The Story Behind Florilegium
This line sat at the bottom of my terminal on July 28, 2026, while I worked. It took about four seconds to read. It was written about four hundred and fifty years ago. This is the story of why it was there.
The worry
Starting in April, what concerned me most was the potential catastrophic risk from AI. I was persuaded that the people taking safety seriously were trying to do it the right way: assembling thoughtful, ethical people to set guardrails, write down rules, and slow things where slowing was needed. I still believe that work matters.
At the same time, one of my best friends from middle school was deep in his own experiment. He was building a new language for formal verification, a way of writing proofs that a computer can check, and he was working on it together with Opus 4.6, a frontier AI model. He came out of that spring convinced that superintelligence would arrive through self-referential systems: systems that can read, evaluate, and improve themselves. For the ethics of the work, he set the highest standard he knew, drawn from the people he most admired. I admired those values too.
Probably he is right about the self-reference.
But the more I sat with it, the more one question bothered me, and it was not about capability at all: should the morality of AI be determined by a few? Even a few with good values? A constitution for a machine, however wise, is still written by a handful of people in one place, in one decade. Everyone else inherits it. I had no answer. The question just stayed, the way real questions do.
The loop we are already in
Around that time I was reading Gödel, Escher, Bach, and I started noticing strange loops everywhere: in music, in proofs, in organizations, in myself. A strange loop is what happens when a system's output feeds back into its own foundations. I came to believe we actually need these loops. A person improves by reading their own life; a field improves by studying its own methods. While self-reference may be how superintelligence arrives, we still need a bridge to connect values between humanity and intelligence.
Here is where the threads of this story connect, and it took me several months to see it. Humans and AI are now inside one loop together. The models learn from what we write, and we increasingly write with the models. If the path to AGI runs through this co-evolution, then whatever shapes human attention and taste is also, quietly, shaping AI.
And look at what is shaping human attention. Short-video feeds, micro-dramas, games tuned by recommendation systems until ordinary people cannot put them down. These systems do not need you to grow. They need you to stay. Every hour inside them flattens a person a little, and the flattened person then writes the flattened text that trains the next model, which generates more of the same, faster. The loop runs downhill.
So, the safety question and the culture question turned out to be the same question wearing two coats. You can set guardrails at the top, on the model. I am glad people are doing that. But you can also work at the other end of the loop: on the humans inside it, one ordinary person at a time.
Building the other end
The occasion, when it came, was small. In a builders' group I am part of, someone posted a challenge: an app for learning a new word every day, useful for everyone, including native speakers. A modest idea. It happened to land while I was carrying the big question, and the two connected, because I had been feeling that bottleneck myself. Not a bottleneck of information. A bottleneck of taste: knowing which word, which sentence, which idea is better, and why.
I knew immediately what I did not want to build: another app. Another icon asking to be opened, another streak asking to be kept, another feed. Learning already carries enough friction. The tool should carry none, and it should be the opposite of an attention trap in its very shape.
As a starter for builders, I put it where one's eyes already fall a hundred times a day: the status line, the single line of text at the bottom of the terminal. One line. One flower a day. Reading it takes a few seconds. There is no feed underneath, no streak to keep, nothing to sign up for. If it says nothing to you today, it costs you nothing, and tomorrow brings a different flower.
The gathering, and the gate
The name came from an old word. A florilegium is literally "a gathering of flowers": the medieval name for a book of passages a reader copied out by hand, what one person, out of everything, chose to keep. For centuries this was how ordinary readers built themselves, one copied passage at a time. People kept these books under another name too: commonplace books. The practice is much older than the syllabus.
For the first shelf, I asked the AI models I was building with what the best things humans ever made might be — a choice that itself aligns with the value of florilegium, not being trapped in a collapsing iteration — and started there: Seneca's letters, Marcus Aurelius, Montaigne, Emerson, Zhuangzi, Ecclesiastes, Sappho's fragments, The Library of Babel by Jorge Luis Borges, and of course, books you import yourself.
Somewhere in these months my own framing shifted. I began by telling everyone "ultimately, it comes down to taste, which is developed by the best things humans ever made." I no longer quite believe that, or rather, I believe it is half. Taste is how you choose. Values are why. Both are gathered along the way.
Whose values, then?
This is where the April question comes back, because I found an answer in the thing itself, afterward, the way you find your reasons in most of what you make.
If the alternative to values determined by a few is values gathered by many, then the gathering has to be legible. You should be able to open the book, read the shelf, see the sources, and audit the choosing. A florilegium is exactly that: a value system you can open. Nothing in it is decreed. Every flower in it was picked by someone, from somewhere, and it says where.
I recently read a line quoted from Leo XIV: "A more moral AI is not enough if that morality is determined by a few." And I have been influenced by the writers describing the intelligence curse: the fear that once machines can do all the work, powerful actors will stop needing regular people, and stop tending them. I cannot solve that. But I can build in the opposite direction. I want to contribute to upskilling people, and, where possible, to awakening them. A tool that assumes ordinary people are worth tending, that their taste can deepen, that their values are worth gathering. And if humans and machines really do co-evolve, then tending the human end of the loop is not charity. It is safety work of a different kind.
At the core of it, I believe in something simple. On an ordinary day, in an ordinary moment, a person can meet a sentence, a song, an artwork that deeply resonates with the day they have just been through. And that becomes the motivation to proceed on their own odyssey.
What I don't know yet
Whether this should speak more languages. The Tang poets are on my shelf in translation, while half of me can read the originals. Whether resonance should be shared: I want to know when a sentence lands in someone's life, and I refuse to build a leaderboard for it, and I have not fully resolved that tension. Whether the next surface is the ear instead of the eye: passages spoken aloud, for the walking parts of an odyssey. If you have thoughts on any of these, I would genuinely like to hear them.
Conclusion: a readymade
A question I keep for the end of dinners. In 1917, Marcel Duchamp signed a factory-made urinal and entered it in an art exhibition as "Fountain." The object was nothing; the gesture was everything: an unauthored thing placed where authority lived, forcing the question of who decides what counts. So what is that gesture in AI, in 2026? A friend and I half-joke about carrying a flint, fire stolen the old way, into a top conference as a safety poster. To see how far we have come from there. And how should we now travel from here?
But some days I think the readymade is already here. We just need more attention, and more humanity.
Florilegium is free and open: the walk · the code. The flowers are public domain. The gathering is yours.