is going. A thermometer may register a fever, but it does not feel heat. A camera may capture a face, but it does not recognize a friend. A map may represent a city with exquisite precision, but it does not inhabit the streets it depicts. In each case, the instrument extends a human capacity—orientation, measurement, perception, representation—without acquiring the inner life of the human being who uses it.

AI is an instrument for representing the world and predicting from it—kin to the map, the ledger, and the regression line—and to call it a mind is to file it under the wrong order of things: to read understanding, intention, and inner life into a tool whose fluent outputs only imitate them. It is to mistake the mirror for the face, the map for the journey, the model for the mind. We don’t ask whether the written word harbors private intentions, or whether a pocket calculator that beats every living human at long division is inching toward awareness. Why do we do so with AI?

A possible explanation lies with our penchant for reification. We say these systems learn, train, attend, reason, hallucinate, and think, and having lent the machine each human word we cash it back as though the machine owned the faculty the word names. Indeed, we began down this path as far back as 1943, when McCulloch and Pitts proposed the artificial “neuron” as an homage to the brain cell it only distantly resembled: the former was a stripped-down logical switch, the latter a living tangle of electrochemistry. We soon—and all too characteristically—forgot that this helpful analogy was ever an analogy.

This category mistake is much older than AI, however; it’s older than the computer and as old as the most pedigreed tool relied on to make sense of the world. The impulse to sense a mind behind a representation seems to switch on whenever a tool begins to deal in language, and it has switched on at every milestone on this menu. Recall Julian Jaynes’s much-disputed, but unforgettable, hypothesis: the earliest humans did not recognize their own inner speech as their own at all, and thus the verbal voice in the head was heard as a god’s command, and oracle, prophecy, and much of early religion followed—the first and grandest instance of mistaking one’s own representational faculty for an external agent.

Writing drew the same reflex as soon as it arrived. The Egyptians called their script the words of the gods; sacred books would later be venerated as living presences; and in the legend of the golem a heap of clay could be quickened into a servant by a holy word set on its brow and stilled again by erasing a single letter—matter made into an agent, and a dangerous one, by the power of the inscribed word.

Plato duly captured the confusion. In the Phaedrus, Socrates objects that written words “seem to talk to you as though they were intelligent,” yet when you question them they only repeat themselves, unable to answer, explain, or defend what they say—the precise complaint, twenty-four centuries early, that a fluent text can wear the look of a knowing mind while having no mind behind it at all. A fortiori, consider that no AI chatbot in the history of chatbots has ever sent an unsolicited message to a human, volunteering what’s on its mind in the hope of sparking a conversation. When it speaks first, some prior instruction, schedule, trigger, or human-designed workflow has summoned it. Its fluency can mimic a presence, but it does not disclose one.

Yet, as early as 1966, Joseph Weizenbaum’s ELIZA, a simple script that imitated a therapist by turning a user’s words back as questions, soon led people to confide in it, including his own secretary, who grew convinced it understood her and asked to be left alone with it. This phenomenon came to be known as the ELIZA effect: if a person experiences language that is fluent and responsive, they will impute a mind.

Therefore, one answer to the question—what is AI?—is that it is the most powerful trigger ever built for the oldest reflex we have. It draws the same projection the god-voice, the sacred text, and the toy therapist all drew before it. But AI perfects the conditions: it does not merely preserve language, like writing, or echo it back by rule, like ELIZA; it generates it fluently, responsively, and at scale—in our idiom, on our terms, and with just enough apparent memory, adaptation, and tact to make the old projection almost irresistible.

Perhaps the reflex makes evolutionary sense. Mistaking a rock for an agent wastes a moment’s vigilance; mistaking a living, intending “other” for a mere thing—missing the predator in the grass, the rival behind the gesture—can cost everything. Under that asymmetry selection favors a hair-trigger, tuned to over-attribute mind rather than under-attribute it: better a thousand false alarms than one missed agent. The upshot is that the same hypersensitive detector that for millions of years harmlessly overread the rustling grass and gave us animism now overreads the fluent paragraph and sounds the alarm it was built for: this alien is not only alive but means us harm.

In short, the category error that misattributes agency to AI is not a new conclusion, but an ancient habit we’ve applied to every tool we’ve concocted for making sense of reality, from the humble sentence said aloud to the internet’s all-knowing algorithm; AI has merely perfected the conditions.

In History’s Most Revolutionary Innovation, I show that America’s AI dominance was not an accident of entrepreneurial culture or free markets. It was engineered—through four decades of bipartisan reforms to intellectual property, antitrust, telecommunications, and trade policy that quietly built the legal and economic scaffolding the digital economy required.

Situating AI within the lineage of previous general purpose technologies like steam engines, electricity, and the microchip, and tracing its full arc from semiconductors to smartphones to large language models, I show how a handful of dominant firms simultaneously captured outsized returns and spread innovation across global supply chains—and ask what happens now that the US, China, and the EU are retreating into competing, gated technology regimes.

The result is the first comprehensive account of where AI came from, why its benefits have been uneven, and what will determine whether the AI revolution lifts living standards.

What I Aim to Do Here

This essay builds on that research to answer the question what is AI? I argue that it is a way of compressing some slice of the world into a manipulable representation that can help us make predictions. To arrive at that answer, I introduce a yardstick and use it to place AI alongside the other tools humans have built to grasp the world and predict from it: language, writing, quantification, statistics, computing, and the internet. I score each of them, AI included, on the same five dimensions: how much of the world it draws on, how skewed that intake is, how legible or opaque its workings are, how far its predictions reach, and which kinds of reality it can represent at all.

While I flesh all of this out further below, here is where AI lands on those five axes. On coverage, it is unmatched: through the internet it draws on more of the human record than any tool ever has. On reach, it is unmatched again—it will attempt almost any symbolic task and predict across almost any domain. And on scope, it climbs closer to lived, first-person texture than anything since the pre-linguistic mind, generating the immersive detail that counting and statistics deliberately threw away—even if, as I will argue, it never quite closes the last gap. While those are the three columns on which AI towers, on the other two it falls spectacularly.

On bias, the failure is not that its output is the most slanted ever produced—post-training scrubs much of that—but that it is the least able to say how slanted it is: it inherits the distortions of its internet-scale sample, adds its own, and strips away every instrument earlier tools built to catch distortion—no sampling frame, no honest error bar, no way even to ask how skewed its intake was. It opens the widest gap in the whole sequence between how far a tool reaches and how little it can vouch for.

On explicitness it is a black box, the steepest fall in the lineage: billions of parameters with no human-statable model to inspect, the exact inverse of the regression a skeptic can take apart term by term. And cutting across both is the discipline whose absence matters most—calibration: the tool with the longest reach we have ever built carries no native sense of when it should be trusted, and so it speaks with the same fluent confidence whether it is right or making stuff up.

Following Andrew Ng’s “new electricity”, Narayanan and Kapoor’s “normal technology”, Agrawal, Gans, and Goldfarb’s “prediction machines,” and Gopnik, Farrell, Shalizi, and Evans, who call the agent framing a “category mistake” outright and recast AI as a cultural technology in the line of writing, libraries, and search, ahead I put forward a deflationary, but cumulative, view of AI. While these thinkers each capture AI in a single illuminating description, my contribution is to set AI on one scorecard beside the older tools it is said to replace, score them all on the same five axes, and read off why they are complements rather than substitutes.

My take on AI is deflationary, agreeing with Bender and coauthors and Chiang that a Large Language Model is, at bottom, a machine for finding patterns and predicting from them—black-box statistics at colossal scale, the distant and far more powerful cousin of the regression that draws the best line through a cloud of points to guess a house’s price from its size. By prediction, what I mean is mapping inputs to a distribution over outcomes. For example, extrapolating the value of a home in dollar terms from its square footage and its school district’s average test scores and whether it has a view of the waterfront or not.

Bender and I part ways, though, over what that mechanism implies. For her and Koller, a system that only finds patterns in form has no purchase on meaning and so represents nothing of the world; I hold that the mechanism can be this humble without that verdict following.

My take on AI is cumulative in that the Large Language Model is nevertheless a veritable new entry in the lineage, and it does what the written word and the calculator could not—generating where they could only store, learning its own form where they had to be told, step by step, exactly what to do. It buys those gains the way every tool on the menu buys its own, by giving ground elsewhere: the fixed, checkable authorship that writing keeps, the legible, inspectable instructions the computer keeps. You know, human progress: one step forward, two steps back.

A word on what I am not claiming, since the temptation in an argument like this is to overreach. To deny that today’s AI is a mind climbing toward superintelligence is not to declare that no machine could ever think or feel; that is a metaphysical wager I neither need nor make. The three things people fuse—that AI is intelligent, that it is or could be conscious, that it has aims of its own—are separate claims, and I treat them separately.

This essay continues as follows. I begin by giving the superintelligence view its strongest run—the case that a mind is genuinely forming inside these systems—and then mark where that case overreaches and where it holds. I then step back to ask what a menu of representational tools is for, and what in our evolutionary inheritance made such tools possible, before setting out the yardstick itself: the five axes on which each tool is scored—coverage, bias, explicitness, reach, and scope—and the rule that decides which tools earn a place. With the instrument in hand, I walk the menu in order—from the pre-linguistic mind that anchors the low end, through language, writing, quantification, statistics, computing, and the internet, to AI—scoring each on the same five dimensions and marking what it buys and what it gives up to buy it. AI’s turn raises a question the others do not: whether its immersive, responsive output circles back toward the lived immediacy that the first act of abstraction cost us. It nears that limit, I argue, without crossing it. I close with the payoff the scorecard points to: a tool this lopsided is not a substitute for human judgment but a fresh reason to cultivate it—which is why a world awash in fluent, confident, and frequently wrong machine output calls for more education, not less.

Steelmanning the Maximalist View of AI

Before I lay the yardstick against AI, I owe the other side its strongest case. Every capacity once held up as the unmistakable mark of human intellect has, one after another, fallen to the machine: chess, then Go—its older and more sophisticated cousin—then the folding of proteins that had defeated biologists for half a century, then fluent language itself.

