In a world where Carthage survived the Punic Wars and shared the Mediterranean with Rome for four centuries, the events are specific enough — and weighted enough — that they produce the kind of compressed references every real civilisation generates.
Sending silver to Cartagena is an act of apparent absurdity that conceals urgent purpose. Hill farms names a prize so trivial it isn’t worth the cost of taking it. Unpaintable is the quiet term of respect for the work that actually mattered, done by the person nobody will thank — the decisive contribution that doesn’t make the temple wall.
“My cousin sailed three days to deliver a jar of olives to a farm.” “Who lives on the farm?” “Ah. Silver to Cartagena.”
“I said ‘we should talk.’ She packed a bag.” “Saltus Pass.” “I meant about the goat.”
A midwife in Gades, overhearing the father being congratulated: “Carthalo’s painting.”
These phrases come from a 125,000-word alternate history called Mare Non Nostrum. The book was produced through sustained dialogue with a language model — months of argument about institutions, incentives, geography, and what happens when two incompatible civilisations are forced into coexistence, neither able to destroy the other. The model generated the prose. I did not write the manuscript’s sentences. But the world that produced these idioms — the events behind them, the logic that connects them, the civilisation dense enough to mint its own metaphors — that is what the argument built.
I want to describe how this works, because the current conversation about AI and writing has no category for it.
Everyone knows about slop. You type a prompt, the machine produces text, you publish it. The text is generically confident and structurally hollow. People are right to dislike it.
Everyone knows about AI-as-editor. You write your ideas, the machine restructures and polishes. Real and useful.
The discourse moves between these two poles, and because it moves between them, anyone who says “I used AI to produce a book” is heard as describing one or the other. There is no third category. I am describing the third category.
Within an hour of the first conversation — through a series of proposals, rejections, and counter-proposals — the dialogue had found its question. Not “what if Carthage won the Punic Wars?” — a premise. But “what does a Mediterranean look like over centuries when two great powers coexist and neither overcomes the other?” — a project. From that structural question, working backwards, we found the divergence point: the First Punic War, and specifically, a minor Carthaginian naval commander whose eleven ships changed the war’s direction. The divergence point wasn’t chosen and then explored. It was discovered, through argument, as the most generative way into a set of questions I wanted to think about.
That pattern — discovery through argument rather than execution of a plan — defined everything that followed.
I arrive at each chapter with a vision for what it needs to do — its role in the arc, its key events, its logic. I discuss that vision with the model. Not “write me a chapter about X.” A genuine exchange about structure and approach, often five or ten rounds before any prose is generated. The generation, when it comes, is shaped by everything the argument has already established. This is not a process where the model produces pages and I select from them. It is a process where thinking happens in the dialogue and prose emerges constrained by what the dialogue has built.
Then I read what the model produces. And the diagnostic work starts. Again and again, the generated text would present a historical development as though it had a single clean cause — and I would recognise that the material so far didn’t support that, that the real causal structure was richer and more conditional. The diagnosis would lead not to a rewrite but to a discussion about what was actually going on underneath, and that discussion would produce the missing layers of causality, the framework that made the event legible. Only then could the prose be rewritten against what the argument had built.
But this makes the process sound one-directional — as though I held a complete vision and argued it into the model’s prose. That happened sometimes. Often, something more interesting happened. I would identify a problem, and the model would propose solutions. I would reject all of them. But the rejections were generative: each wrong answer illuminated the space of possibilities, and the illumination sparked a direction I hadn’t considered. Many of my best ideas came not from independent vision but from reacting to suggestions I didn’t like — each wrong answer sharpening my sense of what was right until the right thing crystallised.
The model’s errors, in fact, turned out to be signal rather than noise. When the world-building was solid, the prose was usually right, because the constraints were tight enough that plausibility and correctness converged. But when there was a gap — an unspecified rule, an undefined incentive — the model would fill it with a plausible guess, and the guess would be wrong, and the wrongness would tell me exactly where the gap was. The structural frameworks that now constrain every sentence in the book — a ten-layer analysis of why Rome stays off the sea for centuries, a mirror framework for why Carthage stays off the land, a source tradition explaining which historical records survive in this alternate world and why — all emerged from this diagnostic process. None were planned. Every one was exposed by the argument hitting a seam that hadn’t been sealed.
Hundreds of conversations. Almost a million words of discussion and reference material. The book is what the argument left behind.
Some art forms have long-established roles for the person who holds the creative vision and works through others to realise it. Film has directors. Music has conductors. Theatre has had someone holding the shape of the whole since the Greek choragos. No one thinks the director’s authorship is diminished because she didn’t operate every camera.
Prose writing has never quite had such a role. Co-authorship between humans can separate vision from execution to a degree, but the norm — and the expectation — has always been that a single person does both. Language models allow the separation to become routine and, if the writer wants it, complete.
The analogy is not perfect. A director works with actors who interpret, who bring independent judgment to their performances. A language model is not that. The relationship is different, and the differences matter. But the structural point holds: you can be the author of a work without having executed every detail of it, and that is true whether you are working through humans, through a model, or through some combination nobody has named yet. If anything, the absence of independent judgment in the model means more responsibility falls on the vision-holder, not less — there is nothing else in the process that is trying to make the work cohere.
The book is mine in the way a film is the director’s. Not because I generated the sentences, but because the architecture that gives those sentences meaning — the choices about what matters, what connects, what the work is for — emerged from my sustained engagement with the process.
What does the director do, specifically?
Hold the vision. Not the text — the vision. I hold the whole work in my head: what has been written, what hasn’t, what the finished work should feel like, how every part connects, the places where one chapter’s implication constrains another’s. That is the understanding of what the work is for, and it lives in the human.
Diagnose. When something is wrong, identifying at what level it is wrong is the core intellectual work. A tone problem? An implied logic problem? A gap in the world that must be specified before anything else can proceed?
Steer. Not by writing prose but by arguing — precisely, specifically — about what the work needs. “Make this better” produces nothing. “This passage implies centralised authority when the point is distributed agency, and the distinction matters because it drives the next three chapters” produces something you can build with.
And refuse. The model produces something adequate. Good enough. And something in you says: not yet. The refusal — another round when the output could pass — is where quality lives. It requires knowing what right looks like before you’ve seen it.
None of this means the human isn’t working in language. The thinking happened in words — hundreds of thousands of them, spoken and typed, arguing about structure, meaning, and implication across hundreds of separate conversations. That body of dialogue, taken together, is a work ten times the length of the manuscript. What’s different is that the thinking doesn’t happen in the final artefact. It happens in the argument that produces the final artefact.
The book’s attribution reads: Ex Dialogo cum Machina. From dialogue with a machine.
Not “by” the machine. Not “despite” the machine. From the dialogue — the sustained, iterative, sometimes exhausting argument between a human who holds the vision and a model whose output is the surface the thinking happens against. The craft is real. It is learnable. It is demanding in ways the slop discourse cannot imagine. And it produced a world dense enough to mint its own idioms — phrases people reach for to describe situations that have no single English equivalent, coined from four centuries of events that never happened, in a Mediterranean that was never Rome’s.
Whether that’s worth reading is not something an essay can prove. The book is there. Come and see.