Markdown - All the Way Down
---
title: Markdown - All the Way Down
layout: post
format: markdown
secretly_also: yaml
turtles: all the way
tags:
- formats
- hubris
- tracksuits
warning: this file is wearing another file as a hat
---
Markdown - All the Way Down
There is an old story, possibly apocryphal and therefore perfect, in which a scientist explains the cosmos to an elderly woman, who informs him that he is wrong and that the world actually rests on the back of a giant turtle. When asked what the turtle stands on, she replies that it is turtles all the way down.
The modern version of this woman runs a large language model. Ask it what the world is made of and it will tell you, with great confidence and a nicely nested bullet list, that it is Markdown all the way down.
She is, annoyingly, more right than the scientist.
Every revolution ships with a file format
Civilizations are remembered by what they wrote on. Clay tablets, papyrus, vellum, the cheap acidic paper of the twentieth century that is currently dissolving in every library on Earth. The medium is not quite the message, but it is the message's coffin, and you can learn a great deal about a culture from the box it chose to be buried in.
Computing is no different, except faster and more smug about it. Every technological revolution arrives holding a file format the way a toddler arrives holding a stick - convinced it is a sword, a wand, and a universal solution. The format is never neutral. It encodes the era's ambitions, its anxieties, and above all its idea of what counts as tidy.
Consider the fossil record.
A brief and unflattering paleontology
The punch card. The Cambrian explosion of computing was a stiff rectangle of cardboard with holes in it. The punch card believed, with the touching sincerity of the very young, that all of human knowledge could be expressed in eighty columns and that anything longer than that was probably frivolous. Entire payrolls, census records, and one or two genuine atrocities were administered through these things. The lesson, learned and immediately forgotten, was that whoever controls the format controls what is sayable.
Plain text and ASCII. Then came the great democratic flatland of the .txt file - no fonts, no styling, no pretension, just characters marching in a line like prisoners. ASCII was the Esperanto that actually worked, and it remains the bedrock under everything that followed, including, you will note, Markdown. Plain text never died. It simply learned to wear costumes.
HTML. The web arrived dressed in angle brackets, promising that anyone could publish anything. This was true, and it was a catastrophe, and a wonderful one at that. HTML was the format of the optimists - hyperlinked, sprawling, structurally unsound, held together by <table> tags doing jobs they were never designed for, much like the early web's business models.
PDF. We cannot skip the cockroach of file formats - lurking, immortal and occasionally spreading desease. The PDF is what print civilization left behind so that it could haunt us. It cannot be edited, it cannot be reflowed, it cannot be reasoned with. You do not open a PDF so much as exhume it. Every government on Earth runs on PDFs, which tells you exactly how much governments enjoy being changed.
Each of these formats won an era and then quietly lost the next one, demoted from revolution to infrastructure to legacy to that thing the intern is afraid to delete.
XML. Ah, XML. The format of committees and serious corporations. XML was what happened when serious men in serious meetings decided that verbosity was the same thing as rigor. It wrapped every scrap of data in opening and closing tags so that a single phone number could occupy the visual footprint of a short novel. XML believed that if a thing was worth saying, it was worth saying three times with a schema. An entire generation of enterprise software was built on the conviction that more nesting meant more truth. We do not talk about SOAP anymore. We have agreed, as a species, to let it lie.
JSON. JSON was the format invented by people who had survived XML and come back changed. It was lighter, meaner, stripped of ceremony, and it ran the entire API economy on little more than curly braces and trust. JSON is the format of the second-wave web - pragmatic, transactional, allergic to comments because comments are feelings and feelings slow down the pipeline.
YAML. YAML was the humane reaction to JSON, which is a sentence that has aged poorly. It promised configuration files a human could actually read, and it delivered this by making whitespace load-bearing, so that a single misplaced space could quietly bring down a production cluster at three in the morning. YAML is the native tongue of the DevOps era, the language in which we lovingly describe the infrastructure that will later betray us. It is so eager to be helpful that it will read the word no and decide, on your behalf, that you meant the boolean false, which is why the country of Norway has been silently deleted from countless config files. YAML wanted to be friendly. It became the format you fear at deploy time.
