What the Renaissance Knew About AI: The Second Theft of Fire
The first time humanity stole fire, it took us four centuries to understand what we had. This time the cost of the fire is falling by two orders of magnitude every two years.
In the Philadelphia Museum of Art there is a painting we could not stop thinking about while writing this. Peter Paul Rubens began it around 1611 and finished it by 1618, and he was so concerned with getting one detail right that he subcontracted it. The eagle tearing into the liver of the chained Titan was painted by Frans Snyders, the finest animal painter in Antwerp. Rubens wanted a specialist for the part of the canvas that mattered most.
The painting is Prometheus Bound. The subject is the oldest story we have about technology.
Prometheus stole fire from the gods and gave it to mankind. Hesiod tells it first, around 700 BCE, in the Theogony and again in Works and Days: the Titan tricks Zeus, smuggles the fire down to humanity hidden in a fennel stalk, and for this gift is chained to a rock where an eagle eats his regenerating liver, forever. The play we attribute to Aeschylus dramatizes the punishment. Every version agrees on the structure. Someone takes a power reserved for the gods. The power transforms the species. The one who stole it pays.
We are a venture builder. We spend our days with founders who are, whether they would put it this way or not, in the fire business. And we have come to believe that the story of the next twenty years is not a new story at all. It is the second time the structure plays out at civilizational scale. The first time, we called it the Renaissance.
We are not the first optimists to reach for Prometheus. The canon we come from — the techno-optimist tradition, Marc Andreessen’s manifesto chief among it — already cast the technologist as the fire-bringer and the builder as the hero of the oldest story. But running the same myth back without adding anything is just cosplay. We are adding the parts tradition leaves out: the eagle, the receipts, and the forty years. The myth does not end at the gift. We follow it all the way to the rock.
This is an essay about that first theft, measured in numbers rather than nostalgia, and about why we think artificial intelligence is the same kind of event running a second time, compressed from four hundred years into roughly a decade. We are going to be optimists about it, but the rare kind who does not lie to you, including about the ways this time might genuinely be different. The fire is real. So is the eagle.
I. What the Renaissance actually was
Start with a correction, because the popular version of the Renaissance is mostly a tourism brochure. It was not, primarily, a sudden outbreak of genius — genius is always available; what changes is whether a society can fund it, reproduce its output, and distribute the result.
We are going to read the Renaissance through three coupled collapses: in the cost of representing reality, the cost of copying knowledge, and the cost of financing both. This is a chosen lens, not the whole story. A real historian would insist on the recovery of classical learning, the scholars who fled Constantinople in 1453, the rise of the tax-collecting state, gunpowder, and the navigation that opened the Atlantic — and would be right. We foreground the three collapses because they rhyme with what is happening now, not because they are the only causes. With that admitted: the artists got the statues, but they were standing on top of an infrastructure revolution, and it is the infrastructure that travels.
Take representation first, because it is the most beautiful and the least understood. Sometime around 1413, a Florentine goldsmith named Filippo Brunelleschi stood in the door of the cathedral and painted the Baptistery across the square on a small panel, using a geometric construction that made the painted building line up exactly with the real one when you looked through a hole in the back. The panels are lost; we know them through his biographer. But the method survived, and in 1435 Leon Battista Alberti wrote it down in De pictura, the first text in the Western tradition to describe linear perspective as a mathematical system. Alberti’s image for it is the one we still use five hundred years later: the painting is a window, a flat surface that becomes a space you could walk into.
This sounds like an art-history footnote. It is closer to the release of a new platform. Perspective and oil were not a cost collapse the way printing and finance were — if anything, van Eyck’s layered glazes were slower and dearer than the tempera they replaced. What collapsed was the cost of a capability: a reproducible, teachable, mathematical procedure for rendering the world as the eye sees it, where before there had been only the flat, symbolic, gold-ground conventions of Byzantine icons. (Van Eyck did not invent oil paint; that is a myth we owe to Vasari. He perfected it.) Once it existed, the procedure could be taught to anyone with a workshop, and a capability that had lived in a few hands became a substrate the whole culture could build on. Leonardo, born in 1452, would push it to the edge of the uncanny; Michelangelo, born in 1475, would put it on the ceiling of the Sistine Chapel between 1508 and 1512. They were not working miracles out of nothing; they were the first great applications running on a new platform.
