The New Renaissance: What the Printing Press Actually Did, and What Open Models Are Doing Now

History does not repeat itself. What repeats is a mechanism, and it is narrower and more interesting than the version people quote on conference stages.

The mechanism is this: every so often, the cost of reproducing knowledge collapses. Not the cost of having an idea — that was never the bottleneck — but the cost of making the second copy, and the thousandth, and putting them into hands that were never meant to hold them. Three things then follow, in order. The institutions built around scarcity adopt the new tool faster than anyone expects. Once they see what it actually does, they try to govern it. And in the gap between those two moves, an unusual number of people build things they had no standing to build.

We are somewhere in the middle of that sequence. The fifteenth century is the closest thing we have to a control group, and it is worth reading properly — because the popular version of that story is wrong in the specific ways that matter most for ours.

What the printing press actually did

The usual telling goes: Europe sat in an intellectual dark age, knowledge was locked away in monasteries, Gutenberg invented printing, the Church panicked, and the modern world fell out of the wreckage. Almost every beat of that is wrong, and the accurate version is more useful.

Europe in 1450 was not dormant. Bologna had run a university since 1088, Paris since around 1150, Oxford since before 1100. Petrarch, usually credited with starting Italian humanism, died in 1374 — seventy years before Gutenberg printed anything. The Renaissance did not begin with the press. It was already underway, and the press arrived as an amplifier for a movement that already had somewhere to go.

That correction is not pedantry. A reproduction technology does not create a culture; it scales whatever culture is already standing. Ideas that had audiences got much larger ones. Ideas with no audience printed just as cheaply and went nowhere.

The second correction concerns the Church, usually cast as the panicking villain. It was an early and enthusiastic customer. Among the first dated products of Gutenberg's press at Mainz are printed indulgences, run off in 1454 and 1455 — before the 42-line Bible was even finished. The institution with the most to lose from cheap reproduction looked at the machine and saw throughput.

Suppression did come, but late, and as a reaction to what the press had already done. Luther's theses circulated in 1517; the first Roman Index of Prohibited Books appeared in 1559. A century separates the invention from the systematic attempt to control it — and that attempt answered the Reformation rather than anticipating it.

Adopt first, govern later, and govern the last crisis rather than the next one. That is the real pattern, and it predicts far better than "the incumbents panicked."

The third correction is scale. Books printed in Europe before 1501 are called incunabula; roughly 30,000 distinct editions survive, from presses in more than 250 towns. In fifty years, one workshop became a continental industry.

And then — the part that almost never gets told — it stopped being free. Venice began granting printing privileges in 1469, handing one printer a five-year monopoly on printing in the city. Privileges spread; by 1710 England had the Statute of Anne and the world had copyright. The window in which cheap reproduction was also unencumbered reproduction lasted about a generation.

Hold on to that one. It is the most load-bearing part of the analogy.

The modern equivalent is narrower than "AI"

"AI is the new printing press" gets said constantly, and it is usually said about the wrong thing.

Calling a hosted chat assistant a printing press has it backwards. A metered API — where the weights sit on someone else's hardware, the terms can change under you, and the model you built on can be deprecated on a schedule you do not set — is structurally closer to the scriptorium than to the press. The capability stays with the provider. It is an extraordinarily good scriptorium, and it is still a scriptorium.

The press analogue is possession. Gutenberg's break was not that reading got easier. It was that a book became an object you could own, carry, copy, and keep after the person who sold it to you stopped liking you.

The thing in modern AI with that property is the open-weight model: a file you download, that runs on hardware you own, that keeps working with the network unplugged, and that nobody can revoke. Llama, Qwen, Gemma, Mistral, DeepSeek — weights on disk. When people say the cost of creation collapsed, that is what collapsed, and it is a different claim from "chatbots got good."

It arrives with the same asterisk the press did. Open weights are not uniformly free. DeepSeek published R1 — a 671-billion-parameter mixture-of-experts model — under MIT. Qwen 3 ships under Apache 2.0. Llama's community licence carries a monthly-active-user threshold above which you need separate permission from Meta, and a handful of well-known models are research-only, forbidding commercial deployment outright.

