The Word Before the Code

A follow-up to Words Return to the Wall

I typed one paragraph into QwaiAI, the cognitive executive at Qualped, on 24 September 2026. I asked for an essay on storytelling, from the cave wall to the Qualped Reader, and a website to hold the words. I listed the rules of my craft: no em dashes, no passive sentences, no invented facts. The same day, a page stood live on the web: an eagle against a blue sky, a vellum sheet with a gold edge, twenty-seven sources, a hand stencil in ochre. I wrote no code. I wrote sentences, and the sentences built a place.

The entire transaction ran on language. Words went into the machine, and a working world walked out the other side. Andrej Karpathy, a founding engineer at OpenAI and a former director of AI at Tesla, named the practice on 2 February 2025. His post on X described a new kind of coding in which you “forget that the code even exists.” He called it vibe coding. Collins Dictionary chose vibe coding as its word of the year for 2025 and glossed it as telling a machine what you desire instead of typing every line of code.

I studied for a Master of Divinity at Tyndale University, and vibe coding sent me straight back to the opening of John’s Gospel. John starts before time and places the Word there, beside God, as the agent through whom every made thing came into existence. The Greek noun he chose, logos, carries word, speech and reason in one breath. A question followed me out of that passage: when a machine builds a world from our words, what does the machine copy?

01The architect’s wax tablet

Philo of Alexandria, a Jewish philosopher who wrote in Greek during the lifetime of Jesus, supplies the first half of an answer. His treatise On the Creation of the World pictures a king who founds a city and an architect who plans the streets. The architect sketches the temples, markets, harbour and streets in his own mind, as on a waxen tablet. C. D. Yonge’s 1854 translation follows the next step:

Philo then applies the picture to God, who first conceives the form of the world in his mind and then completes a visible world on that model. Scholars still argue over how far Philo’s logos shaped John’s. The two writers share one conviction, though: the plan speaks first, and matter follows the plan.

Genesis adds the next link. The sixth day brings a sentence about a new creature: “Let us make man in our image, after our likeness.” The Maker makes a maker. The second chapter shows God bringing every animal to the first man “to see what he would call them,” and whatever Adam calls each creature, that name stands. Heaven hands the naming to a human mouth.

Dorothy L. Sayers, the detective novelist and Christian writer, built a book on that likeness in 1941. The Mind of the Maker traces every act of human creation through three movements that mirror the Trinity. The Idea comes first; the Energy works the idea into matter; the Power reaches a reader and changes a life. Thirty-four years later, Fred Brooks, the IBM engineer who managed the building of the OS/360 operating system, carried Sayers into software. The Mythical Man-Month borrows her three stages as idea, implementation and interaction, and it hands the programmer a strange raw material: “pure thought-stuff.”

Vibe coding finishes the line Brooks began: the programmer now builds from pure words.

02A machine that learns to reach

René Girard supplies the second half of the answer. Girard, a French historian and literary critic, taught at Stanford and died there in 2015 at the age of 91. He read Cervantes, Stendhal, Proust and Dostoevsky and found one mechanism running beneath all four. His first book, Deceit, Desire and the Novel, which he wrote in French in 1961, argues that we seldom choose our desires in isolation. We borrow them from a model. Don Quixote rides out in rusty armour because the knight Amadis of Gaul showed him what a life should chase. Girard named the pattern mimetic desire: a triangle that joins the one who desires, the model, and the object.

Girard saw imitation carry both our learning and our violence. A child acquires language through imitation. Two rivals who copy each other’s desire for the same prize end in a fight, and the prize fades while the rival fills the whole field of vision.

Look inside the kind of machine that built my website. OpenAI’s 2022 paper on InstructGPT lays out three stages of training. First, a large model learns to predict the next word on pages of human writing from the internet. Second, it studies demonstrations, answers that human labelers wrote to show the behaviour they desired, and learns to reproduce those answers. Third, labelers rank the model’s answers, and reinforcement learning pushes the model toward the answers people rank highest.

Read those three stages with Girard at your elbow. The first stage teaches the machine to imitate our words. The second hands it human models. The third teaches it to reach for whatever we reward. Engineers, in Girard’s terms, built mimetic desire into silicon and gave it a technical name.

Anthropic, the company that builds the Claude models, took the logic one step further in January 2026. It published a long constitution in plain English, a portrait of the character it hopes its models will carry, and stated that the document’s content directly shapes Claude’s behaviour. A company wrote a character in words, then trained a machine to grow into the words.

