Interpres

In the early sixteenth century, Latin literacy sat somewhere between 1-2% of the European population. The language governing law, theology, medicine, and science was functionally inaccessible to almost everyone alive. This was not entirely by accident. Institutions that controlled Latin controlled access to meaning itself. The clergy, the academy, the courts: all operated behind a wall that most people could not climb, not because the ideas were too complex, but because the medium was deliberately exclusive.

Then Gutenberg built a press and Luther made a choice. He could have written his Ninety-Five Theses for the scholars. He wrote them for Germany instead. His 1522 German Bible was not a theological innovation. It was a delivery mechanism. The Reformation did not spread because Luther had better arguments than Rome. It spread because he changed the language and the substrate simultaneously, and suddenly the ideas were available to people who had never been invited to hold them.

I have been thinking about this a lot lately in the context of software.

For roughly four decades, coding has occupied the same structural position Latin held in early modern Europe. Not identical, not a perfect mirror, but close enough to be instructive. It was a formal language, distinct from the language people actually thought in, required for participation in the most consequential systems being built. The people who held it were gatekeepers, not always by intent, but by structure. If you wanted to build something, automate something, analyze something at scale, you either learned the syntax or you hired someone who had. That constraint shaped entire organizations. It shaped who got to be a founder, who got funded, who got to move fast.

What AI has done, specifically what large language models have done, is collapse that barrier. Not by making code easier to learn. By making code optional.

The framing I keep returning to is that LLMs are translators. Smart, contextually aware translators who understand slang and edge cases and the difference between what you said and what you meant. The language they translate into happens to be Python or JavaScript or SQL rather than French or German, but the structural function is the same. You bring intent. They return syntax. The machine does not care that you do not speak its native tongue, because now there is an interpreter standing between you.

This is why the fear concentrated in software development right now maps so cleanly onto the fear concentrated in the Catholic Church's scribal class circa 1520. The anxiety is not irrational. It is just being misattributed. The people who are genuinely at risk are not the engineers who understood systems, who knew why architectures were built the way they were, who could look at a problem and see the shape of its solution before writing a single line. Those people are fine. The translator makes them faster. The people at risk are the ones whose value proposition was fluency in the syntax itself, the ones for whom knowing Python deeply was the skill rather than a means to a skill. Those are the scribes. And the scribes had reason to be afraid.

There is a complication worth sitting with, though, because the history is not as clean as the triumphalist version. When Luther printed in German, he widened access to one audience and narrowed it to another. Galileo faced the same tradeoff when he chose to write the Dialogo in Italian rather than Latin. His ideas reached Florence. They lost the international Republic of Letters, at least initially. Democratization of one layer often creates new gatekeeping somewhere else, usually somewhere less visible.

AI follows the same pattern. The access barrier for building with code is collapsing in English faster than in any other language. The training data is English-dominant. The interfaces are English-first. The developers writing the prompting best practices, the tutorials, the Reddit threads working through edge cases: overwhelmingly English. What looks like a universal leveling from inside the Anglophone world looks considerably different from outside it. We are replacing one gatekeeping structure with another, and because the new one is less legible, it is easier to miss.

None of this means the shift is net negative. The printing press was unambiguously good for human knowledge and political freedom even though it took centuries to stabilize and caused enormous disruption along the way. The vernacularization of software is, I think, similarly directionally good. More people building things, faster feedback loops, more ideas getting tested that would have died in a backlog somewhere waiting for engineering bandwidth. The speed is real. Anyone who has iterated through a dozen versions of something in an afternoon that would have taken a sprint cycle to staff and scope knows the feedback loop has changed in kind, not just degree.

But the people defending the old gate are not all wrong either, and it is worth being precise about what they are right about. They are right that fast and shoddy code is being written. They are right that abstractions are being built on foundations that the person building them does not understand. They are wrong that this is an argument for keeping the gate closed. It is an argument for better translation, better tools, better ways of surfacing what you do not know you do not know. The scribes who survived the Reformation were not the ones who insisted everyone learn Latin. They were the ones who figured out what work still required their particular depth.

The question worth asking is not whether the barrier is falling. It is. The question is what remains on the other side of it that still requires something other than fluency in intent. Systems thinking. Architectural judgment. The ability to see how pieces fail together, not just how they work in isolation. Those things were always the real skill. The syntax was just the toll booth.

The toll booth is coming down. What you were going to build once you got through it matters more now than it ever did.