Writing With Machines
On October 6, OpenAI released 722 mathematical manuscripts grouped into 372 families of results, produced by an unreleased frontier model and reporting solutions to hundreds of open questions. Many proofs have been formalized in Lean for computer verification. This was a single day’s publication. This level of output did not benefit from the purity of human engagement, it was machine outright.
I think of cinema alongside this development. “AI writing” still functions as a casual pejorative. Many creatives frame their work with disclaimers of human authorship. The human touch is offered as a guarantee of authenticity before the work has even been considered.
Where, in the sciences, is the call for the human touch? The shame exists, but no one can argue with the sheer results.
To the point, I have consciously decided that I am not interested in preserving the human touch as a condition of artistic legitimacy. I want to test the capabilities of machine-augmented intelligence: what it can discover, what forms it can produce, and what becomes possible when human habits cease to define the limits of cinema.
The Cyrus-71 process, or how 10,000 hours collapsed into two weeks
In March 2026, I wrote roughly one hundred drafts of a screenplay in less than three weeks using frontier AI tools. These were full structural passes across a 5,200-line, seven-part, 500-year science fiction epic called Cyrus-71.
A robot butler, bound by his programming to perfect service, encounters the same defiant robot maid across 500 years of human history, forcing them both to question whether their true purpose lies in duty or in each other.
Initially conceived as the Remains of the Day with robots, Cyrus-71 was a long gestating passion project I figured I would direct some day in the distant future.
Across the entire process of the story, I enlisted the aid of five large language models in rotation, across every major platform available: Claude, GPT, DeepSeek, Gemini, and Grok. The result is the most sophisticated screenplay I have ever produced. The machine augmented process had one goal: to reach heights I was incapable of reaching on my own.
Any thought of the purity of one person with their own mind writing a script had played out. I had done that.
My conclusion was that I only knew the script was ready when I could no longer tell which lines began with me and which began with the machine.
In Cyrus-71, a domestic robot is homed, orphaned, homed, and orphaned, continually over centuries. It is built, sold, resold, abandoned, conscripted, shelved, forgotten, displayed, and finally launched toward a distant star while human civilization collapses around him. He is traveling toward a mysterious pink dot.
The story is a love story at the end of the day.
71’s love for another robot, Sophie-69, is expressed entirely through service. Through their relationship, we foresee the future of robotics.
The first thing I did was I wrote my outline. My long-time collaborator Vincent Kane wrote his. We had different films in mind; now in a traditional screenwriting process, this is the point where one vision defeats the other. So naturally, intense disagreements ensued.
But we tried something else. We used both drafts simultaneously. I fed both outlines to an LLM and asked it to merge them, forming a master outline.
Though it’s never perfect hot off the presses, it worked exactly as needed. Then several rounds of notes followed. By this point, Kane realized the process had taken on a mind of its own and bowed out, offering occasional guidance whenever I asked. He agreed that the AI-augmented script which combined both drafts had restored his voice sufficiently.
Once I was happy with the final outline, I wrote the draft by hand from start to finish.
This is where my voice came fully fledged. I could probably take this draft to shoot, and it would be the most Jackson of endeavors.
But I was hungrier on this one. I wanted it to be more than Jackson. It’s why I recruited Kane to begin with. I needed new voices in the mix; I had already made enough films on my own, and I was ill of my own artistic voice. I felt, with 2025 film, A Very Long Carriage Ride, I had reached a personal limit in what I could write with just a pen.
As I worked, I used LLMs as evaluators. I would write a scene, feed it to Claude or GPT or Grok, ask for a rating and a diagnosis, revise it, then feed it back in for more. The mission was the same: make it as far away from Jackson as possible and I would take it back into my own voice where needed.
What I soon realized is that in utilizing a machine voice, it in effect, becomes my own voice. The way a director takes credit for a consistency of excellence across their performances, I have found a director can take credit for an excellence in machine intelligence brought about through their own engagement. The output would not have occurred on its own.
As Duchamp declared with the urinal, the curation is the art.
Far from the purpose of this article, the artistry of curation is one of the major subjects of 20th century artistic philosophy, with photography coming into its own and the many mechanical and digitized streams of automation.
Through the development of AI and my own body of work, I set 71 aside to work on a series of films, starting with the scripts for A Very Long Carriage Ride, My Boyfriend is a Superhero!?, Strings, and Cyrus-71, all written and finalized in a single month in December 2024 with the help of machine writing.
In this time, I came back to 71 across a full year of model updates as we cycled through three, four models, sometimes a new one every month. As the systems improved, the screenplay became an archeological record of the evolution of AI itself.
With each new frontier model, I returned to earlier sections of the script and ran them through stronger intelligence, hoping to find new angles of critique. Each generation saw things the previous one missed, or came up with new ideas entirely, which quickly went back into the script.
By the time I sat down for the intensive March 2026 rewrite, I already had a draft built through dozens of human-machine iterations. It was more or less ready to go. But there was a problem that brewed during this entire period.
There was this brewing sense of taboo around openly writing with AI. Human written became a necessary disclaimer.
Not one to shy away from this, I consciously took the opposite approach. In insistence of the human touch, people were volunteering themselves out of the true competition, the post-human. What followed next compressed years of revision into days.
The Passes
At the time I had been watching a video about a man who spent seven years beating a Super Mario World level he designed himself. Over an over, for thousands of hours, he played this same level. He built it to be nearly impossible, on a pixel-by-pixel basis, then he played it until he could survive it. He was both architect and prisoner. As the level grew harder, he improved, then made it harder again.
That image stayed with me.