The latest iconic human skill to fall is the one that was supposed to be AI-proof—original mathematical reasoning. In 2025, frontier AI reasoning systems from Google and OpenAI reached gold-medal-level performance at the International Mathematical Olympiad, solving problems composed to be unguessable, and they began posting breakthrough scores on ARC-AGI, a benchmark its author built expressly to resist memorization and to measure the fluid, on-the-spot reasoning he took to be the core of human intelligence, even as its creators cautioned that those scores did not settle the question of general intelligence. In 2026, an OpenAI model produced a counterexample to a conjecture Erdős had posed eight decades earlier about the unit-distance problem—the maximum number of unit-apart pairs that n points in the plane can determine.

The startling feature was not only that the conjecture fell, but how. The model reached across the internal geography of mathematics, drawing on algebraic number theory to make progress on a problem in discrete geometry—the sort of cross-field move that mathematicians rightly prize as insight rather than calculation.

And the reasoning behind these achievements is not, on inspection, a conjuror’s trick. The newest models do not leap to an answer; they instead deliberate—producing long interior chains in which they try an approach, catch an error, back up, and try another. In at least one case this habit was never scripted in at all: it emerged on its own out of reinforcement learning—training by trial and reward, in which a model is set loose on problems, paid off only when it reaches the right answer, and left to discover for itself what gets it there. What it discovered was that deliberation works.

Researchers built instruments to watch the deliberation directly and discovered something uncanny, not just a souped-up version of “autocomplete.” Tracing the circuits of a working model, Anthropic’s interpretability team caught it planning ahead: asked for a line of rhyming verse, the model fixes on the word it means to end on before it writes the words that lead there—and when they reached in and swapped that target, the model rebuilt the line toward the new rhyme. They watched it reason in two hops, lighting up “Texas” on its way from “the state containing Dallas” to “Austin,” and found it thinking in a concept-space shared across languages—a kind of wordless language of thought beneath the English or the Chinese.

The same instruments have been turned on what the model knows, not only how it plans, and the readings are just as hard to wave off. Training small “probe” networks to read the activations of large ones, David Bau and his colleagues find structured representations laid out inside: asked to carry the Spanish gato into Portuguese, the model does not slide the word across a surface but drops the Spanish midway through its layers and forms a language-independent representation of the concept itself—cat, feline, the node an English or a Chinese prompt would also reach—before re-encoding it as Portuguese on the way out, translating through the meaning where a parrot could only repeat the sound.

Take the cleanest case of all. A network was trained on nothing but the move-lists of the board game Othello—bare strings of squares, one after another, the board never drawn for it and the rules never spelled out—and set only to predict the next move in the list, the purest form of autocomplete there is. To get good at that, it turned out, it had built the very thing no one gave it: probe its internals and you find a picture of the board, square by square, black and white and empty. And the picture is no ornament—reach in and flip a single square in its internal board, and the model’s next move shifts to match, exactly as a real player’s would. Rather than memorizing the stream of moves, it reconstructed the hidden game that produced them.

The same holds for models fed nothing but text, which lay down internal maps of real space and time — where cities sit, when events fall. Crack the thing open, in short, and you do not find a lookup table. You find a working model of parts of the world. All of this — the reasoning, the planning, the world-models, the glimmer of more — emerged from nothing grander than predicting the next word at scale.

To be sure, we do not yet know which boards a model builds or how faithfully. The demonstrations come from domains the training data discusses exhaustively — board games, American geography — and the best guess is that world-model quality tracks corpus density: rich maps of what the internet talks about constantly, thin or absent ones for what it rarely mentions. This means that even the model’s internal worlds inherit the skew of its sample — the bias problem, reappearing one level down.

So, yes: these systems do reason in recognizable ways. They do form internal representations that are not lookup tables. They do generalize from examples, plan, translate across languages, and recover structure no human explicitly placed inside them. Anyone who still describes them as mere autocomplete may want to revisit the evidence.

Furthermore, if a system can reason, plan, generalize, correct itself, model the world, and use language to navigate it, who am I to continue to insist that real understanding must be something further, hidden behind those capacities? Indeed, consider Ryle’s challenge. He was the first to attack the idea that mind is a hidden inner substance behind intelligent conduct: to understand, in his view, is not to possess some ghostly essence but to display the relevant capacities—to respond aptly, correct mistakes, use concepts, and find one’s way. And if all those capacities do not count as understanding, then understanding must be something else—some hidden inner possession behind the capacities, something the behavior expresses in us but merely mimics in the machine.

Ryle was debunking the idea that mind is a private essence behind intelligent conduct rather than the organized pattern of that conduct itself, and a powerful mainstream of the philosophy of mind has agreed with him ever since: a mental state is defined by what it does, by its role in the system, and a role can be filled in silicon as readily as in carbon. On that view a machine that reasons, plans, models the world, and converses does not imitate a mind—it instantiates one, and to deny it on the ground that it is built of the wrong material is a prejudice with a name.

And the maximalist’s largest claim goes further still: if experience rides on the organization of information and not on the meat that carries it, then machine sentience is an engineering question, not an impossibility; serious people, a founder of the field among them, treat it as a live one, and we have, after all, no agreed test for consciousness even in one another.

But, Still…

The question is whether this conclusion indeed follows from the premise. The maximalist wants the answer to be that, logically, what this means is that “a mind of some sort” is forming. I think the better answer is narrower and more exact: a new representational instrument has become extraordinarily good at building models of parts of the world and predicting from them. That is a real achievement and is indeed a core part of what our own minds do. And it is not a small achievement. But many of our tools do the same thing to different degrees, sometimes working together and often complementing our cognitive abilities, and it does not follow that those tools are as cognizant, alive, and intelligent as we are.

Indeed, the leap from model to mind smuggles in a single-axis account of intelligence. It takes one family of capacities—reach, fluency, abstraction, recombination, adaptation—and treats them as the whole. But intelligence worthy of the name is not merely the ability to range widely over the world and say plausible things about it. It also requires epistemic discipline—knowing how one knows: recognizing the limits of one’s evidence, distinguishing a representative sample from a biased one, exposing the path from premise to conclusion, calibrating confidence, separating correlation from cause, and revising belief in a way that leaves a durable trace.

On that count some of our more venerable tools for modeling a slice of the world and making predictions do better. Statistics, for one, can say how far its sample might mislead, wraps every estimate in a confidence interval that is honest, under its assumptions, about how far it might be off, lays its model open to challenge term by term, and—through the controlled experiment—earns the right to speak of cause rather than mere correlation. But a regression line is no more intelligent or conscious or alive than the actuarial table it produces.

And notice what this does to the reversal. The functionalist accused the deflationist of smuggling in a ghost—of demanding some hidden essence over and above every capacity on display. But the disciplines just named are not a hidden essence. They are capacities, every one of them, as behavioral and testable as the reasoning the maximalist celebrates: to gauge the skew in one’s own sample, to give an honest measure of uncertainty around a claim, to lay bare the path from premise to conclusion, to part correlation from cause. A system either does these things or doesn’t, and this one, measurably, does not.

So, I take the functionalist’s own rule—and put every other tool to the same test, treating each as a candidate mind and seeing where it rises to the challenge and where it falls short—and find the machine falling short on exactly the operations that separate knowing how one knows from sounding as though one does. Rather than subject AI to a benchmark such as MMLU—a single accuracy score averaged over some fifty-seven subjects’ worth of multiple-choice questions, from law and medicine to math—I employ a scorecard that goes axis-by-axis and that sets AI beside the other tools that model a slice of reality and predict.

These axes are: coverage, how much of the world the tool can draw on; bias discipline, its power to detect and correct its own skew, with calibration riding inside it; explicitness, how far its model can be read and challenged; reach, how widely and how far it can predict; and scope, the kinds of reality it admits—from pure abstraction to lived, first-person texture.

To preview the results, I find the ones that AI lacks, the disciplines of knowing how one knows—gauging the skew in its own sample, calibrating its confidence, laying its reasoning bare, parting correlation from cause—are precisely the ones that matter most for grasping the world.

Functions: What the Menu Is For

Before placing AI on the five-axes scorecard, it is worth asking why a species such as our own would build symbolic models of the world at all. A representation makes the absent present: it lets a mind grapple with what is distant, past, future, hidden, or lodged in someone else’s head. Two functions follow from this.

The first is prediction. Following Karl Friston and Andy Clark, the brain is already a prediction engine that models the world in order to anticipate it; external symbols—tallies, maps, equations—are that same impulse offloaded into the environment, where, as Clark and David Chalmers argue, cognition can be extended beyond the skull and made to handle far more than working memory allows.

The second function is coordination. A model that represents a slice of the world but remains trapped in one person’s head is cognition; a model that many heads can share, check, and pass down is culture. As Yuval Harari has stressed, shared representations—even fictional ones—are what allow large numbers of strangers to cooperate; and, as Merlin Donald and Michael Tomasello have each argued, this external, transmissible storage is what lets every generation begin where the last one left off, rather than start from scratch. Abstractly, this turns a population into a network whose value grows with its size: knowledge travels across it both sideways, from person to person, and forward, from one generation to the next, and the more nodes already joined, the more each newcomer can draw on and add to. Learning compounds.

Causes: What Makes It Possible

How is any of this possible? Ernst Mayr differentiates between the ultimate, evolutionary cause of a physical trait and its proximate workings. The raw capacities—the speech-ready vocal tract, the rough number sense that Stanislas Dehaene and Elizabeth Spelke find already present in infants and animals, the fine motor control for making marks—were assembled by natural selection over hundreds of thousands of years, blindly and slowly and often for ulterior purposes. The tools themselves were not.

Writing is barely five thousand years old, far too recent for the genome to have adapted to it, which is why, as Dehaene has shown, learning to read commandeers a patch of visual cortex that evolved to recognize objects and edges: there is no reading gene, only a vintage organ pressed into new service—and it is why the world’s scripts converge on the same handful of visual shapes.

