The Iteration Gap
Structure before sophistication. Every time.
I. Masaccio to the Model
In 1425, a young painter named Masaccio finished a fresco in a small chapel in Florence called the Holy Trinity, and for the first time in the history of Western art, a wall appeared to open into a room that was not there. He had used a new mathematics of vanishing points and orthogonal lines — linear perspective, recently formalized by the architect Brunelleschi — to make flat plaster behave like depth. It took roughly two hundred more years before Gutenberg’s press, in the 1440s, made the reproduction of ideas cheap enough to outrun the reach of any single scribe or church. Another three hundred and thirty years passed before Watt’s steam engine, refined through the 1770s, converted heat into reliable mechanical work and detached human productivity from muscle and water for the first time in history. A little over a century later, in the 1880s, electricity began to arrive in homes and factories, and barely twenty years after that, the automobile put the average person in motion at a speed no one in history had traveled outside of a horse or a train.
Keep extending the list and the gaps keep closing. Powered flight, 1903. Nuclear fission demonstrated, the 1940s. The transistor, 1947. The commercial internet, the 1990s. The smartphone, 2007. The transformer architecture that underwrites every large language model in use today, 2017. Plot the distance between each entry against the year it happened, and the curve does not rise in a straight line. It bends upward, and over the last two centuries it bends hard — centuries between entries, compressing to decades, then to single digits. It is tempting, looking at that curve, to conclude something simple: that the rate of pivotal change in human history is accelerating, without qualification, without limit, and that whatever produced the transformer will produce the next pivot faster still. This essay takes that intuition seriously enough to test it — and finds that the honest answer is both more interesting and more useful than the straight line it seems to describe.
II. The Stalled Century
The clearest challenge to that straight line does not come from a skeptic of technology. It comes from an economic historian who spent a career measuring it. Robert Gordon’s The Rise and Fall of American Growth (Princeton University Press, 2016) makes a case that inverts the intuition entirely: the single largest transformation in the material conditions of ordinary human life did not happen recently, and it did not accelerate into the present. It happened between 1870 and 1970, and by any measure of lived experience, it has not been matched since. Electric light replaced darkness. Indoor plumbing replaced the well and the outhouse. Antibiotics replaced a world where a scraped knee could kill. The automobile replaced the horse; household appliances replaced hours of daily manual labor; radio and then television replaced silence. Gordon’s provocation is a thought experiment: take a person from 1870 and set them down in 1970, and they would be catastrophically disoriented — nothing in their material world would prepare them for what they’d see. Take a person from 1970 and set them down in the present day, and they would recognize almost everything around them. The car looks different but is still a car. The house has the same plumbing, the same wiring, the same appliances in kind if not in brand. The one object that would genuinely confuse them is the phone in every pocket.
Peter Thiel made a version of this same complaint in 2011, in a line his venture firm, Founders Fund, later adopted as the subtitle of its own investment thesis: “We wanted flying cars, instead we got 140 characters.” The line became the header of the firm’s 2017 manifesto, “What Happened to the Future?”, written by partner Bruce Gibney — a document arguing that a generation of the most capable engineers in the world had been diverted from aerospace, energy, and biotechnology into comparatively trivial software. Thiel’s favorite illustration of the complaint is concrete and checkable: with the retirement of the Concorde, the time it takes to cross the Atlantic by air went up for the first time in the history of transatlantic travel, rather than down. Gordon supplies the data underneath Thiel’s provocation. Between them, they describe the same shape from two directions — a hundred years of extraordinary material transformation, followed by fifty years in which the transformation seemingly moved almost entirely into one narrow domain: information.
This is where most commentary on the subject stops, and where it goes wrong. The standard explanation for the divergence is a story about talent and nerve — that engineers chose the easy, fundable, low-risk path of software over the harder, slower, more consequential work of building the physical future, and that with better incentives or braver founders, the stalled domains could simply have kept pace. It is a compelling story. It is also incomplete, because it locates the cause in a choice, as though a different generation, differently inspired, could simply have chosen otherwise. The deeper explanation isn’t about where talent went. It’s about why talent went there — and the answer is a variable most people never think to name: the cost of a single iteration.
III. The Iteration Gap
Consider what it actually takes to improve something. You try a version, you observe what happens, you adjust, you try again. Every domain of human progress runs on this loop, and every domain runs at a different speed, because the loop’s cost is set by physics, not ambition. A software engineer can write a function, compile it, and know within seconds whether it works. That loop — attempt, feedback, correction — can run millions of times a year, for pennies, on a laptop. A battery chemist testing a new electrolyte formulation waits weeks for a cell to cycle through charge states before the data even exists to evaluate. A drug developer waits years for a trial to reveal whether a compound does what the model predicted, because no simulation yet substitutes for a human body’s real response. These are not differences of degree. They are differences of kind.