The second collapse scaled the first. Around 1450, Johannes Gutenberg combined movable type with the press, and the cost of copying a page fell off a cliff. The numbers are the whole argument in miniature: a skilled scribe could produce a few pages a day, a print shop around 1600 up to roughly 3,600 impressions. By 1500, within a single human lifetime of Gutenberg’s press, printing had spread to 282 towns across Europe, and the presses had produced, by the most-cited estimates, on the order of twenty million books. Over the following century, output rose roughly tenfold again, into the range of 150 to 200 million. For thirteen centuries before printing, European book production had grown at about one percent a year. Then it didn’t.
And here is the fact that should make every founder sit up, because it is as close to a controlled experiment as economic history gets. The economist Jeremiah Dittmar, in the Quarterly Journal of Economics, compared European cities that adopted printing presses before 1500 to otherwise-similar cities that did not. Between 1500 and 1600, the printing cities grew sixty percent faster. He gets at causation with an elegant trick: a city’s distance from Mainz, where the technology was born, predicted how early it got a press but, he argues, had no other channel through which to affect its later growth. Take it seriously and the press produced a meaningful share of the divergence, not merely a correlation with it. It is a strong result for the genre, and like every result of the genre it rests on a premise specialists still contest. But the direction is not in doubt: cheaper reproduction of knowledge, compounded over a century, left a durable mark on the wealth of cities.
The third collapse is the one venture capitalists should find most familiar, because it is us. None of this funded itself: Renaissance Italy ran on a financial stack that was, for its time, as radical as the art. In 1494 a Franciscan friar named Luca Pacioli printed the Summa de arithmetica in Venice, and in it described double-entry bookkeeping, the “method of Venice.” Pacioli did not invent it; he was writing down what Italian merchants already did. But printing it turned a guild secret into a public technology, and double-entry bookkeeping is the cognitive tool that lets capital see itself: assets and liabilities, in balance, legible, auditable. The merchants of Florence had been running on it for a century.
They had also solved capital mobility. The bill of exchange let a merchant deposit money in Florence and have it paid out in Bruges in a different currency, with the lender’s profit hidden inside the exchange-rate spread between the two cities. This mattered enormously, because the Church banned lending at interest, and the bill let you earn interest without technically charging it. Cambium non est mutuum, the lawyers said: exchange is not a loan. It moved capital across a continent without moving a single gold coin, inside the letter of canon law — financial engineering of a very high order, and the bloodstream of the whole period.
And sitting at the center of it was the firm we would now call the prototype of everything we do. The Medici Bank, founded in 1397 by Giovanni di Bicci de’ Medici, grew into a network of around ten branches (Florence, Rome, Venice, Geneva and then Lyon, Bruges, London, Milan, and more) structured, in the words of the great historian Raymond de Roover, so that it “resembles nothing so much as the modern holding company.” Each branch was a separate partnership; the local manager put in capital and was paid in profit share, not salary, while the center held control. That is a venture portfolio’s shape. The Medici took the surplus that double-entry and the bill of exchange made legible, and they allocated it into banking, into politics, and, famously, into art. Cosimo de’ Medici funded the rebuilding of San Marco and, from the 1440s, endowed its library, often called the first public library of the Renaissance. His grandson Lorenzo recorded in 1471 that the family had spent 663,755 florins on buildings, charity, and taxes since 1434. He added that he did not regret a florin of it.