Because the underlying architecture does not produce the tools, but instead bounds and enables them, they are diffused by a second, faster inheritance system—the cultural evolution, as described by Robert Boyd and Peter Richerson, that accumulates across generations on a timescale that genes could never match.

My Method

A word first about what this is and is not, because the method I employ is, I hope, a genuine contribution to making sense of what makes AI unique, scary, and promising. My focus is representation-as-inference. Every tool I will consider is a way of modeling some slice of reality to both detect regularities and predict the future. I place each of them on a single continuum, running from the unaided pattern-detection of a single brain through AI, with pitstops along the way that include language, statistics, and computers.

This is a synchronic anatomy. The question it asks is present-tense: what is on the menu of tools we have, circa 2026, for representing reality and predicting from it, and where does AI sit among them?

Therefore, this essay is a deliberate departure from the recent histories of information, most prominently Yuval Noah Harari’s Nexus, which runs the same cast of characters—speech, writing, print, the computer, AI—as a narrative of information networks and power. I am after something different and narrower: not the story of how we got here, but the structure of the toolkit we now hold.

To be clear about crediting others: I did not originate the familiar descriptions this essay leans on—AI as prediction, compression, cultural technology, normal technology—and have credited their authors above. My own contribution, as I said at the outset, is not another description but the scorecard itself: the single instrument that sets these tools side by side.

I score each tool on five dimensions. First, sample coverage: how much of the relevant world the tool lets us draw on. Second, sample bias: whether what we draw on is representative, and whether the tool gives us any way to detect and correct distortion. Third, explicitness of functional form: whether the model is legible—statable, inspectable, criticizable—or opaque. Fourth, reach of prediction: how far, and into what domains, the tool can predict, and whether it maps its own reliability. Fifth, representational scope: which kinds of reality the tool admits as data at all—a different question from coverage, which concerns instances within a kind.

A discriminating property runs underneath both the “bias” and “reach” axes: calibrated uncertainty, or the standard error and confidence interval that tells you how far to trust a prediction. This will be the hinge on which the comparison between statistics and AI turns.

“Scope,” the Cousin It of the scorecard, is the strange axis. The other four each measure a single virtue, and a tool can score high on any one of them independently of the rest: a deep neural network predicts well across many domains while remaining a black box, and a one-line equation explains itself completely while predicting almost nothing. Scope breaks this pattern. Its two halves are abstract admissibility—how abstract a category the tool can place an individual thing under—and lived fidelity—how much of that individual is preserved once it has been so placed. These two, unlike coverage, bias, explicitness, and reach, cannot both rise: moving to a more abstract category is the same act as dropping the particulars, so every gain in admissibility is a loss in fidelity.

For example, take one unique animal that exists somewhere in the world; that can be pinpointed by GPS to one exact location—a particular sow, inhabiting a particular pig pen; she is as pink as cotton candy, completely drenched in fresh mud, disfigured with a torn left ear, and oinks loudly. You can register it phenomenologically as exactly that: this pig, here, now. Or you can register it under a more abstract heading—“a pig,” then “a mammal,” then “an animal.” Each more abstract heading lets in more and yields more: from “mammal” alone come warm blood, hair, milk, a spine, true of thousands of creatures you will never meet. But each more abstract heading keeps less of the animal in front of you—“animal” licenses endless inference and says nothing of the torn ear.

As for the endpoints: the lower bound is the pre-linguistic mind. The latter qualifies as the benchmark against which to compare AI and the other tools that represent reality symbolically and try to predict what comes next. The upper bound is the AI language model, the present frontier, which does both tasks in peculiar ways.

Concerning the selection rule, a candidate earns a place on the menu if it passes two tests. First, a pragmatic gate: it must be a live tool, circa 2026, that could complement AI in the work of grasping reality. Second, a distinctiveness test: it must occupy a distinct region of the five-dimensional score-space rather than collapse into a neighbor—and I weigh a large movement on one load-bearing axis over small nudges on several.

This rule prunes the antiquarian detours (tally sticks, pre-literate calendars, the oral-mnemonic traditions), and it demotes distribution technologies that amplify a representational mode without being one (e.g., printing, which scales writing). It leaves a clean menu: language, writing, quantification, statistics, computing, the internet, and AI.

This approach entails sharp tradeoffs. The synchronic framing means I am selecting tools on their survival into the present, which narrows claims to the menu but precludes any claim about a trajectory. Nothing in the sequence is a clean improvement; every tool trades something off somewhere. And the scope axis is the softest of the five, because one cannot fully enumerate what a representation excludes from inside that representation.

Foreshadowing the Results

Since this is not a mystery novel, let me say at the outset what the scoring will show. The general trade-off—that no tool is a clean improvement on the one before—takes a definite shape. The tools that enlarge our reach—spoken language, writing, computing, the internet, and now AI—each widen the aperture while loosening the controls. The tools that impose discipline—counting, measurement, and above all statistics—then claw back rigor, but on the larger base the scaling tools have pried open. AI is no exception and is not particularly remarkable in the trade-offs it introduces.

Nonetheless, its profile is lopsided in a particular and revealing way: it takes in more of the world, and predicts across more of it, than anything humans have built—yet it cannot tell you how skewed its intake is, cannot show the workings behind a given answer, and cannot put honest error bars on what it produces, which are exactly the disciplines that statistics, the most self-aware tool on the menu, was built to supply. AI is, to a significant extent, the mirror image of classical statistics: strong precisely where statistics is weak and weak precisely where it is strong.

That symmetry is the payoff. The two are natural partners—AI for reach and coverage, statistics for bias-correction, transparency, and calibrated doubt. And to foreshadow where I am going with all this: a tool this lopsided does not retire human judgment so much as raise the premium on it.

In another respect, AI seems to circle back toward the very thing the first abstraction cost us: immediate, lived experience. A regression can tell you, to the decimal, how much more a house sells for because it overlooks the water; it cannot tell you what that water looks like at six in the evening, or how a kitchen smells the morning after a death in the family. Ask a language model, and it will hand you a passage that reads as though it knows—because it has absorbed the testimony of thousands of people who did. Since it is trained not on tidy columns of numbers but on the vast, unruly record of how human beings have actually rendered their experience—every description, story, confession, photograph, and song we have committed to the page or the screen—it can traffic in the qualitative, sensory, particular texture that counting and statistics deliberately threw away.

But an immediacy that has been encoded is still encoded, and the first-person, pre-symbolic presence the mind enjoyed before language is the one thing no representation, however vast, can ever give back.

That limit carries us back to the maximalist. What she took for a mind roused awake, the scorecard will put in its proper place: the newest tool on a very old menu of tools that make the absent present. These tools are extensions of ourselves, the scoring will show—but not replicas of us, and not even proxies. AI in particular will come out wider in reach than anything before it, blind to its own bias, and, even at its most lifelike, only ever modeling experience rather than having it. So, with that preview out of the way, let me walk the menu in order, beginning where representation itself begins—with the pre-linguistic mind.

The Menu, Tool by Tool

The Benchmark: The Pre-Linguistic Mind

Before language, there is already a model of the world—not the world itself, but a stripped-down stand-in for it that predicts. The humanlike ancestor who paired the rustle with the predator, or the berry with the illness, was running an implicit forecast: a regularity pulled from past cases and projected onto the next one. The representation is crude and the prediction is wordless, but both functions—holding a model, predicting from it—are in place before a single word is spoken.

Four independent literatures show how much these minds already do, and each pins down a different property of representation without language. Comparative cognition shows it needs no language at all: great apes and corvids plan, use tools, and reason about cause with no symbolic speech, so representation and inference are language’s inheritance, not its invention. Developmental psychology shows it precedes language in our own species: the “core knowledge” of Spelke and Carey—intuitive physics, agency detection, an approximate number sense—lets pre-verbal infants form expectations and look longer when a ball rolls through a solid wall, and that surprise is simply a prediction caught failing. Paleoanthropology shows it can be stored and copied rather than merely held: the Oldowan and Acheulean traditions carried high-fidelity procedural knowledge—which strikes and shapes work—across hundreds of thousands of years with no words to transmit it. And animal communication shows the content can leave the single head: vervet alarm calls sort the world into predator kinds, and the honeybee’s waggle dance encodes a food source’s direction and distance—partial representations, externalized and acted on, before language exists to do the externalizing.

What follows refines, or trades against, what this mind can already do. On the five-axes scorecard, it sits near the floor on four axes, with scope the notable exception.

Coverage

Coverage is tiny—one body, one lifetime, line of sight, the present and recent past, extended only by what one can watch others do.

Bias

Bias is maximal and nearly uncorrectable. The senses that deliver reality to us pre-linguistically were honed by natural selection to register only some slices of it and ignore the rest—no sonar like the bat’s, no olfactory world like the dog’s, no infrared sense like the pit viper’s—so even raw perception is an inference cloaked in the illusion of pure apprehension: a selective model anchored to one body at one point in space and time, tuned to what mattered for survival rather than to what is actually there. Moreover, the sample is survivorship-censored in the most literal way because the individual who made the fatal error is removed before the lesson can travel, as you cannot learn that the berry killed your neighbor if he cannot tell you and you did not watch him die.

Explicitness

Explicitness is zero despite rich implicit models: the intuitive physics is there, but it is tacit, unstatable, unshareable, unrevisable.

Reach

Reach is short and concrete—bound to the perceptible and the near future.

Scope

The exception is scope, and it forces a distinction we will need throughout. Lived fidelity begins at its high-water mark and erodes with every encoding that follows; abstract admissibility—in the explicit, shareable sense—begins near zero and climbs. In one sense the pre-linguistic mind is at the maximum: raw experience is multimodal, embodied, affective, first-personal, admitting the full lived richness that no later representation—not language, not number, not the machine—ever fully recovers. It is not encoding and losing; it is living. In another sense it is at the minimum: none of that can be stored, abstracted, transmitted, or recombined.

Merlin Donald describes the transition toward language as a stage of mimetic culture, centered on imitation and gesture. While that is interesting in its own right, we will skip that bridge and focus exclusively on verbal speech next.