This is the actual mechanism behind the great divergence, and it explains something the talent-flight story cannot: why the divergence compounded rather than merely persisted. Capital is not sentimental. It follows return, and return follows iteration speed — a domain where you can attempt and fail a thousand times a year will simply outcompete, dollar for dollar, a domain where you get three tries a decade. So capital flowed to software not because engineers were seduced by easy money, but because software was, structurally, the higher-yield bet — and once it began winning, it won more resources, which shortened its loop further, which won it more resources still. Meanwhile the slow domains fell further behind not because they were unfashionable but because they were, in the strictest economic sense, out-competed on cycle time. Gibney and Thiel saw the outcome and named a symptom — the Concorde’s retirement and the lengthening Atlantic crossing — but what they described as a failure of nerve was in fact a rational market sorting by a variable nobody had named.
Gordon’s data lands in the same place the iteration-cost argument predicts. The domains that transformed daily material life before 1970 are largely the slow-loop domains — energy, transportation, manufacturing, medicine — while the domain that has visibly compounded since is the fast-loop one: information.
But there is a second constraint beneath the first, and it is the one that matters for everything that follows. Software’s raw material is information: a bit, once discovered, is free to copy, free to test in simulation, and bound by no law of nature that limits how much of it can exist. The physical world obeys no such generosity. A battery’s energy density is bounded by the chemistry of the elements involved, full stop — no amount of clever iteration invents a loophole in the periodic table. A fission reaction proceeds at the rate the isotope allows. These are not engineering problems awaiting a better idea; they are walls set by thermodynamics, and no volume of intelligence, however fast it thinks, moves a wall it did not build.
This is the distinction the flying-cars complaint elides, and it is the one an honest reckoning with the singularity has to hold onto: iteration speed is a solvable problem — you can throw compute, simulation, and parallel experimentation at it, and AI is, right now, doing exactly that. Physical constraint is not solvable in the same sense; it can only be worked within. The question of whether the current acceleration crosses over from Gordon’s stalled material world into something genuinely general, then, is not a question about how smart the systems get. It is a question of how much of what still stands in the way is a slow loop waiting to be shortened, and how much is a wall.
IV. The Magic Mirror
The iteration-cost argument explains why capital chose software over hardware. It does not yet explain a further, more uncomfortable split — why capital, once inside software, gravitated so heavily toward a particular kind of software: the kind that captures attention rather than the kind that extends capacity. The same variable does the explaining. If iteration speed sets the ceiling on how fast a domain can compound, then engagement is the fastest-iterating substance ever built, because the feedback loop is not months or even hours — it is milliseconds. Every scroll, every hover, every half-second of hesitation before a swipe becomes a data point that retrains a recommendation algorithm in something close to real time. Nothing in energy, biology, or even productivity software iterates at that speed. Attention was never chosen out of malice; it was selected because it offered the tightest, cheapest, fastest loop capital had ever encountered — a loop faster than the one that produced the transformer.
The consequence is a lucrative asymmetry Tim Wu traced in The Attention Merchants (2016): the more precisely a platform can measure and monetize a fraction of a second of human attention, the more efficiently it converts that fraction into revenue, and the more capital it attracts relative to a slower, harder business trying to extend human capability rather than capture human focus. Frances Haugen’s 2021 disclosures to the Wall Street Journal, drawn from Meta’s own internal research, made the mechanism explicit rather than inferred: the company’s studies found that Instagram measurably worsened body-image issues for roughly one in three teenage girls who already felt bad about their bodies, and that safer algorithms consistently produced less time on platform and less ad revenue — a direct, internally acknowledged trade-off between user wellbeing and engagement-driven return. It is worth noting the finding is not without its critics; some researchers have pointed out that self-reported harm in survey data is a weaker form of evidence than clinical measures of wellbeing, and that other studies find more neutral or even positive associations for many users. But even the more skeptical reading concedes the core structural point Haugen’s documents revealed: the company optimized for the loop that returned fastest, and measured the human cost of that loop closely enough to document it internally, whether or not the aggregate effect fits a single headline.