We want to be careful here, because this is the kind of figure that gets romanticized into nonsense. That number is not an “art budget”; it bundles buildings, alms, and taxes, and it is Lorenzo’s own accounting of his own family. And the motives behind it were not ones a modern allocator would recognize: Medici spending was driven by salvation anxiety (Cosimo sought a papal indulgence for the wages of usury, and built churches partly to buy down his time in Purgatory), by the Aristotelian ideal of magnificence that made conspicuous expenditure a civic duty, and by the politics of a family that ruled Florence without holding a crown. They were not chasing a return on the chapel. But the point survives the caveat, and is sharpened by it. A class of people who had grown rich on financial innovation chose to allocate an extraordinary share of that wealth into culture, knowledge, and beauty — for reasons of soul and status rather than spreadsheet — and the civilization we inherited is substantially the output of that decision. Not that the Medici were investors, but that capital, pointed at talent and patient enough to wait, compounds into things no quarterly mind would ever fund.
So that is the first theft of fire, with the romance scraped off: three coupled collapses — representation, reproduction, capital — that together carried Europe from the medieval to the modern. The artists were the visible miracle. The infrastructure was the actual one.
II. The fire is burning again, and it is burning faster
Hold that three-part structure in your head — representation, reproduction, capital — and look at what has happened since November 2022. The same three collapses are underway, and two of them are happening faster than anything in the historical record.
Begin where the parallel is most exact, which is reproduction. The printing press collapsed the cost of copying a fixed text. Generative AI is collapsing the cost of producing a non-fixed one, at a rate hard to believe even after you check it. According to Stanford’s 2025 AI Index, the cost of running a model at the capability level of GPT-3.5 fell from twenty dollars per million tokens in November 2022 to seven cents per million tokens by October 2024: a reduction of more than 280 times in under two years. The independent researchers at Epoch AI, looking across many capability thresholds, find that the price to reach a given level of performance has been falling somewhere between nine-fold and nine-hundred-fold per year, with a median around fifty-fold annually.
The 280× is the collapse in the cost of matching a fixed 2022 capability — a commoditization curve, the price of yesterday’s intelligence in free-fall. The frontier has also gotten cheaper, but far less dramatically, and the very best models sometimes cost more than their predecessors, not less. So this is not “intelligence got 280× cheaper” in some absolute sense; it is something stranger. A level of machine capability that was, in 2022, expensive enough to ration is now nearly too cheap to meter. That is what the printing press did to the book — not make the frontier of human thought cheaper, but make yesterday’s thought, the fixed and finished text, almost free to reproduce.
That earlier collapse, from a scribe’s months to a press’s hours, played out over decades as the technology spread to those 282 towns. The cost curve of machine intelligence is doing something of the same order of magnitude every year, and in its fastest segments faster: the printing press with the time axis compressed by one to two orders of magnitude. Adoption has tracked the cost. ChatGPT reached an estimated 100 million monthly users within about two months of launch, at the time the fastest-growing consumer product anyone could remember. By OpenAI’s DevDay in October 2025, it had 800 million weekly active users. The presses reached 282 towns in fifty years; this reached most of the connected human species in under three.
The first collapse, representation, is the subtler one, because it is the one the Renaissance is actually famous for. Perspective and oil gave humanity a new, reproducible substrate for rendering a convincing reality. That is, in a structural sense, what foundation models are: a new substrate for producing text, image, sound, code, and video that no individual hand produced. The analogy to perspective is at the level of the substrate — each, once installed, sits underneath an explosion of specific creation. Alberti gave painters a window; the model gives anyone a surface onto which language can be turned into artifact. What that move does to the people who used to do the rendering is the harder question, and we take it up in Section V; for now the point is only that the substrate is new. We are early in understanding what the great applications on it look like — at best in the years before Leonardo — but the platform is unmistakably installed.
The third collapse is, once again, capital, and once again it rhymes with the Medici. Private investment in generative AI reached $33.9 billion globally in 2024, more than eight times its 2022 level; total private AI investment in the United States alone hit $109 billion that year. The forecasters have put numbers on the prize that are hard to take literally: Goldman Sachs estimates generative AI could raise global GDP by around seven percent, roughly seven trillion dollars, over a decade; McKinsey puts the annual value across dozens of use cases at $2.6 to $4.4 trillion. Serious people think these numbers are too high, and we will give the most serious of them his say. But the shape is the Medici shape: a new class of capital, made legible and mobile by new financial technology, allocating itself at scale into the creators working on a new platform.