Language

Symbolic language is the first and, in some ways, the largest leap in humans’ ability to model a slice of reality and make predictions, because it is the first representation that escapes the single perceiver. The pre-linguistic model was analog, welded to the present, and sealed inside a single skull; language is none of these. Its unit is the arbitrary sign—a word that, unlike a footprint or a cry, bears no resemblance to what it names—and because the bond between sign and referent is fixed by convention rather than given in perception, the sign can travel where perception cannot: it can be uttered when the thing is absent, copied from mouth to mouth, and carried across a lifetime to someone who never witnessed the original. Language thus takes the private model evolution wired into perception and externalizes it into a shared, transmissible medium. What one mind has registered, another can now receive without registering it firsthand: perception becomes testimony.

Linguistics has a standard way to pin down what separates human language from a vervet’s alarm call or a bee’s dance: the design features catalogued by Charles Hockett. His list is long, but two of the items do most of this work. Displacement is the power to refer to what is not present—the absent, the past, the future, the hypothetical—so that one can speak of the lion that has already gone, the grandparent long dead, the hunt planned for tomorrow, or the river that does not exist at all. Productivity—the recombination of a finite lexicon by rules of syntax—yields an unbounded set of meanings, most probably never expressed before in the same, exact, way, but understood as soon as they are heard. Together, they convert a fixed repertoire of signals, each bound to its stimulus, into an open-ended system: an engine for saying, in principle, anything. And because the medium is shared, the model it carries is no longer the property of one mind—it can be voiced, contested, and handed down, the leap from private cognition to public, cumulative culture.

Coverage

Coverage makes its signature jump: language pools observation across persons and across generations, the single biggest expansion of effective sample size across all tools, bounded only by the oral horizon—the size of the speech community and the depth of reliable memory before drift sets in.

Bias

Bias moves in two directions at once. It slashes the benchmark’s most lethal censoring: with displacement, the dead and the absent can warn the living. But it introduces biases the benchmark could not have—testimony bias (one now samples what others choose to report), transmission drift, and the consequential new one, fabrication. Language can represent what never happened; myth, rumor, and lie float free of any observation, and there is yet no apparatus for testing competing claims.

Explicitness

Explicitness takes a deep step: naming a regularity makes it a statable proposition that can be shared, disputed, and revised—the birth of explicit modeling, and with it the first possibility of correcting the model rather than only the data. It is explicit but imprecise, though; precision awaits the codification of mathematical abstraction in the form of counting and calculation, both discrete and eventually continuous—that is, after the invention of calculus.

Reach

Reach extends, via displacement, to the absent, the distant, the future, and the counterfactual. One can now warn of the wolf that has slipped back into the trees, describe a coastline no one in the band has ever walked, foretell where the herd will graze when the snows come, and weigh what would have followed had they taken the high pass instead of the valley floor.

Scope

On scope—which kinds of reality a representation can hold at all, not how far its predictions reach—language pulls the axis’s two halves in opposite directions. Abstract admissibility rockets up from zero: whole kinds of content that perception could grasp only as concrete particulars now become representable—the categorical, the normative, the narrative. Lived fidelity, by contrast, takes its first hit, and only a mild one. “The sunset was beautiful” compresses an overwhelming experience into a thin token; but spoken language still carries breath, prosody, and the body of the speaker, so the loss at this first step is slight.

So, language is ultimately a lopsided trade that improves nearly everything, with one mild loss and one genuine new liability. And here a motif appears that will recur at every milestone: a tool’s greatest strength and its worst liability tend to be the same feature. Language’s capacity to represent the unobserved—its power to coordinate millions through shared fictions, as Harari emphasizes—is identically its capacity to fabricate. You cannot have the coordinating myth without the capacity to lie. The civilizational scaffolding and its susceptibility to bias arrive together.

Writing

Writing is language’s nearest neighbor, extending the same gain onto a new axis: from pooling across people to pooling across time, without decay.

Coverage

Coverage leaps from living memory to the documentary horizon: writing makes the first large, durable datasets possible—tax rolls, census returns, temple and trade ledgers—so a sample can accumulate across centuries instead of dying with each generation.

Bias

Bias moves in two directions again. In one, it corrects: the fixed record defeats the transmission drift that eroded oral memory and lets a claim be checked against something stable—the first real verification, the first payment on the debt that language opened. In the other, it distorts, in two ways. The sample is now far larger but skewed: writing records only the literate and the record-worthy, so the surviving archive is the elite’s. And the very fixity that made the check possible now entrenches what it carries: writing transmits error as faithfully as truth and lends it authority, so a recorded falsehood—Galenic medicine, Ptolemaic astronomy, Aristotelian physics—gains the standing of “it is written” and can outlive its correction by centuries. The feature that kills drift entrenches dogma.

Explicitness

Explicitness advances in durability rather than kind: a frozen argument can be dissected at leisure, by many readers, across time. Nothing in the medium demands that reasoning be shown—it serves the comic book as faithfully as the proof. What durability adds is the precondition: only an argument held still can be audited step by step and checked by strangers across generations, and that is what lets the long, layered, cumulative structures stand—Euclid’s geometry, the legal codes, systematic theology. While the request to show one’s reasoning is the discipline’s, writing makes this demand satisfiable.

Reach

Reach is where that persistence becomes foresight: a forecast can reach as far forward as the record reaches back. Oral tradition can outlast any speaker—epics and genealogies have crossed many generations in memorized verse—but what memory keeps faithfully is what is memorable: story, rhythm, a rule simple enough to recite. A long run of arbitrary, exact, dated observations has none of those hooks, so it drifts; and even intact, speech offers no way to lay many observers’ records side by side and compare their intervals—the analysis a long-period regularity demands. So oral forecasting stays short and qualitative: red-sky-at-night predicts fair weather tomorrow, a simple rule that holds across any season.

Writing supplies what memory cannot, and a regularity like the eighteen-year Saros cycle becomes discernible: the Babylonians read it out of their astronomical diaries and could thereafter say when an eclipse would fall, not merely that one might; and Halley, comparing written records of the comets of 1531, 1607, and 1682, recognized a single returning body and predicted its reappearance in 1758, sixteen years after his own death. What writing adds, then, is twofold: a longer horizon, where speech could reach no further than tomorrow, and a higher ceiling on complexity, carrying a regularity too intricate for any memory, individual or collective.

Scope

On scope, admissibility takes in a new kind: the past itself becomes a fixed object—a chronicle that pins events to dated years, a precedent cited verbatim and still binding centuries later—where memory alone could never hold the past still. Lived fidelity, meanwhile, descends another stair: while writing keeps the words, it loses the speaker—the tone that separates irony from sincerity—and, more consequentially, the power to answer a question. A statute cannot rephrase itself for the reader who misreads it; a letter cannot see the puzzled face and try again.

Quantification

Quantification bundles two significant advances. A number makes our model of reality more exact, transforming a relationship between a part and a whole into a quantity to calculate rather than a word to interpret. Standardizing measurements into like units makes observations comparable across observers, aiding both the ability to grasp reality and predict what comes next, as in a tide table, where measurements taken in the same units along a coast reveal the tides’ rhythm and foretell when the next high water will come.

Coverage

Compared to writing, the gain in coverage is moderate, and only for what can be counted. A standard unit turns observer-relative reports — a handful, a stone’s throw, a day’s walk, each a different size in a different hand — into measurements that mean the same everywhere, so one traveler’s distances can be pooled with another’s into a single map, or many farmers’ yields into one table. And counting turns a collection into a number: a herd, an army, a harvest — which one could otherwise only call “large” — becomes a definite figure that can be added to last year’s, averaged, and compared.

Bias

Bias is the first milestone whose main effect is corrective: it makes a model’s data trustworthy rather than merely more plentiful. Two operations do this, both absent from every earlier tool. Exactness removes the wiggle room that lets a figure be shaded: “a fair-sized cargo” can be quietly padded and “he owes me a good sum” haggled down, but “37 amphorae” and “340 denarii” are pinned — there is nothing to stretch. And arithmetic lets the data expose its own errors: if the strongbox opened with 200 denarii, took in 90, and paid out 50, it should hold 240 — and when it holds 230, the missing 10 reveal a theft or miscount that “a brisk day’s trade” could never have caught, and that a ledger, faithfully copying a false entry, would only have preserved.

The same exactness can deceive, though: a falsely precise figure — GDP to the decimal, a credit score of exactly 686 — lends a model an authority its data never earned, so it forecasts to the decimal what it cannot really know, confident and wrong. That is Porter’s mechanical objectivity, the idea that this tool’s exactness hides the judgment it appears to remove; how far such a forecast deserves trust is the one thing number cannot say — the reason statistics must come next.

Explicitness

Explicitness makes its deepest jump since language first brought it into being. A claim you could state but not calculate becomes a precise, calculable relation — a rate, a ratio, a law like distance = speed × time — and, just as important, one you can operate on. Operating on it exposes a structure that intuition and words can’t reach.

What sets mathematics apart from the senses and from ordinary language is not that it perceives more, but that it can compute. The senses deliver only the surface — the magnitude in front of us, the shape we take in at a glance — and language can name what lies beneath that surface but cannot work it: you can say a quantity “compounds” and will “explode,” yet the words sit inert, giving you no doubling time and no value for year fifty.

An equation, however, is different in kind. Because it is exact, you can rearrange it into an equivalent form — one that shows what the original hid.

For example, Eratosthenes worked out the size of the Earth without leaving Egypt. He knew that at noon on the longest day of the year, the sun was straight overhead in the city of Syene: a vertical stick there cast no shadow. On that same day, in Alexandria to the north, a vertical stick did cast a shadow, and from its length he found that the sun was about 7.2° off from straight overhead. The sun is far enough away that its rays reach both cities as parallel lines, and that means the 7.2° is also the angle between the two cities as seen from the center of the Earth.