Enter the humble pound sign
Markdown was born in 2004, a modest creature, the work of a writer who simply wanted to type on the internet without drowning in HTML tags. The premise was almost insultingly reasonable - text should be readable as text, even before anything renders it. A heading should look like a heading. Emphasis should be marked with a small honest asterisk rather than a <strong> and its closing twin. The raw file and the finished file should be cousins, not strangers.
It was designed for humans writing for humans. README files. Blog drafts. Notes to oneself. It asked for almost nothing and gave back something legible. For about fifteen years this was its entire, unremarkable career - the format of programmers and bloggers, beloved in a niche, ignored by the world.
And then the machines learned to read.
How a note-taking format conquered artificial intelligence
Here is the part the elderly woman with the turtle understood instinctively. Large language models did not adopt Markdown because a committee chose it. There was no summit, no standards body, no press release. Markdown won the way water wins - by being the path of least resistance and then becoming the shape of the riverbed.
The reasons, stripped of mysticism, are these.
- It was everywhere in the training data. The models learned to write by reading the internet, and the internet's thinking-out-loud - its READMEs, its forum posts, its documentation, its half-finished tutorials - was overwhelmingly written in Markdown. The machine simply learned the dialect its parents spoke.
- It degrades with dignity. A broken Markdown file is still a readable file. Mangle the syntax and you get text with some stray asterisks, not a stack trace. Mangle XML and you get a parser screaming into the void. Forgiveness is an underrated virtue in a probabilistic writer.
- It is structured enough to mean something and loose enough to mean it without trying. A model can produce a heading, a list, a code block, and a table without ever leaving the comfort of plain text. It is the format equivalent of a tracksuit - presentable enough to go outside (or to a wedding if you are a Russian), but you could also sleep in it (unless you are a Russian - hey, it's your track suit for special occasions!).
So the chatbots speak Markdown the way the rest of us breathe. The bold, the bullets, the tidy little headers - this is not a feature someone added. It is the native accent of the thing.
The sardonic part, as promised
Notice what this says about the era, because the format always says something.
The age of XML believed that truth was a matter of structure, and so it built cathedrals of nesting. The age of JSON believed truth was a matter of efficiency, and so it stripped everything to the bone. And the age of artificial intelligence has chosen, as its native tongue, the one format that was explicitly designed to look like it is barely a format at all.
This is extremely on brand. Our most powerful software pretends to be a casual conversation. Our most elaborate statistical machinery pretends to be just chatting with you. And the format it reaches for, instinctively, is the one whose entire design philosophy is to disguise its own structure as plain honest text. Markdown is the format that performs effortlessness, which makes it the perfect costume for an intelligence that would very much like you to forget how much machinery is humming underneath.
There is a second, sweeter irony. Markdown was invented so that humans could write easily and read the raw file comfortably. Today the raw file is mostly written by machines and rendered for humans who will never see the asterisks at all. The original use case has been turned exactly inside out, like a sock, and almost nobody noticed.
All the way down
So here we are. Everything that makes todays agentic engineering is written in Markdown. This very document is Markdown, written about Markdown, very possibly produced by a model that thinks in Markdown 😉, to be rendered into something readable by software that assumes Markdown, for a reader who will judge it by how clean its headers look. It even opens with a little block of YAML up top, because a Markdown file these days is rarely just a Markdown file - it is a Markdown file wearing another format as a hat, and that hat has opinions about Norway.
The pound signs do not stand on anything. The bullet points do not stand on anything. It is, as the woman said, turtles the whole way down, except the turtles are little asterisks and the cosmos is a tracksuit.
Every revolution gets the file format it deserves. Ours got the one that pretends it is not there.
And honestly, given the alternatives, we got off easy. Imagine if it had been XML. I survived it and I am grinning through tears.