We are part of that class, so we will not flatter ourselves by claiming the Medici’s mantle; the comparison is theirs to lose, not ours to award. We claim only the job description. The Medici did not carve the David; Cosimo paid for the workshops, the libraries, and the long patient years in which a Donatello or a Brunelleschi could become himself. That is the job, and it is humbler than the myth makes it sound: not the genius, but the person who funds the conditions for genius and waits. It has not changed in six hundred years. Only the medium has — and, if we are honest, only some of us do it with anything like the Medici’s taste.
III. The eagle is also real
Now we have to talk about the eagle, because an essay that stops here is exactly the kind of optimism we promised not to write. The honest critique of the techno-optimist canon is that it cherry-picks: it tells the story of the loom and the tractor and the spreadsheet, all the technologies that displaced tasks and grew employment, and quietly skips the parts where the fire burned the people who carried it. We are not going to skip those parts, because the myth does not let us. Prometheus is chained to a rock. The gift and the punishment are the same story.
So here is what the first theft of fire actually cost, in the same currency of verified fact we have used throughout.
The printing press did not only carry literacy and science. It also carried the witch hunts. The Malleus Maleficarum, the witch-hunting manual first printed around 1486, became one of the early bestsellers of the press, running through 28 editions by 1600. A 2024 study in Theory and Society tracked witch trials across 553 cities in Central Europe over nearly three centuries and found that each new edition of the Malleus was followed by a rise in trials, spreading city to city like a contagion of ideas. That is an association, not a proven cause — witch-hunting had plenty of other engines, from confessional war to weak local courts — but the diffusion pattern is hard to unsee. The European witch hunts killed, by scholarly consensus, somewhere between 40,000 and 60,000 people. The same machine that carried Copernicus carried that. The press was not, in itself, a force for enlightenment; it was a force for replication, and it replicated whatever was put into it, including the worst of us. That should sound uncomfortably current to anyone watching what large language models do with the contents of the internet.
It also armed the Reformation, and through it, a century of war. The economist Jared Rubin showed that cities with a printing press by 1500 were at least 29 percentage points more likely to have turned Protestant by 1600. Luther’s pamphlets sold more than 300,000 copies between 1517 and 1520. We will not blame the Thirty Years’ War on Gutenberg, since the chain from a printed pamphlet to a battlefield runs through too much politics to be laid at the feet of a machine, but the machine was the necessary condition. Cheap reproduction of ideas does not discriminate between the ideas worth reproducing and the ones that get people killed; it is a cost multiplier on whatever a society already believes.
And the Renaissance did not lift all boats. It concentrated wealth, nearly permanently. In one of the most remarkable studies in all of economics, Guglielmo Barone and Sauro Mocetti took the Florentine tax records of 1427 and matched the surnames to taxpayers in Florence in 2011, across roughly twenty generations and six centuries, and found that the families who were rich then are still measurably richer now. The fortunes of the Renaissance did not disperse; they compounded, down the bloodlines, past the fall of the Medici, past Napoleon, past the unification of Italy, into the twenty-first century. The fire warmed some hands far more than others, and it kept warming them for six hundred years.
We are going to argue, a few sections from now, that these fires also widen participation — that more people end up able to make things that matter. Hold both against each other, because they are both true and they are in tension: the same revolution can broaden who gets to create while concentrating who gets to keep the proceeds. An honest optimism does not resolve that contradiction by pretending one side away. It lives inside it.