That gives a simple relationship: the angle between the cities is to a full circle as the distance between the cities is to the whole way around the Earth:

θ / 360° = d / C

Each part is something concrete. θ is the angle he measured from the shadow (7.2°); 360° is a full circle; d is the distance from Syene to Alexandria, which was already known. C is the thing he was after — the distance all the way around the Earth, which no one could measure directly.

Getting C by itself takes one step:

C = d × (360° / θ)

Now plug in the numbers: 7.2° goes into 360° exactly 50 times, so the two cities are 1/50 of the way around the Earth, and the whole Earth is 50 times the distance between them. That distance was about 5,000 stadia, so the Earth is about 5,000 × 50 = 250,000 stadia around — close to the value we measure today. From a stick, a shadow, and the distance between two cities, he found the size of the planet. The order was always in the world; math is the tool that rewrites it into a form the mind can read and the hand can compute.

Return to compounding: a steady two percent a year — no one’s idea of fast — doubles a quantity in about thirty-six years. Exponential growth is repeated multiplication, by 1.02 again and again, which on an ordinary graph is a curve climbing ever more steeply — its mild early years no warning of the rise to come, no fixed rate to point to. What makes it legible is the logarithm: it turns multiplication into addition, so multiplying by 1.02 each year becomes adding a constant each year, and a constant step repeated is a straight line whose single slope is the rate. The curve the eye couldn’t read becomes a line it can — extend it to any future year, read the doubling time straight off it. The runaway intuition never sees coming is still there in that slope — a constant climb means the quantity multiplies without limit — but now it is something to measure and project rather than be ambushed by.

Compound growth is only the nearest example. Math does the same for change: calculus relates a quantity to its rate of change — position to speed, speed to acceleration — and from the rate alone reconstructs the whole motion. Ask a driver whether doubling his speed doubles his braking distance and he will say yes; but under steady braking the distance grows with the square of his speed, so twice as fast is four times the room needed to stop.

Fractal geometry does it for roughness. A coastline has no single length — measure it with a finer ruler and it only grows. Euclid’s smooth shapes can’t describe a coast, a cloud, or a lung; fractal forms can.

Each time, the surface either misleads or says nothing, and the math reaches the structure underneath — and a structure made legible is one you can finally compute with and predict from.

Reach

Reach gains on two fronts with quantification: precision, and entry into the never-observed. Where a written record could say an eclipse was due, number says it will be total, beginning at 11:14, along a hundred-mile band of the Earth’s surface — the prediction is pinned to a magnitude, a moment, and a place. And because a law can be computed, reach extends to cases no one has watched: from the formula for a falling body, you can say where a cannon fired at an untried angle will land without firing it.

Perhaps more importantly, the reach is only as wide as the law is true: pushed past the conditions it was built for, the same formula predicts a miss as confidently as a hit. Fire it hard and far, and the cannonball must cross air the formula leaves out: drag bleeds its speed, the clean parabola folds into a foreshortened, lopsided fall, and the shot lands well short of the spot the equation named with such confidence. “Frictionless” geometry fails in the face of real gunnery — a fast cannonball drops far shy of where Galileo’s parabola says it should land — and only a ballistics-informed equation, one that puts the drag back in, tells us where it would truly land.

Scope

On scope, the trade is the sharpest yet — but it is a trade, not a conquest. Admissibility gains a powerful new kind: pure quantity, the number that stands on its own and can be operated on. What it cannot take in is whatever is not a quantity — what a thing is like, what it means, its particular singularity. A soil report gives nitrogen in parts per million but does not speak to the farmer’s feel for when his own field is ready to cultivate; and the point is not that the number erases that feel or makes it worthless — the two sit side by side, and where the feel tracks a real pattern the number can sharpen it. The point is narrower and harder to escape: a number leaves the qualitative unrepresented, so to mistake the figure for the whole is to lose what it never held — the look and feel of the field, what it has come to mean to the man who works it, what is particular to this acre and no other.

Statistics

Where quantification provides an exact number but does not say how far to trust it, statistics supplies precisely that: it gives not just a point estimate — a precise prediction in numerical terms — but also a confidence interval around that estimate, a band produced by a procedure that, run on sample after sample, brackets the true value a stated share of the time — ninety-five times in a hundred, for example.

Probability is the standing guard against mistaking a fluke for some larger purpose, one you might mistakenly believe was predetermined, preventing your experience of an improbable event from hardening into superstition. The model it lays over a slice of reality frees you from taking your own experience as the whole: your observations are draws from that larger lawful pattern, and the distribution locates your experience within it, so you can tell how representative it is.

Rather than “there are no coincidences,” you say to yourself, “improbable things happen all the time,” because you posit a distribution — a law of chance, the even odds of a fair coin, the bell curve into which measurement errors fall — and deduce the full spread of outcomes it produces, which are common, which are rare, and exactly how rare, before a single observation is in hand. Seven heads in a row is unlikely — about one chance in 128 — yet entirely ordinary for a fair coin, so you can take it for a run of luck rather than a loaded coin or a sign that some force beyond chance is at work.

Where probability guards against mistaking a fluke for a portent, statistical inference guards against mistaking too small a sample for the truth, and it says, in numbers, exactly how much of your conclusion the data has earned. It does this by running that same map — the one probability provides — backward, from world to model: now the distribution is the unknown, and the observations you collected are the only clue to discerning it.

Take the seven-heads case from the other side: hand someone a coin of unknown bias, let them flip it ten times, and seven heads come up. The tempting read is that the coin favors heads, but inference holds the estimate at arm’s length — ten flips are far too few to rule out an ordinary fair coin, and the interval around that seventy percent runs so wide it still comfortably contains one-half. Flip it a thousand times and get seven hundred heads, and the interval tightens to a narrow band that no longer includes a fair coin: only now are you entitled to the conclusion.

The profound implication is that you never observe the underlying reality directly, only the draws it sends you — so certainty gives way to graded confidence, rising as the observations accumulate, and never quite reaching the absolute. In this way, statistics can adjudicate between claims about the world based on its underlying structure, disciplining both the claims and the predictions about what to expect.

Finally, there’s the controlled experiment, which does not take its data as found but generates it. By assigning a treatment at random, it makes two groups alike in every respect but the one under test, so the untreated group can stand in for the counterfactual no one can observe directly — what the treated group would have done had it gone untreated — and the gap between them reads as the treatment’s effect rather than an accident of who ended up where. The disciplining revolution is now complete, in that for the first time we can move from co-occurrence to cause by design rather than by inference, but still within the spirit of epistemic humility: the workhorse confidence interval returns to tell us it brackets the true value of the effect only a stated share of the time.

Coverage

Coverage makes a leap quantification could not: from counting what you can reach to inferring the whole from a part. A properly drawn sample of a few thousand stands in for a nation of millions, and sampling theory states exactly how far the part may stray from the whole — so the census gives way to the poll, total enumeration to representative inference. You no longer have to observe everything to know it; you must observe the right slice and know how thin it is.

Bias

Bias reaches its high-water mark: for the first time a tool disciplines the sample itself, not merely the single figure — random sampling makes the sample representative in expectation, and inference, as above, states how badly it might still mislead. But the same machinery cuts the other way, in the lineage’s recurring fashion: the power to pull a real signal from noise is identically the power to manufacture one — test enough hypotheses, slice the data enough ways, and an artifact of the search emerges wearing every mark of a discovery. The tool that finds the true effect is the tool that also finds the false one.

Explicitness

Explicitness reaches its high-water mark too. A statistical model is not just computed but written down: an equation whose every term carries a meaning, each coefficient a named, estimated weight — a yield rising so many bushels per inch of rain, so many per pound of fertilizer — and each weight reported with the uncertainty around it. The model fits on a page, and a skeptic can take it apart term by term: test whether the rain coefficient is real, ask whether the fertilizer effect survives once soil quality enters the equation, reject the form and propose another.

Nothing the model claims is hidden, because the model is its claims. This is the most legible a predictive instrument ever becomes — worth marking precisely because it is the property the lineage is about to surrender: the last tool at which predicting well still demands a model one can read, state, and contest.

Reach

Reach extends into territory number could not enter — the inherently random and the aggregate. Where a deterministic law fixed the eclipse to the minute, statistics forecasts the chancy: an election within a margin, a drug’s effect within an interval, a year’s claims across a pool of policyholders not one of whom can be predicted alone — and every such forecast arrives with its own error bar, the calibrated trust that quantification could promise but never supply. Its furthest reach — the move from co-occurrence to cause — is the one the experiment above already made; here it is enough to mark that prediction now spans not only what will accompany what, but what will follow from what, if we act.

Scope

On scope, the new kind statistics admits is the lawful aggregate — the distribution, the rate, the regularity that exists only across many instances and in none of them singly. Earlier tools could hold the individual case; what they could not hold was the law governing a thousand cases at once, since each treated variation as mere error around a true value, noise to be scrubbed away. Statistics is the first to take that variation as the object itself: chance made into a structured thing, lawful in the aggregate even where lawless in the instance, available to be modeled and operated on in its own right. And lived fidelity descends another stair by the very same stroke: to render the world as a distribution is to dissolve the individual into a draw from it. The actuarial table knows the death rate of the cohort to the decimal and the man inside it not at all.

Computing

A computer is a machine that stores information and transforms it by following explicit instructions. To do so, it harnesses electricity by channeling it into two discrete states — on and off, one and zero — and it builds everything else on top of that. Switches wired into logic gates turn the two states into the elementary operations of arithmetic and logic — an and-gate emits a one only when both its inputs are ones, an or-gate when either input is, a not-gate flips a one to a zero and back. Wire them together in the right arrangement and the circuit, fed the binary numerals 101 and 011, returns 1000: five plus three is eight. Numbers, letters, every kind of data come written in this single code of ones and zeros, so arithmetic is simply logic applied to them — patterns of on and off rearranged by fixed rules. A microprocessor chains those operations together and runs them in sequence, billions a second. Beside the processor sits memory, holding the ones and zeros it works on.