This is the mechanism underneath what looks, from the outside, like vanity. The selfie, the follower count, the influencer economy — these aren’t best explained as a sudden collapse in collective character. They are the visible surface of a capital allocation problem: the fastest possible iteration loop happens to run on a very old piece of human wiring, the impulse to check one’s own reflection and ask, in effect, whether we are the fairest of them all — and the platforms most successful at monetizing that impulse are, structurally, the platforms fastest at learning to trigger it more often. Natasha Schüll’s research on slot-machine design, later adopted explicitly by technologists like Tristan Harris in describing his own former employer’s design practices, uses the same language the casino industry uses: variable-ratio reinforcement, the same unpredictable-reward schedule that makes a slot machine harder to walk away from than a vending machine. The mirror does not merely reflect. It has been engineered to answer back, unpredictably, just often enough to keep us asking.
This matters for the argument, not just morally but structurally. Section III established that capital follows return, indifferent to whether the domain it funds extends human capacity or merely captures it. That indifference is the same one Karp’s complaint gestures toward from the enterprise side and Gordon’s stalled-century data confirms from the historical side: markets are extremely good at finding and compounding the fastest loop available, and only accidentally good at ensuring that loop is the one civilization actually needs. Three hundred million people scrolling for an hour a day is, from a pure returns perspective, indistinguishable from progress — it produces exactly the revenue growth and exactly the engagement metrics that attract exactly the capital that a genuinely transformative slow-loop technology cannot compete for. The tragedy embedded in the flying-cars complaint is not that Twitter was funded instead of a flying car. It’s that the same iteration-cost logic that once built the printing press and the steam engine, now unmoored from any bottleneck but human attention itself, is just as capable of building a magnificently efficient mirror as it is of building anything that moves humanity forward.
V. The Testable Crossover Framework
If iteration cost is the real variable, then the crossover question — does AI’s compounding loop reach out and accelerate the physical world, or does it stay contained within information — becomes something better than speculation. It becomes a rule with a falsification condition: AI should compress timelines in proportion to how computational a domain’s bottleneck is, and should barely move a bottleneck that is physical or experimental. Run that rule against three live cases, and the pattern holds with useful precision.
Chips: the strongest case, with an honest asterisk. In 2021, Google published a Nature paper claiming a reinforcement-learning system called AlphaChip could generate chip floorplans — the physical arrangement of components on silicon — comparable or superior to human experts, in hours rather than the months floorplanning has traditionally required, and the method was reportedly used in Google’s own TPU accelerator chips. This is close to the purest version of the crossover thesis: chip floorplanning is a combinatorial optimization problem, almost entirely computational, with no experimental waiting period between attempt and result. It should be exactly what AI compresses fastest — and DeepMind’s account is that it has, with the technique now extended across the industry.
But the honest version of this case includes a real controversy: a 2023 meta-analysis by chip-design veteran Igor Markov argued Google’s results were not reproducible as published, and that the RL method actually underperformed simpler existing techniques — human designers, simulated annealing, and commercial EDA software — once methodology gaps were filled in independently. Google has stood by its results. The dispute remains unresolved in public. This is worth raising not to undercut the chip case but to sharpen it: even in the domain where the iteration-cost framework predicts AI should have the easiest time delivering a clean, verifiable win, the actual evidence is contested. That is worth sitting with before assuming the crossover is simple anywhere.
Biotech: the partial case, exactly as the framework predicts. AlphaFold, DeepMind’s protein-structure prediction system, is the clearest unambiguous AI success in a physical science — a 2024 Nobel Prize was awarded for its development, and DeepMind reports researchers using it have seen submissions of novel experimental protein structures increase by more than 40 percent. That’s a real, computationally bottlenecked problem — predicting a 3D shape from a sequence — genuinely solved. But the framework predicts this success should stop exactly where the bottleneck stops being computational, and it does: AlphaFold predicts structure, not function in a living system, and it carries no information about how a protein behaves in solvent, in the presence of drugs, or under the structural flexing real proteins undergo — meaning it accelerates the early “hit and lead generation” stage of drug discovery while leaving the actual clinical trial, which is bound by the slow, irreducibly physical process of watching a real human body respond, essentially untouched. Partial crossover, cleanly split along the line the theory predicts.
Energy: the weak case, and the reason SMRs still need a capex thesis rather than a software disruption story. Nuclear power is instructive precisely because it is not computationally starved — it is regulation- and construction-starved. As of 2026, only two small modular reactors operate commercially anywhere in the world, both with longer timelines and lower initial output than projected; in the U.S., the Department of Energy is putting $900 million behind SMR projects still targeting first power in 2030 or beyond, with most of the recent progress reported in the sector being permitting activity, supply-chain development, and site preparation — not the kind of bottleneck a faster model dissolves. Fusion researchers report real experimental progress, but by their own account the work remains incremental and engineering-heavy, still without a clear commercial end state. This is the case that most needs the physical wall, not the iteration loop: no amount of AI-driven design optimization shortens a multi-year regulatory review or accelerates the physics of plasma confinement. It’s the domain where a capex-and-infrastructure framing is doing the real work, because the constraint genuinely is capital and physics, not intelligence.