We include all of this for a reason beyond honesty for its own sake: the contemporary eagle is being measured right now, and a serious person has to hold it next to the seven-trillion-dollar number. The International Monetary Fund estimates that around 40 percent of jobs globally, and about 60 percent in advanced economies, are exposed to AI. The study by Eloundou and colleagues at OpenAI, published in Science, found that around 80 percent of the U.S. workforce could have at least 10 percent of their tasks affected by large language models, and about 19 percent could see at least half their tasks affected. Be precise about a word the optimist and the doomer both abuse: exposure is not job loss. The IMF is careful to say that roughly half of exposed jobs may be augmented rather than replaced, and the Eloundou paper measures which tasks could be touched, not which workers will be let go. But “exposed” is not “safe” — the disruption will be real, it will fall unevenly, and the people it falls on will not be comforted by a chart showing that the loom worked out fine in the end.
And there is a skeptic who thinks the whole prize is far smaller than Goldman and McKinsey claim. Intellectual honesty requires us to hand him the microphone and let him actually talk, not just wave at him. Daron Acemoglu of MIT, one of the most cited economists alive, modeled the macroeconomic effect of AI and concluded it would raise total factor productivity by no more than about 0.66 percent over ten years, and less than 0.53 percent once you account for the fact that the hard tasks get automated last. That is a rounding error next to Goldman’s seven percent. His method is disciplined and unglamorous: only about a fifth of work tasks are even exposed to AI, only a fraction of those are profitably automatable this decade, and the cost savings on each are modest, so the macro arithmetic cannot get very large. And he goes further in a way the optimists never quote: AI also creates bad new tasks — deepfakes, manipulation, addictive feeds, automated fraud — which he estimates could pad measured GDP by around two percent while actually reducing human welfare by roughly three-quarters of a percent. His AI is not a smaller version of ours. It is a different object, one that can grow the numbers on the page while making life worse.
We think he is wrong, and we owe you the reason: a forecast that cannot say why it disagrees with its best opponent is not a forecast, it is a sales pitch. The weakness in Acemoglu’s case is structural, and Philippe Aghion (a Nobel laureate himself) has named it: the task-by-task accounting only measures AI doing existing jobs more cheaply. It assigns essentially nothing to the creation of entirely new tasks, products, and industries — and nothing to AI accelerating the production of ideas themselves, the input to all future growth. That is the whole engine of the very analogy this essay is built on. The printing press’s measured “task automation” was the scribe; its actual contribution was the university, the journal, the newspaper, the scientist — categories of work that did not exist at the old price and that no task-level model in 1500 could have counted. Aghion’s own replication, with the idea-production channel switched on, puts the plausible TFP contribution an order of magnitude above Acemoglu’s. Acemoglu is the most rigorous accountant of the world AI leaves unchanged. We are betting on the world it brings into being — exactly the part his method, by construction, cannot see.
IV. The forty-year warning
Which brings us to the hardest and most useful piece of history we know, the one we would staple to every breathless AI deck including, sometimes, our own.
When electricity arrived, it did almost nothing for forty years.
Generating stations were lighting New York and London by 1881. And yet, as the economic historian Paul David documented in his famous essay “The Dynamo and the Computer,” the productivity payoff did not arrive until the 1910s and 1920s — roughly four decades later. The hard data is starker than the story: the economist Warren Devine found that electricity supplied less than 5 percent of American manufacturing horsepower in 1899, only about 50 percent by 1919, and roughly 75 percent by 1929. For a generation, factories had electric motors but used them as drop-in replacements for their old steam engines, bolted to the same central drive shaft. The gains only came when a new generation of managers tore up the factory and redesigned it around what electricity actually made possible: small motors at each workstation, machines arranged by the logic of the work rather than the geometry of the drive shaft. The technology was ready in 1881. The organization took forty years to catch up.
This is the pattern economists named “general purpose technologies” — steam, electricity, the computer — and they all share it. Robert Solow won a Nobel Prize partly for noticing, in 1987, that “you can see the computer age everywhere but in the productivity statistics.” (The quote is from The New York Times Book Review, not the New York Review of Books, despite what half the internet will tell you; an essay this concerned with verification should get its own epigraphs right.) The lesson is not that the technology fails — electricity did not, the computer did not. The lesson is that the fire and the new order it enables are separated by a long, frustrating gap, filled not with better technology but with painful human and institutional reinvention.