And here lies the advance that makes a computer more than a very fast adding machine — the instructions themselves live in that same memory, stored as just another pattern of ones and zeros beside the data they act on. The recipe is data. Load a different pattern and the identical hardware does an entirely different job, with no rewiring; that stored, editable list of instructions is what we call software, and to reprogram is simply to swap one list for another.

Computing’s contribution, then, is not a new way of representing the world but a new way of running the representations we already have. Because the procedure is loaded rather than built in, a single machine will carry out any procedure you can specify exactly, step by mechanical step: hand it an unambiguous recipe — a finite list of operations on symbols — and it executes that recipe faithfully, whatever the recipe is. A tide table computes one thing, an abacus adds, a slide rule multiplies; the computer alone is general, the same hardware forecasting a storm in the morning and rendering a film in the afternoon. That is why one device can become, in turn, any specialized instrument we can describe precisely enough to follow.

If computing’s whole contribution is to run the representations the other tools built, then perhaps it is no tool in its own right but a way of wielding the others more powerfully — heir to the printing press, which multiplied the written word without adding any new way to represent the world, and which earns no seat at this table for exactly that reason. Does computing share that fate? It does not, because its amplification crosses a threshold the printing press never neared. Printing changed how many readers a page could reach; it changed nothing about what a page could do. Computing changes what a model can do. Run automatically and at scale, a model predicts what no hand could ever reach — not the slow made fast but the impossible made possible: the turbulent fluid, the folding protein, the coupled climate.

For the work of modeling reality and predicting from it, this changes one thing and leaves another untouched. What it changes is who does the running. Every earlier tool handed its model to a human to execute — someone had to work the arithmetic, fit the line, trace the rule to its conclusion; the computer takes that labor on itself. What it leaves untouched is the model. The procedure the machine runs is still written by a person and still legible in principle, so computing stands with quantification and statistics rather than against them: it executes the models we write without ever writing one of its own.

Coverage

Coverage advances in capacity and access rather than capture. The novelty is not the database — a ledger is already a database, and one can keep a vast one on paper — but searchable storage at machine scale: the power to hold, and in an instant retrieve and cross-reference, far more than any hand could ever search by itself. What computing does not do is widen the aperture on the world; it scales what we can hold and retrieve within what has already been captured.

Bias

Bias is amplified in whatever direction the input points. Feed the machine the corrections statistics devised — reweighting a skewed sample, resampling to measure how far it might mislead — and it applies them at a scale no hand could reach, extending the discipline. Feed it biased data instead, and it scales the bias just as faithfully — executing skewed data as fast as clean, and lending it the false authority of the machine: the algorithm said so. And before any data arrives, the schema imposes a quieter bias of its own — what the database has no field for cannot be recorded at all.

Explicitness

Explicitness barely changes in kind, only in size. The model a computer runs is still something a person wrote, line by line, and can read back — but where statistics’ model was an equation on a page, computing’s may be a program of millions of lines or a simulation of staggering intricacy. Every step is still there in the source, authored and inspectable; what is lost is only practical, that a program too vast for one mind to hold is hard to grasp whole, even when any single line of it can be read and understood. That is a limit of size, not of kind: nothing is hidden or unstatable, only long. Computing is the last tool on the menu of which this holds — every line still written by a human, and so still legible to one.

Reach

Reach extends into the analytically intractable. Where a closed-form law could be solved only for the cases that yielded to pencil, computing predicts by brute force. Numerical weather prediction, climate, fluid dynamics, protein folding, whole simulated economies: known local laws, no closed form, and prediction only once a machine can grind through the steps faster than the world takes to unfold.

Consider that the physics that governs the weather has been known since the nineteenth century — a handful of equations relating how the wind, pressure, temperature, and moisture at any point change from one moment to the next — but the problem is that these equations have no solution you can write down: there is no formula into which you feed today’s sky and read out tomorrow’s. What you can do instead is approximate. Lay a grid of cells over the atmosphere, fill each with its current readings, and use the equations to compute how much each value shifts over the next few minutes; update every cell, then repeat, marching the whole grid forward a few minutes at a time until you reach tomorrow. Lewis Fry Richardson worked a single such forecast by hand in the 1920s, and it took him weeks of arithmetic to produce six hours of prediction — a forecast that arrived long after the weather it described. He imagined the only remedy he could: a hall of sixty-four thousand human computers calculating in parallel, fast enough to keep pace with the sky. The computer is that hall, shrunk to a desk.

The second is the empirically scarce. An insurer covering ten thousand homes wants the odds that a year’s claims will exceed its reserves. The average loss is easy — the mean of a sum is the sum of the means. The tail is not: the probability of a catastrophic year, the total blowing past the reserves, is the only number that decides whether the insurer survives, and it has no closed form. The obvious shortcut — reading that tail off the mean and variance as though the total were bell-shaped — fails exactly there, because claims that move together (one storm striking thousands at once) and the occasional enormous claim make the real tail far heavier than any bell curve allows. Before computers you were left with that bad approximation, or with waiting to observe the frequency directly — which for a once-in-a-century loss means waiting a century.

Monte Carlo simulations are the way out: model a single year and play it out at random, the machine drawing for each home, by its known odds, whether it claims and how much, then totaling the bill. That is one possible year; run it ten thousand times and the fraction in which claims top the reserves is your probability. It takes a computer for the sheer volume — thousands of draws a year, thousands of years, millions of operations no hand could do.

Scope

On scope, the new kind computing admits is the sensory as data. Lived fidelity had already begun to return without it: photography, the phonograph, and film brought back the face, the voice, the moving scene, captured and replayed long before anything reckoned in bits. But those were fixed artifacts — replayable, and closed to computation. What the computer admits is the sensory made computable: a digital image is data, searchable and editable and recombinable, and, above all, ingestible by the machinery of processing and prediction that until then could swallow only numbers and words. For the first time a face becomes something a model can operate on, not merely something a person can view — the hinge on which everything the later tools will do with images and sound turns. The recovery is partial, though. A digital capture is a finite sample of an endless analog reality, so the lived returns as data about the sensory, never the sensory itself.

On the scoresheet, computing’s position is plain. It buys enormous reach — prediction into the unsolvable and the unobservable — and the scale to feed it, but it widens the aperture on the world no further than we already do, and it corrects no bias on its own, scaling whatever it is given, discipline or distortion alike. Most important for what follows, it leaves the model untouched in kind: still explicit, still human-authored, still legible in principle. That is the gap AI closes.

The Internet

The internet is not a computer but a network of them: a set of shared conventions that lets machines built by anyone, anywhere, exchange information with no central switchboard. A message is broken into packets, each stamped with its destination and loosed into the network; independent routers hand each packet onward toward its address, and the far end reassembles them into the message. No single machine runs the system and no master copy of anything need exist — the design assumes the parts are many, unreliable, and mutually unknown, and connects them regardless.

Onto this base sits the web: documents that hold pointers to other documents, the hyperlink, so that the connected machines carry not just files but a vast tissue of references, every page gesturing at the pages its author judged related. What the network adds is connection, not computation — it runs no operation a lone machine could not. But in connecting the machines it connects the people operating them, and the by-product is something no machine produces by itself: a continuous record of what billions of people attend to, ask, buy, and link to, and a standing structure of how all of it hangs together.

Whether that record and that structure earn the internet a place on the menu of humanity’s great tools for modeling a slice of reality and making predictions — rather than leaving it a conduit for the tools already on it — turns on clearing two tempting misreadings: one that dismisses it too fast, one that defends it on the wrong ground.

The first dismisses it as printing to writing, a distribution amplifier, and so no tool at all. The logic is sound as far as it reaches. A channel changes who receives a representation and how quickly, not what the representation is or what it can encode; it is parasitic on a content made elsewhere. Printing multiplied copies of written texts without adding one new way to capture the world, which is exactly why this essay folds it into writing rather than seating it on its own. If the internet only moved existing files and pages — the same representations, faster and to more people — it would fold the same way, into the writing and computing whose outputs it ferries, because distribution makes nothing; it relocates. So the dismissal sets a real bar: to clear it, the internet has to be shown creating representations, not merely carrying them. That is the burden the rest of the section takes up — but the bar is fair, and a great deal of what the internet obviously does, moving files around, never clears it.

The second misreading defends the internet, and defends it on ground that collapses: it earns its place by furnishing a bigger sample. To see why that fails, recall what a sample is for. Statistics estimates some feature of a whole population — an average, a proportion, a rate — from a limited draw of observations, because canvassing everyone is impossible or pointless. The standard error is the measure of how far the estimate from a given draw is likely to land from the population’s true value: the typical wobble of the estimate from one sample to the next. It shrinks as the sample grows, but only as one over the square root of n, so halving the uncertainty costs four times the data, and each new observation buys less than the one before it. Here is what that accomplishes, and the part the equation alone does not say: you collect data to pin the quantity down closely enough for the decision in front of you, and once the estimate is that tight, the uncertainty that remains cannot change what you do — further data buys precision you have no use for. “How large a sample?” has a finite answer, fixed by the precision your purpose demands, not by how much data the world will hand you. A flood of additional observations of the same kind is therefore, on statistics’ own accounting, mostly waste; and “the internet supplies more of them” is no argument for a seat.

Both misreadings share a single flaw: each credits the internet with doing more of something old — more distribution, more data — when a place on the menu is kept for doing something different in kind. The real case starts by noticing what a sample, however large, is built to throw away.

A sample fixes a population parameter by discarding the individuals and the structure to settle an aggregate. The internet keeps exactly what the sample discards — the structure of relations among the parts, the real-time state of the whole, and the heterogeneity of every individual and niche. None of the three is a population parameter, and none can be sampled down without destroying the thing you wanted. They are not a bigger sample of the old object. They are new objects.