That is not a scandal. It is Venice in 1469, arriving right on schedule. The privileges always show up. Which is why the licence, not the benchmark score, is the first field worth reading — and why every model in the model library carries its actual licence with the commercial-use verdict spelled out, rather than a blanket claim that the whole catalogue is "open source."

What it costs to own a copy

Here the essay has to stop being an essay, because the genuinely new thing about this shift is that it is checkable.

A model on disk is not an abstraction. It is a specific number of gigabytes that either fits in your graphics card's memory or does not. Quantization — storing each weight in four or five bits instead of sixteen — shrinks that file three- to fourfold at some cost in quality, and is most of the reason any of this is a consumer story at all.

But the arithmetic is unforgiving, and it is where the excitement quietly breaks. Some models sit comfortably on a mid-range card. Some need the best consumer GPU made. Some need two of them, or a server, and no amount of enthusiasm changes that. Mixture-of-experts models are stranger still: they must hold every parameter in memory but read only a fraction per token — which is why a very large model can run surprisingly fast and still refuse to load at all.

This is the honest form of the democratization claim. Not "anyone can run anything now" — that is false, and most people repeating it have not checked. The real claim is smaller and much better: the constraint moved from permission to arithmetic. Nobody has to approve you. You just have to have enough memory.

You can check yours in about a minute. The GPU and VRAM checker takes your card and a model and returns a verdict; the per-model VRAM calculator shows the working. If the answer is no, the local vs cloud cost calculator prices the alternative. Every formula behind those numbers is published, because a number you cannot trace is just a rumour with a decimal point.

That is a mundane paragraph, and it is the whole point. Around 1470, "can I own a copy of this book?" became a question answered in money rather than status. In 2026, "can I own a copy of this model?" is answered in gigabytes. Both are enormous downgrades in the power of gatekeepers. Neither is magic.

The specialist was an industrial invention, not a medieval one

The second half of the argument — that the age of the narrow specialist is ending and the generalist is returning — is right. It is just usually told with the wrong villain.

The Middle Ages were not an age of hyper-specialization. Guild training produced close to the opposite: an apprentice became a journeyman became a master, and a master made the whole object. Medieval society was rigid about status — who you were permitted to be — but a craftsman's work was broad. The narrowing came much later, and we can date it.

Adam Smith opened The Wealth of Nations in 1776 with a pin factory, where making a pin had been split into about eighteen distinct operations performed by different hands. Taylor turned that observation into a management system in 1911; Ford put it on a moving line in 1913. The twentieth century then built its schools, careers, and job titles on top of it.

So the specialist is roughly 250 years old, not a thousand — which changes what is actually ending. Not some primordial human condition. A specific industrial bargain: you become a reliable component, and in exchange the system guarantees you a slot.

The Renaissance polymaths are the counterexample precisely because they predate that bargain. Brunelleschi trained as a goldsmith, engineered the dome of Florence's cathedral across sixteen years, and in 1421 received what is often called the first modern industrial patent — for a boat. Alberti wrote the founding treatise on perspective, another on architecture, and invented the cipher disk that made him a father of Western cryptography. These were not people with unusual permission. They were people working before the walls went up.

When the routine execution layer gets cheap — and that is precisely what has just happened — the bargain stops paying. What remains is the part that was never a component: deciding what is worth building, and recognising when the output is wrong.

Taste is the scarce input now

"Taste" sounds like a soft word for a hard shift, so let me make it concrete in the domain this site actually lives in.

Running a model locally involves a chain of judgments no tool makes for you. Which model family suits this job. Which size. Which quantization, and how much quality you will trade for the memory. Whether the context window you need survives on your card once the KV cache is accounted for. Whether the licence permits what you intend to do commercially. Whether a benchmark that looks decisive was measured on your workload or on somebody else's.