We tell you this against our own interest. If you are an entrepreneur and your pitch is that AI will transform your industry in twenty-four months, the history says: probably not, and the people promising it are selling the dynamo bolted to the old drive shaft. The transformation is coming. The timing is the thing the optimists get wrong, every single time, in the same direction. Plan for the forty years even as you build for the next two. And notice what that forty years is made of: not better technology, but human beings reorganizing themselves around it. Hold onto that, because it is about to become the crux of the whole argument.
V. The objection we are most afraid of
Here is where an honest version of this essay has to stop and confront the thing that could break it. Every analogy we have drawn rests on a single load-bearing assumption. We chose the printing press as our central fire — and the printing press is the safest fire we could have chosen, because it only ever replaced the scribe, never the author. It collapsed the cost of copying a human thought; it never once threatened to produce the thought.
AI’s defining move is precisely the one the printing press never made. It does not just reproduce judgment more cheaply; it generates judgment. And “the work moves up the stack to higher-order tasks” — the optimist’s reflex, the one we are about to lean on ourselves — assumes there is always a higher stack for displaced humans to climb to. The Renaissance escalator always had a higher floor because the machine could not think: the scribe could become a proofreader, an editor, a scholar, precisely because the press couldn’t do those things. If this fire can climb the stack faster than people can, the escalator runs the wrong way. That is the real “this time is different,” and it is not a doomer’s fantasy — it is the one objection that, taken seriously, dissolves the happy ending we keep borrowing from history.
So we are not going to borrow it. We are going to argue for it, on AI’s own terms, and tell you where our confidence stops.
We think the higher floor is real, for three reasons — one structural, one historical, one we are watching happen in real time. We will name which is which, because that honesty is the difference between an argument and a hope. The structural reason first, and it is the load-bearing one: the work hardest to automate is not the highest-skilled, it is the least specifiable. AI is extraordinary at tasks you can describe and verify, and clumsy at the ones where defining the goal is the job — deciding what is worth doing, holding responsibility when it goes wrong, earning another human’s trust, exercising taste under genuine uncertainty. Those are not a thin residue at the top of the stack; they are most of what a senior person in any field actually does, and they expand to fill whatever the machine clears beneath them. The historical reason, which we hold more lightly because it is the very induction this section exists to question: every prior fire enlarged the set of things worth doing faster than it automated the old set, and the constraint on human work has almost never been a shortage of tasks but a shortage of cheap enough tools to make new ones viable. And the third reason, the one we trust most because we are watching it rather than predicting it: the early evidence on generative AI in real workplaces shows the productivity gains landing largest on the least experienced workers — the novices, the ones with the least to fall back on — which is augmentation pulling people up the stack, not automation knocking them off it.
And here is where our confidence stops, because the same evidence that gives us the floor names its own limit. That novice-lifting result comes from one rigorous study, in one occupation, over one stretch of time. The aggregate is not the transition: even if the new higher-stack work exists, the scribe is not automatically retrained as the editor, and the gap between losing the old job and reaching the new one is measured in exactly the painful, generation-long human reorganization the last section described. The distribution is not guaranteed to be kind — Acemoglu’s warning that these gains can concentrate while wages stagnate is live, and the Florentine six centuries are a long shadow. So our optimism is real but bounded: we believe the higher floor exists, we do not believe people reach it automatically, and the entire policy and institutional task of this era is shortening the climb. That is a narrower claim than “it’ll be fine because it was fine before,” and the only version we can actually defend.
VI. Why we build anyway
Two more reasons hold us on the optimist’s side of the line, beyond that bounded case for the floor — and they are not the reasons the brochure gives.