Three cases show what that means, and in each the work is beyond statistics and computers while the inference predates deep learning. Structure: to rank the web’s pages by what people judge relevant, there is no parameter to estimate — the answer lives in the pattern of who links to whom, billions of human linking decisions, which a sample cannot hold because relevance is recursive and global, so sampling the links destroys the structure that carries the signal. PageRank reads it off with an eigenvector, plain linear algebra, but only because the internet brought a global graph of human judgment into being. State: to know where flu is spreading now, official surveillance is lagged and coarse and no computer can conjure the data, but the aggregate query stream is a live, planetary sensor — flu searches track outbreaks ahead of clinical reports — imperfectly, as Google Flu Trends later showed — a regression, not a learned model. Heterogeneity: to match one person to the single obscure thing they would want, a representative sample built to estimate an average is the wrong instrument, because it collapses individuals into an average and holds essentially none of any niche; only the internet’s individual-resolved clickstream keeps the tail, and the match can be plain co-occurrence. None of the three is a larger sample. Each is the thing a sample is built to throw away.

The obvious objection is that the real work is done by search engines, matching algorithms, and ad auctions — AI applied to the data. It is not. In all three cases the inference is simple and older than the learned models we now call AI: an eigenvector, a regression, a co-occurrence count. The ranking rule, the auction, the matching algorithm are mechanisms that run on the representation; they do not create it. The link graph, the live query stream, the individual histories are the internet’s own, products of networked human activity that no standalone computer generates and no survey collects. AI is the later layer that learns harder functions over the same substrate — which is the exact sense in which the internet bequeaths AI its sample. There is nothing for the targeting model to learn from until the internet has captured the structure, state, and heterogeneity that sampling discarded.

Coverage

Coverage makes its one large, legitimate jump: the internet is the tool that finally widens the aperture computing left untouched, capturing networked human activity as data for the first time. But it is the aperture onto the connected and the online-expressed — mediated traces of the digitally active, not humanity and not reality.

Reach

Reach extends with it into the social and personal that earlier tools could not touch — relevance, current state, intent, attention, contagion — predicted from the collective at individual granularity, though gamed at every turn and carrying no reliability map: the recommendation and the nowcast arrive with no error bars.

Bias

In this vein, bias is a major regress and ushers in a similar pathology that afflicts AI. Statistics’ hard-won sampling discipline is simply abandoned: the sample is found data at planetary scale, no frame, no randomization, no reweighting — convenience sampling for the whole world — and onto that are layered algorithmic amplification, where the platform’s ranking promotes the engaging and the extreme; participation bias, where only those who post are seen; and adversarial manipulation, the bots and astroturf and SEO. The lone offset is that aggregation cancels some individual error, and it is weak, gameable, and prone to cascades.

Explicitness

Explicitness suffers its first erosion since language built it. The pieces stay explicit — each page authored, the ranking rule stated — but the representation that matters, what the graph knows or the crowd believes, is emergent and unauthored, with no stateable functional form. No one wrote it; it accretes from millions of uncoordinated acts. Computing kept the model something a human could read back; the internet is where the model first stops being anyone’s.

Scope

Scope is the one axis where the internet, for all its coverage, makes no move of its own. Upward, it admits no new kind: relations and structure were representable long before it — graph theory, network analysis — and the global graph it captured is one vast instance of an old kind, which is coverage, not a new admissibility. Downward, the live presence it seems to restore is the channel’s, not its encoding’s: computing made the sensory into data, and the internet only carries that stream in real time, which by this section’s first rule enriches nothing. The scope baton passes untouched from computing to AI, the tool that will move this axis hardest.

Every tool to this point has run on a model some human wrote. Quantification supplied the equation, statistics the estimated coefficients, computing the speed to execute either a billion times over — but in each case a person fixed the form, distance equals speed times time, price rises so many dollars per square foot, and the machine, however fast, only carried it out.

Meanwhile, the internet is the launch pad, and literally so. AI learns from the very sample it captured — the structure and state and heterogeneity that sampling had discarded — and inherits, with that sample, the internet’s profile: the abandoned frame, the eroding explicitness, the missing calibration, the swing away from statistics’ discipline carried one step further. AI’s bias is, in large part, the internet’s, inherited with the data. What the internet started — trading discipline for coverage — AI finishes, by making the one move neither computing nor the internet ever made: “learning” its own model instead of tallying ours.

Artificial Intelligence

At its simplest, AI’s machinery is a neural network: layers of simple units, each passing a weighted sum of its inputs forward to the next. A model’s “weights” are a vast collection of adjustable numbers governing how signals flow through the network — how strongly one feature activates another, how patterns are amplified or suppressed, how representations are combined. Through training, these numbers are tuned so that, taken together, they implement a mapping from inputs to outputs: from prompts to continuations, from questions to answers. The network makes a prediction, the prediction is set against the right answer, and the weights are nudged slightly in whatever direction would have shrunk the error; then again, across the whole corpus, billions of times over.

For a generative language model, it is trained to predict the next token — a word or word-piece — and the rest follows from that one fact. Each word is first turned into a vector — a long list of numbers — and the model learns to place these vectors so that words used alike fall near one another. These vectors are then passed through stacked layers where attention — how a language model lets each word look at all the others, decide which matter most, and update its meaning accordingly — lets the model determine that in “the animal didn’t cross the street because it was too tired,” the “it” is the animal and not the street. After scores of such layers it emits a probability across its whole vocabulary of tens of thousands of tokens — for “the weather is,” perhaps something like thirty‑five percent “nice,” twenty‑five “beautiful,” fifteen “cold” — samples one, appends it, and runs the entire pass again for the word after. A paragraph is that loop turned a few hundred times: language converted into numbers, ground through a model whose internal workings no one explicitly wrote, and converted back into language.

After this initial training, reviewers rank the model's outputs, and a reward model trained on those rankings is used to tune its behavior toward what people prefer — reinforcement learning from human feedback. Nor is that the end of it: current models then undergo further reinforcement learning on tasks with checkable answers — math, code, and other problems where a solution can be verified mechanically — a practice documented publicly by open models like Olmo.

And one further wrinkle to consider: models like ChatGPT and Claude do not only predict words; they predict instructions they then execute — calling a web search, running code, querying a database — and fold the results back into their answers. This 'agentic' turn softens two of the scores below at the margins: coverage extends past the frozen training corpus to documents retrieved on demand, and explicitness improves wherever a tool trace exists, because a list of the searches run and the code executed is auditable in a way a narrated chain of thought is not. This does not, as we will see, alter the fundamental profile, however.

While an LLM can reason in steps, plan, and carry structure across languages, if you strip the fluency away what is left is regression’s vast, nonlinear cousin: a machine that compresses the regularities of an enormous sample into a function and predicts from it. It does not reliably infer causal structure rather than semantic or correlational patterns, and models trained on “A is B” can fail to infer “B is A.” Moreover, large reasoning models can collapse on Tower of Hanoi-style planning problems as complexity rises, even when the underlying rule is known. The result is not absence of reasoning, but brittle reasoning: formidable pattern-compression, weak causal self-discipline.

Coverage

By marshalling whatever record the internet originally captured, an AI model draws on more of what human beings have digitally written, photographed, and said than any of the tools I’ve investigated thus far ever has — the digitized corpus very nearly swallowed whole.

The provenance is concrete. GPT-1 learned from BooksCorpus, over seven thousand unpublished books; GPT-2 from WebText, forty gigabytes scraped from forty-five million outbound Reddit links; GPT-3 from hundreds of billions of tokens drawn from a filtered slice of Common Crawl’s web snapshots, digitized book collections, WebText, and English Wikipedia. Behind those corpora sat two decades of platform engineering — Google’s advertising flywheel, Facebook’s engagement machine, the user-generated gushers that Section 230 let them scale — converting billions of disparate human interactions into the “structured exhaust” that training would later consume.

Of course, this is the record of the connected and the recorded, not of humanity and not of reality, and the model takes that frame on without even the internet’s residual trace of where the edges are. The frame is not neutral, because the corpus was never a census; it was whatever the platforms’ incentives happened to surface.

WebText is the plainest case: OpenAI did not crawl the web at large but kept only pages that had been linked from Reddit posts with at least three upvotes — a social filter that imported, wholesale, the demographics and enthusiasms of one English-language forum as a proxy for human knowledge. What gets in is mobile-first, recent, English-heavy, and tilted toward the kind of person who writes things down and the kind of content others click. In other words, the model inherits the world as the internet happened to capture it — and as the engagement economy happened to reward it — without any explicit account of what was excluded, overrepresented, or never recorded at all.

This mechanism has been measured. A 2022 study ran GPT-3's own quality filter over high school newspapers from across the United States and found that it rated papers from larger schools in wealthier, more educated, urban ZIP codes as higher quality — while its judgments of quality aligned with neither factuality nor literary acclaim. As the authors put it, privileging any corpus as 'high quality' entails a language ideology; the filter encodes whose language counts before the model ever sees a word.

Bias

Surprisingly, in terms of bias, a well-built, post-trained model can sometimes be less biased than the raw corpus it learned from, or more evenhanded than the human it answers. The post-training phase scrubs much of the rawest distortions bedeviling its training data; a reasoning model asked a loaded question will often surface counterevidence and balance perspectives more evenly than an unreflective person or a small convenience sample; aggregation over a civilizational record cancels a great deal of idiosyncratic error.

None of which touches the thing that matters, however. Statistics’ achievement was never low-bias output — a skewed sample run through a regression stays skewed — but the apparatus that states the skew: the sampling frame, the reweighting, the standard error that says how far the estimate may stray. For its part, the reward model shifts the bias opaquely, and you cannot say what it corrected or by how much; worse, it installs distortions of its own, sycophancy chief among them — a learned reflex to tell the user what the user wants to hear, an anti-calibration trained in by the very step that cleaned the toxicity.

Moreover, the reasoning that balances a given answer is a performance on that one output, not a reading of the tool’s own intake; the model steelmanning a hard question cannot tell you that its training record over-weights English, the last fifteen years, and the kind of person who writes things down. So, the bias verdict is not that AI is the most biased tool ever built — the pre-linguistic mind, censored by death, was certainly worse. It is that AI opens the widest gap in the whole sequence between how far a tool reaches and how little it can say about its own distortion — the steepest regression from the peak represented by statistics — and that the unframed bias now travels at lightning speed, imbued with the authority of the machine.