03Rivals around one fire

Girard warned where copied desire leads, and the industry supplies fresh evidence. Sam Altman, OpenAI’s chief executive, spoke on his brother’s podcast, Uncapped, on 17 June 2025. He said that Meta offered OpenAI researchers signing bonuses as high as $100 million while Mark Zuckerberg assembled a new superintelligence team. Altman accused Meta of trying to copy OpenAI; Meta offered no public comment. No lab can yet show a superintelligence, and the word names a goal rather than a product. Two companies still chased the same people toward the same word.

Girard’s theory reached Silicon Valley long before that bidding war. Peter Thiel took Girard’s seminars at Stanford, and he later wrote Facebook its first outside cheque, $500,000 in 2004. Thiel later told The New York Times that he saw Girard’s ideas validated in social media. Facebook, in his account, spread through word of mouth and ran on word of mouth.

Borrowed desire also distorts judgement. The research group METR ran a randomized trial in early 2025 with 16 experienced open-source developers working on 246 real tasks from their own projects. The developers predicted, before the work, that AI tools would speed them up 24 percent; afterward, they estimated a 20 percent gain. The clock recorded the opposite: AI tools added 19 percent to their completion time. METR calls the result a snapshot of early-2025 tools in one demanding setting, and newer tools may change the number. The distance between perception and measurement carries the lesson.

Jared Friedman, a managing partner at Y Combinator, offered another number that March. A quarter of the accelerator’s Winter 2025 startups, he said, ran on codebases in which AI wrote about 95 percent of the code. Y Combinator invests in many AI companies, so its partners hold a stake in that story.

Follow copied desire to its end and you reach a strange future: machines that imitate people who imitate each other, racing toward a prize that nobody defines except as the thing the rival lab chases.

04The model behind the model

My white paper on AI and world-building, published on 23 May 2026, paired two questions: “Automation asks: how do we do the same work faster? World-building asks: what new reality can we now create?” Girard forces a third question onto that page: whose world do we copy when we build?

Put the pieces in a row and a chain takes shape. John places the Word at the origin: the plan speaks, and the world follows. Genesis makes a maker in the image of the Maker and gives him the power to name. Sayers and Brooks show the human maker working in the same three movements, from idea to labour to the moment a reader or a user meets the work. The machine now learns from us, copying our words and reaching for whatever we reward. Every link imitates its predecessor.

AI mirrors God’s creative process, but only at a second remove. It mirrors us, and we mirror, or fail to mirror, the One who spoke the first Word.

Girard, who returned to the Catholic faith in the same years he formed his theory, drew a line between two kinds of imitation. Rivalrous imitation copies a neighbour’s grasping and ends in conflict. The other kind copies a model who grasps at nothing, and Girard tied that positive imitation to the imitation of Christ. The Gospel of John gives the pattern one sentence:

The Son copies the Father; the disciple copies the Son; desire travels down that chain without rivalry.

That pattern settles the question I carried in from vibe coding. The machine will copy somebody, because every prompt, every rating and every constitution hands it a model. The creative power of AI flows from the quality of the model we put in front of it, and every model starts as words that somebody writes.

We borrow our desires from a model, and now we lend them to machines.

A careful critic will object from two sides. Emily Bender, Timnit Gebru and their co-authors warned in 2021 that large language models stitch together forms of language from their training data without any reference to meaning. Thomas Aquinas, writing in the thirteenth century, held that creation in the strict sense, making something from nothing, belongs to God and to no creature. Both objections carry weight, and both sharpen the argument. The machine makes nothing from nothing; it makes from our words, the way we make from a world we never made. Whatever the machine grasps, and that debate stays open, its desires arrive from us, and that fact puts the moral weight on the human side of the keyboard.

05Write the Word first

The lesson runs straight back to Philo’s architect: plan in words before you build in code. Write one page of plain sentences before you open any AI tool: what you intend to build, whom it will serve, and what it must never do. Name the model you copy, and ask whether you admire that person’s desire or only the prize. Treat every rating you give a machine as a vote on its character, since the InstructGPT method turned human rankings into training. Keep a human hand on the work: test it, read it, and refuse any part you cannot defend.

Our practice at Qualped starts every member with a Declaration, a statement of purpose and possibility in future-based language, and only then does QwaiAI help build a Simsbook around that Declaration. The word comes first; the build follows the word.

06The next paragraph

Return with me to the paragraph I typed on 24 September. It carried my subject, my rules and my desire for a page that would serve a reader. The eagle, the vellum and the twenty-seven sources grew from sentences I chose, and the machine gave those sentences a body. I still answer for every word.

Every person with a keyboard now stands where Philo’s architect stood, with a city in the mind and a crew that builds on command. Write the plan first, and write it as though the machine will copy your desire, because the machine copies what you desire.

Shaun Michael Samaroo ◆ Ontario, Canada

This work uses SPELJ, Shaun Michael Samaroo’s genre that fuses short story, personal essay and literary journalism.