The first is about the shape of history, and we offer it with a caution bolted on. The first Renaissance did not emerge from comfort; it emerged from a system under catastrophic stress. The Black Death killed something between a third and half of Europe between 1347 and 1351, and the labor scarcity it left behind — fewer hands, more land and capital per surviving worker — helped crack the rigid medieval order open and, over the following century, roughly doubled real wages across much of Europe. The caution: a pandemic is not a technology, and we will not dress forty percent mortality up as a tidy instance of Schumpeter’s “perennial gale of creative destruction.” He coined that phrase in 1942 for market-driven churn, not plague, and the durable wage gains were strongest in northern Europe, not the Italy of the Medici. The honest, narrower version still holds: renewal at civilizational scale has tended to follow disruption no one would have chosen, and to be legible as renewal only from the far side. That should make us humble about the second renewal, not triumphant.
The second reason is the one the cost curves make almost arithmetic, and it is the deep logic under the higher-floor argument. When the price of a fundamental input collapses, demand rarely stays fixed and lets employment fall; it expands, and the expansion tends to create more work, at a higher level, than it destroyed. When books got cheap, Europe did not read the same number with fewer scribes. It read orders of magnitude more, built universities to house them, and invented whole professions — editors, publishers, journalists, scientists — that could not have existed at the old price. Literacy in Western Europe roughly doubled between 1500 and 1700, not because people changed but because the price of the written word did. The mechanism that makes cheap intelligence frightening — that it does so much more for so much less — is the same mechanism that, run at this scale before, has enlarged rather than shrunk the space of valuable human work. That is not a law of nature; it is the way the bet has paid every previous time it has been run.
And beneath both sits a third thing, the one we believe most and can prove least, so we will mark it as conviction rather than dress it as fact. Every previous theft of fire widened the circle of people who could participate in creation. Perspective and the press took the making of knowledge and beauty out of the monastery and the guild and handed it — unevenly, incompletely, but really — to a far larger share of humanity; the Medici pointed capital that had been locked in land and bloodline at raw talent instead. Each of these revolutions, underneath the violence and the inequality and the forty-year lags, bent toward a world where more people got to make things that matter. This is the same arc we flagged the tension in earlier: it widens participation even where it concentrates proceeds, and both happen at once. We are betting that AI is the most powerful instrument yet built for the participation half of that arc — for taking the capacity to create, which has always been scarce and gated, and making it abundant and open. We could be wrong. The eagle is real and we have shown it to you. But we would rather be among the people who carried the fire down and bore what came with it than among the ones who left it on the mountain because the gods said no.
VII. Down the mountain
We opened with a painting of a Titan in chains. We should be honest about why that image, and not a triumphant one, is the right emblem for the work.
Prometheus is not punished for failing. He is punished for succeeding — for giving humanity a power the gods wanted to keep. The chains are not the opposite of the gift; they are the price of it, paid knowingly, because the gift is worth it. That is the founder’s bargain in its oldest form. But here we have to be more honest than the myth flatters us into being, because it has a convenient asymmetry and so does the canon we come from. In the heroic telling, the fire-bringer is the one chained to the rock, nobly bearing the cost. Reread the actual story and the arithmetic of this essay together, and it does not hold: in Hesiod, the lasting cost of the fire does not fall on Prometheus at all — it falls on humanity, through Pandora and the jar of evils that the gift came bundled with. And the contemporary eagle, by our own figures, does not feed on the people who fund the fire. It feeds on the worker whose tasks are exposed, the cohort caught in the forty-year gap, the family whose trade the cost curve erased. The venture builder is not the one in the chains; we are closer to standing on the mountain, watching. An honest version of the bargain does not get to cast itself as the noble sufferer. It has to say plainly: the fire is worth bringing down, and the cost of it lands mostly on people other than us, and that is precisely why the responsibility for how it lands is ours to carry and not to wave away. Rubens hired a specialist for the eagle because he understood the punishment was half the painting. We do not get to paint only the fire.