VI. What This Means for “The Singularity”
Taken together, these three cases refuse to answer the singularity question the way it’s usually posed — as a binary, will-it-or-won’t-it, coming-or-not. They answer a better question: where does the loop close, and where does it hit a wall it cannot move? Chips show a domain where the loop should close fastest, complicated by a live dispute over whether it actually has. Biotech shows a domain that closes exactly halfway, with the dividing line falling precisely where the bottleneck shifts from computational to physical. Energy shows a domain barely touched, gated by regulation and thermodynamics rather than cleverness.
This reframes I. J. Good’s original 1965 “intelligence explosion” concept — the idea that a sufficiently capable system improves itself recursively, compounding without limit — as something narrower and more useful than the popular version of the idea suggests. The recursive loop is real and already partially running: AI is measurably accelerating parts of AI research itself, which is the one domain where the entire bottleneck, top to bottom, is computational. But extending that loop outward into biology, energy, and materials is not automatic; it happens exactly to the degree those fields’ bottlenecks are informational rather than physical, and the evidence so far shows real but bounded crossover, not an unbounded one. The honest answer to “will the singularity happen” is: it is already happening, narrowly, inside the domains built entirely from information — and it is not yet clear how much further it reaches, because that depends on a wall we do not get to negotiate with.
VII. The Institutional Threat
Which brings the argument back to where it started — to a Palantir CEO on CNBC in July 2026, accusing the leading AI labs of running a token-pricing model that oversells capability while extracting enterprise IP; and to the deeper question underneath that complaint, of whether the AI market resolves into one dominant player or a stable, differentiated set of players. The singularity debate and the market-structure debate turn out to be the same debate, viewed from two altitudes.
If AI’s compounding loop stays largely confined to computational bottlenecks — the domain that includes code, language, and increasingly chip design — then the market structure question resolves the way software markets historically have: fast iteration, low capital intensity, rapid displacement, and real risk of the kind of superscalar dominance that swallows or forces merger among smaller players, because the loop that lets one lab pull ahead is exactly the fast, cheap, compounding kind. But if large parts of the economy remain gated by physical bottlenecks — energy, biology, manufacturing, the domains this piece has just shown resist the loop — then durable dominance is far less assured, because chip fabrication capacity, reactor capex, and fusion physics don’t compound the way a model weight update does. The market stays a mixed economy: monopoly risk in the fast-loop layers, stubborn capital intensity in the slow-loop ones. That is a very different world than either “one company wins everything” or “software eats the physical world too.”
The institutions built to govern this — antitrust frameworks, procurement rules, national-security review processes, enterprise contracts — were built for the 1870–1970 pace Gordon documented, when a pivotal shift took a generation to fully arrive and institutions had time to adapt around it. They are now being asked to govern a domain that, in its computational core, compounds on a timescale of months. That mismatch — not the abstract question of whether a hard takeoff someday occurs — is the threat that is actually live right now. Structure before sophistication applies as much to the institutions meant to hold this transition as it does to anything built beneath them.
Notes
1. Bruce Gibney, “What Happened to the Future?”, Founders Fund manifesto, 2017.
2. Robert J. Gordon, The Rise and Fall of American Growth: The U.S. Standard of Living Since the Civil War (Princeton University Press, 2016).
3. Tim Wu, The Attention Merchants: The Epic Scramble to Get Inside Our Heads (Knopf, 2016).
4. Jeff Horwitz, “The Facebook Files,” The Wall Street Journal, September 2021; Frances Haugen, testimony before the U.S. Senate Commerce Subcommittee on Consumer Protection, October 5, 2021.
5. Natasha Dow Schüll, Addiction by Design: Machine Gambling in Las Vegas (Princeton University Press, 2012); Tristan Harris, Center for Humane Technology.
6. I. J. Good, “Speculations Concerning the First Ultraintelligent Machine,” 1965.
7. A. Mirhoseini et al., “A Graph Placement Methodology for Fast Chip Design,” Nature 594 (2021): 207–212; Igor L. Markov, “The False Dawn: Reevaluating Google’s Reinforcement Learning for Chip Macro Placement,” arXiv, 2023 (disputes reproducibility of the above).
8. Google DeepMind, “AlphaFold: Five Years of Impact,” 2026; 2024 Nobel Prize in Chemistry awarded for AlphaFold’s development.
9. U.S. Energy Information Administration and American Society of Mechanical Engineers reporting on SMR and fusion status, 2026.
Structure before sophistication. Every time.