Every one of those is answerable. None is automatable, because each is a trade-off, and a trade-off needs someone who knows what they are optimising for. That is taste: not aesthetic preference, but a calibrated sense of what matters here.

The same shape appears wherever the tools got cheap. Anyone can generate a thousand words or a working script. Almost nobody can say which thousand words were worth generating, or which of that script's assumptions will fail in production. The scarce skill was never production. It is judgment — and judgment just became the whole job.

Abundance and the return of craft

The familiar version of this argument says photography threatened painting, and painting answered with Impressionism, Cubism, and Abstract Expressionism. The instinct is right; the chain is far too neat.

What actually happened is sharper. Daguerre's process was published in 1839, and the trade it destroyed was specific: the portrait miniature, a whole profession of painters who made small likenesses for people who wanted to remember a face. A camera did that job better, faster, cheaper, and within a generation the profession was gone.

Painting did not die, because painting had been doing two jobs and lost only one. Once it no longer had to be the record, it was free to become the interpretation. The first Impressionist exhibition opened in Paris in 1874 — thirty-five years after the daguerreotype. That is what adaptation looks like: slow, and driven by a job being taken away.

The lesson is not the comfortable "art always survives." It is that a technology takes the mechanical half of a craft and leaves the judgment half — and the people who believed their value was the mechanical half have a very bad decade.

Quite a few professions are in their portrait-miniature phase right now. Feeds fill with generated text that is fluent and says nothing, because fluency was the mechanical half and it is now free. What becomes scarce is the thing generation cannot fake: having actually checked. Having a position. Publishing a number with its uncertainty attached instead of a confident round figure nobody can trace.

The part of the pattern nobody enjoys

Every one of these transitions was expensive for the people living inside it.

The century after Gutenberg gave Europe the Reformation, a sequence of religious wars, the collapse of the scribal economy, and eventually a censorship apparatus. The century after the pin factory gave us industrial cities and the labour movement they made necessary. The liberating parts and the destabilising parts were never separate events. They were the same event.

The governance half is already running this time, on a far shorter clock. The EU AI Act went from proposal to phased obligations in a few years and has already been amended. UNESCO has catalogued nine distinct regulatory approaches lawmakers are using worldwide. What took the printing press a century — from tool to systematic attempt at governance — is taking this perhaps five years.

Faster is not automatically better — the 1559 Index answered a crisis that had already happened, and much of today's rulemaking has the same backward-looking shape. But the compression is real, and it means the window that stayed open for a generation after Gutenberg is likely to be shorter this time. The open-weight ecosystem exists right now. Treating it as permanent is the one clearly available mistake.

What to actually do about it

The tempting ending is to ask what you will create. Good question, slightly cheap one — it does not tell you where to stand. The more useful version, drawn from the pattern rather than the mood:

• Own something; do not only rent it. Rented capability is real capability right up until the terms change. Download weights. Run one thing locally, even something small, so you know what your own hardware can actually do — Ollama takes about ten minutes.
• Check the arithmetic before you believe the story. Including this one. "It runs on consumer hardware" is testable in a minute, and it is frequently false.
• Read the licence before the benchmark. Venice started granting privileges in 1469. The restrictions on this generation of tools are already written down, in the model cards, and hardly anyone reads them.
• Get broad on purpose. The bargain that rewarded narrowness is 250 years old and it is ending. Two adjacent competences beat one deep one once the execution layer is free.
• Do not confuse volume with work. Generation is free now, which is precisely why it is worth nothing. What you checked, what you decided, and what you were willing to be wrong about in public is what is left.

The tools have changed and the canvas is genuinely enormous. But the fifteenth century did not reward the people who were excited about printing. It rewarded the people who had something worth printing — and who noticed that the window would not stay open forever.

It is open now.

Keep going

• GPU & VRAM checker — what your card can actually run
• The model library — licences, sizes and commercial-use verdicts
• Set up Ollama and run your first local model
• Local vs cloud: the cost calculator
• How every number on this site is calculated
• Why running a local LLM is cheaper and more secure than you think