The Renaissance was the first time we did this at the scale of a civilization, and we have spent five centuries living inside the world it made — richer, more literate, more capable, and more unequal, founded on a catastrophe and carried by a machine that replicated our worst alongside our best. It was worth it. Almost no one who lived through the plague and the wars and the witch trials would have been able to say so, and we have the privilege of saying it only because we are standing on the far side of the four hundred years. That privilege should make us humble about the second Renaissance, not triumphant. The people living through the disruption rarely get to see the renaissance. They just get the disruption.
But the fire is already burning, the cost curve is not going to reverse, and the only real choice in front of the people reading this is what they feed into the flame. The press reproduced the Malleus and Copernicus with perfect indifference; the model will do the same with whatever we give it. The technology has no opinion about whether it carries the witch hunt or the Scientific Revolution. The people building it do not have the luxury of that neutrality. That is the whole responsibility of building in this decade, and it is the one thing in this essay that shows up on no cost curve and in no forecast.
So we will not end the way the optimists we came up admiring end — with a borrowed slogan and an open invitation, as if the water were warm and the only question were whether you are brave enough to get in. The water is not simply warm. It is a fire, the thing in the old story that you carry down the mountain knowing it will cost something, and knowing the cost will not fall first or hardest on you. The second theft is already underway. It will be faster than the first, and stranger, and the gap between the spark and the new world will feel like forever even though, by the clock of history, it will be an instant. There is room on that mountain for the people willing to carry the fire down and answer for what it does on the way. The prize is real and the eagle is real, and the builders worth following are the ones who have looked steadily at both and started down the slope anyway, with their eyes open and the weight understood.
That is the work. It is the oldest work there is. It is time to do it honestly.
— Ogulcan Ozyavuz · Partner, DICE Venture Builder
A note on sources and method
This essay makes a deliberately aggressive argument, and we hold ourselves to the rule that an aggressive argument has to be an honest one. Every quantitative claim above is drawn from a primary or peer-reviewed source and was verified before publication; where a figure is contested, self-reported, or an estimate, we have said so in the text rather than smuggling certainty we do not have. The historical cost-and-harm claims — the witch-trial diffusion, the Reformation effect, the six-century persistence of Florentine wealth, the forty-year electrification lag — are included precisely because the techno-optimist tradition tends to omit them, and an argument that omits its own best counter-evidence does not deserve to be believed.
Key sources. On the printing press: Buringh & van Zanden, Journal of Economic History (2009); Dittmar, Quarterly Journal of Economics (2011); the British Library Incunabula Short Title Catalogue. On Renaissance finance and patronage: Raymond de Roover, The Rise and Decline of the Medici Bank (Harvard, 1963); Richard Goldthwaite, The Economy of Renaissance Florence (Johns Hopkins, 2009); Luca Pacioli, Summa de arithmetica (1494). On Renaissance art: the Vatican Museums; the Musée du Louvre; the Philadelphia Museum of Art; the Metropolitan Museum’s Heilbrunn Timeline. On AI cost, scale, and economics: Stanford HAI, AI Index Report 2025; Epoch AI; Eloundou, Manning, Mishkin & Rock, Science (2024); Brynjolfsson, Li & Raymond, Quarterly Journal of Economics (2025); Goldman Sachs Global Economics (2023); McKinsey Global Institute (2023). On the costs and the cautions: Rubin, Review of Economics and Statistics (2014); Doten-Snitker et al., Theory and Society (2024); Barone & Mocetti, Review of Economic Studies (2021); Pamuk, European Review of Economic History (2007); David, American Economic Review (1990); Devine, Journal of Economic History (1983); the IMF (2024); Acemoglu, Economic Policy (2025), with the new-task-creation rebuttal from Aghion & Bunel, AI and Growth: Where Do We Stand? (Federal Reserve Bank of San Francisco, 2024). On the myth: Hesiod, Theogony and Works and Days; Aeschylus (attributed), Prometheus Bound.
This essay was researched and drafted with the assistance of AI tools, with all factual claims independently verified against the sources named above.