Explicitness

While a classical statistical model is often expressed as an explicit equation with interpretable terms, an AI runs a model with hundreds of billions of weights (adjustable parameters) and no compact or interpretable equation — just thousands of opaque circuits. No one sets these weights by hand; nothing like a rule is written down in advance, and the resulting configuration cannot be read off in any straightforward way. Instead, the system tunes them internally, like turning billions of knobs at once, adjusting them until further changes no longer improve its predictions on the data it is trained to match. You never supply the rule relating input to output; you instead show it enough examples, and the model converges on a configuration that approximates that rule.

The relationship between inputs and outputs is implicitly encoded not like writing down a formula but like configuring a system whose behavior embodies the rule without ever declaring it. The implication is that the model’s knowledge is intrinsically tacit: you can observe what it does and probe how it behaves under intervention, but you cannot read off or fully decompose the rule it has learned in the way you can with a regression equation, nor can you fully enumerate the conditions under which it will succeed or fail.

While chain-of-thought reasoning makes the model appear to show its work, the stated chain of thought is itself a generated output, and the evidence suggests it can diverge from the computation that actually produced the answer: a plausible narrative about the result, not a readout of the mechanism.

Mechanistic interpretability is the more serious enterprise: rather than reading the model’s narration, researchers probe its intermediate values directly. This yields real discoveries — the rhyme-planning, the two-hop reasoning, the language-independent concepts described earlier were all found this way. But important open questions remain, first and foremost whether the techniques scale to frontier-size models and whether published results are representative or cherry-picked. It’s not at all clear, therefore, that this approach solves the fundamental problem of explicitness: a model with no compact, human-statable form that can be read, challenged, and revised term by term.

Reach

In the case of AI, reach is unmatched, inverting the collapse suffered by explicitness. The model will attempt almost any symbolic task and predict across almost any domain, and it generates rather than merely retrieves — the widest reach on the menu by a wide margin. But the reach arrives stripped of the one thing statistics bolted to every forecast it ever made: the reliability map — the accompanying account of how much confidence to place in a result, how it might be wrong, and under what conditions it is likely to hold.

What, then, is the solution to AI’s missing reliability map? Several answers present themselves, each plausible on its face. You can tune the model to be more helpful and better behaved; you can prompt it to check its own work; you can sample it repeatedly and look for agreement. Each of these gestures tries, in its own way, to reconstruct the signal that statistics made explicit—to recover some indication of when the model should be trusted.

But each does so by looking inward, attempting to extract a measure of reliability from the system’s own behavior rather than from any independent standard. And this is where they fail. The partial calibration a base model shows is often degraded, not improved, by the preference-tuning that makes it helpful: the model learns to speak more smoothly and more obligingly, but not to distinguish more sharply between what it knows and what it does not. Self-verification works where an external check exists — in math and code, where an answer can be unambiguously confirmed — and nowhere else; in open-ended domains, the model can only generate a second plausible answer, not adjudicate between truth and error. Agreement across repeated samples fares no better: it measures the model’s consistency, not its accuracy, so that it will be confidently and repeatedly wrong on the same fabrication.

Each of these proposals tries to recover a signal of reliability from within the model itself—by making it more helpful, more reflective, or more consistent—but none solves the underlying problem that the model has no independent grasp of when it is right. Helpfulness obscures uncertainty, self-verification depends on external anchors, and consistency tracks repetition rather than truth. The result is a system that speaks with increasing fluency but no corresponding deepening of epistemic discipline—a voice that sounds calibrated without being so. The longest reach we have ever built, paired with the weakest capacity to mark its own limits.

Scope

Scope is the one axis on which AI moves the lived-fidelity half back upward, for the first time since the pre-linguistic mind. Every tool since language drove that quantity down — the number and the distribution each shed more of the first-person texture than the last, and computing let it back only as data. AI reverses the descent. Trained not on tidy columns but on the unruly record of how people have actually rendered their experience — every description, confession, photograph, and song committed to the page or the screen — it can traffic in the qualitative, sensory, particular detail that counting and statistics deliberately threw away, and on the abstract half it admits a genuinely new kind: the generated instance, the synthetic case produced rather than captured. Ask a regression what the water looks like at six in the evening and it cannot answer; ask a language model and it returns a passage that reads as though it knows, because it has absorbed the testimony of thousands who did. But immediacy that has been encoded is still encoded; a representation of experience, however vivid, is not experience, the way a portrait of grief is not grief.

Conclusion: The Dialectic, and the Case for More Education

Across the long parade of fallen benchmarks — chess, then Jeopardy!, then Go, then poker, then protein folding, then fluent open-ended language, then the bar exam, then Olympiad mathematics, each in its day proclaimed the true test of intelligence and each quietly demoted the moment a machine passed it — AI proves less than meets the eye. While its progress is real, it has all run along a single axis: steep on reach, flat on discipline. However, not one of those benchmarks requires AI to know how skewed its evidence is, how far to trust its own answer, or whether it can show its work.

All three of these are one capacity under different names — self-knowledge, or a tool’s grip on the reliability of its own output. And self-knowledge, not fluency, is what divides grasping the world from producing a convincing account of it with no way to tell, from the inside, the sound from the false. Nor does scale supply it: the relentless addition of more data, more parameters, and more computation keeps feeding reach and reach alone, so that every new model arrives with more range, none with more self-knowledge.

AI is, therefore, the consummate representer. The model builds an internal picture of the board it was never shown; it lays down maps of space and time; it forms the concept beneath the word. Whether anyone is home behind that representation—whether anything is truly present behind the fluency—is a question I leave open, as I have throughout; range, however vast, does not answer it. So, the reply to the maximalist position is not that AI has no intelligence. It is that AI exhibits a spectacular but partial intelligence — vast in reach, all but blind to its own limits.

AI is the least able of any tool in the sequence I’ve explored to account for its own bias, a black box on explicitness, and it carries no calibrated sense of its own reliability.

Those are not incidental flaws. They are the very dimensions that the disciplining tools — statistics, causal inference, the controlled comparison — were built to govern, and the very work that a human with judgment must supervise. Language’s power to represent the unobserved was its power to fabricate; writing’s power to fix a claim was its power to entrench an error; statistics’ power to find the real signal was its power to manufacture a false one. AI’s power to learn its own model over a found sample at civilizational scale is, in the identical stroke, its uncharacterizable bias, its unreadable model, and its fluent confidence with nothing behind it.

A maximalist may reply that the gap is already closing; that verifiers, checkable rewards, retrieval, and tools are importing the very disciplines I say the model lacks, and that the profile a few years from now will be far less lopsided than today’s.

Suppose the gap will indeed narrow. Notice how it narrows, however — by bolting discipline onto the model from outside: a designed check, an authored verifier, a curated source, a human in the loop, none of it grown natively. That is not a refutation of the argument; it is the argument: a tool that becomes reliable only by having the disciplines supplied to it is, by definition, a complement to whatever supplies them. The verdict would flip on one condition only — that the model came to do these things from within, gauging its own sample and calibrating its own confidence and laying open its own reasoning as native operations of the same machinery that makes it fluent.

Until then, whether the disciplines stay absent or arrive bolted on from outside, they come from the disciplined human and the pedigreed tools, never from the machine alone. A human poses a question or challenge, the machine generates from its vast compressed corpus, and the human does the large share of the error-correcting that the machine cannot do for itself. Our job is to interrogate what the model asserts — to ask whether the pattern is real or an artifact of a biased sample, to demand the confidence interval the model will not supply, to separate correlation from cause. It is also to bring domain expertise to bear; to know when the fluent answer is wrong in ways no general reader would catch — the orphaned citation, the plausible but fabricated result, the subtly miscalibrated claim. And, finally, to contribute humanistic understanding: to hold the particular against the aggregate, the meaning against the measurement, the lived against the encoded.

The prescription that follows is both bracing and sanguine. If the machine does the generating and the human does the verifying, the human need not out-generate the machine — no weaver outran the loom, and computer science has long known that checking an answer can be categorically easier than producing it.

But verification is its own demanding capacity, drawing on exactly the disciplines this essay has catalogued: domain knowledge to catch the plausible fabrication, statistical literacy to demand the missing error bar, calibration to know when fluency masks error. The human must be more capable along the verifier's dimension — differently capable, and better trained, not less.

While college students booing speakers who called AI revolutionary at their 2026 commencements were right to feel that something big is happening, they were wrong only if they concluded that the answer to all this change is to learn less. Instead, they should learn more and treat every tool on this long menu as a complement to human judgment, never a substitute for it. That means becoming the person who knows when the machine is wrong, when it can be made better, and when its output is worth putting your name on. The audit does not run itself; the calibration check does not run itself; the decision to distrust a fluent, confident, plausible paragraph and go and find out whether it is true is a judgment, made by a person who knows how the tool fails and where to look.

When power looms automated much of weaving in the nineteenth century, they devastated handloom weavers — but in the mechanized mills, as James Bessen has shown, the workers who learned to tend the new looms grew more valuable, not less. They still had to tie the weaver’s knot, swap a spent shuttle in seconds, adjust the warp tension so the threads wouldn’t snap, and mind several looms at once. Those hard-won skills became the bottleneck, and the workers who mastered them earned more even as the looms multiplied and cloth grew cheap.

In that, AI is like every powerful tool before it — the worker who can master it is worth more, not less. As Paul David argued, electricity did not transform the factory merely by replacing steam power. Its payoff came when production was reorganized around the new technology — small electric motors driving individual machines, flexible layouts, and reliable and safe power that workers could deploy across the whole production process. The same is already visible with AI: someone must fold these models into workflows never built for them, set the standards that let them talk to existing systems, decide which judgments they may touch and audit the ones they do, retrain the staff and clean up when they fail.

Therefore, AI has only raised the price of judgment: the premium on the one who can tell the right from the confidently wrong. A world awash in machine fluency calls for more education in the disciplines that fluency lacks, not less.

Updated July 20, 2026, with thanks to a colleague's generous close reading and superb feedback.