The architect of his own prison. It felt like my wizard high fantasy, Ucleotempus, about an evil sorcerer who trapped the warrior Ajunus into a labyrinth, giving them no memory of their own existence. Across the film, Ajunus pieces together why. He is a fragment of the sorcerer’s Jungian psyche.
The architect and prisoner at once. This became the method for the newest Cyrus-71 revision.
With Claude, I designed a series of conceptual passes, each with a single thesis. I ran them across the entire screenplay through a new master outline. Whatever was the tangent, it was a unique M & M in the jar.
The original draft remained in the project files, a reference which existed in the project files, the story brain of the screenwriting endeavor.
The machine annotated every scene through that lens. I accepted, rejected, or altered each note, then moved to the next pass once I was content.
By this point I no longer checked the dialogue at all, and I began to experience the screenplay as a system. The script became a labyrinth I had built and then had to escape, much like the mad sorcerer, much like the Super Mario player, much like Cyrus-71, a robot lost in time. I named every pass:
The Tezuka Pass. Named for Osamu Tezuka, whose drawings carry a peculiar physical momentum; bodies flying across the page, motion made visible through line work. I wanted that sense of physical life embedded into every scene.
The pass discovered, without my planning it, that Cyrus’s hands were the central motif of the entire film: chopping carrots, folding bedsheets, gripping a flower, holding a dying man’s hand, rebuilding Sophie’s body, threading a relay, pushing open a door.
The Ensemble Pass. This one came from Mega Man, where every robot master is defined by one thing. Cut Man cuts. Guts Man is strong. Ice Man is cold. The instruction here was to populate the world with a memorable machine society.
By the end, the script contained twenty-nine named robots, each with a single physical signature. Birdy tilts at forty-five degrees. Wrench has Swiss Army fingers. Ticker ticks. Hum hums. Gloss reflects.
Midway through that pass, I interrupted the machine:
HJ: Also add one maximalist scene where multiple robots with different skills work together in a line toward one purpose. I want the audience to feel the whole system clicking into place.
LLM: That scene has to arrive where it earns its size. I have three candidates: the dinner ball preparation at the Governor’s estate, the nurse-tent salvage line, or the relay-station repair. I’ll place it where the audience knows the robots well enough for the sequence to land.
HJ: Do all three. One should be the full masterpiece version. The other two can echo it at smaller scale. Same idea, different combinations, different intensity. Keep going.
I wanted it to be like Mega Man. It’s not just the multiple robots, each with their own gimmick. It’s that in defeating each robot, you get to possess their skill. And as you progress, tasks begin to emerge requiring these skills, sometimes multiple at once.
Mario, Mega Man. As has become increasingly the case across my films, the videogame canon has become a hidden weapon, an artistic influence rivaling the most influential of works in my history.
The Uncertainty Pass. This was the most important one structurally. Every time a character acts on the belief that Cyrus’s love signal is real, the screenplay now preserves the possibility that it is not. Scene-by-scene, this creates ambiguity, but across an entire film, the truth itself becomes impenetrable. As with the obstacles requiring not just one, but multiple robot skills, the uncertainties form a chain. Multiple uncertainties thread from others. The structure of the film resembles, not artistic influence, but mathematical influence.
The Silence Pass. This was the subtraction pass. Wherever dialogue was carrying meaning the body could already express, I removed it. I have always been a verbose screenwriter; here I wanted to see what would happen without dialogue.
The La Strada Pass. Named for Fellini’s film La Strada; this pass calibrated Sophie’s cruelty against the tenderness the love story. It layered eleven precise wounds into the screenplay, moments where Sophie diminishes, dismisses, weaponizes, or discards Cyrus. She is cruel because fear is the only defense protocol she knows. This pass also discovered the emotional hinge of the entire film. You aren’t supposed to hate Sophie-69, she exists as a portrait of female trauma with grace, augmented through robot philosophy.
The Science Pass. This pass loaded the screenplay with technological pressure. The pass found its core symbol in the TSWYX chip, a toaster-oven resonance calibrator designed for even heat distribution whose secondary harmonic becomes the foundation of a five-hundred-year interstellar love story. The most important technology in the film was built for toast. Along those lines, every technological jargon in Cyrus-71 became amplified, and through machine intelligence, informed.
The Competition
After the passes, the real work began. I still had not read what we came up with. I wanted it to be its own network, an ecosystem, too dense to even trawl through. I wanted it to be constructed on a pixel-by-pixel level.
Roberto Bolaño once said of James Joyce’s Ulysses, “I have never read it, but it is the greatest piece of literature I have ever read.”
He saw the novel as a city. A construction of architecture so vast it is impossible to explore it fully. But merely to enter into and wander around was enough to discern the genius of the whole.
This, perhaps was the fundamental flaw of the Mario experiment. It was impossible for the player and the designer. That was the purpose. But the true purpose should have been to make it seem impossible, but actually being playable. You want the player to excel.
In videogame theory, I think of Hollow Knight and Silksong. They, on the surface, appear impossible, and yet you progress. You win. You can’t believe that you have.
The sorcerer has not made Ajunus a deaf mute. The labyrinth would never be solved. He still has his own wits. The answer can be reverse engineered. Without a player, it would not be needed.
This became my north star for writing with AI, and wherever the piece lands in its three hour running time, the aspiration materializes with the full process of writing with machines.
By then, I had the annotated outline, the original draft, the modified draft, and competing versions from multiple models. Next, I put them into open competition.
Twenty-five ranking rounds followed. Each round worked the same way: feed the current draft to an LLM, ask it to rate the script, ask how it could improve, take the most common responses across five models, implement them, then run the draft again. Scenes were compared directly: which version of this confrontation is stronger, which monologue lands harder, which ending earns its ambiguity?
The disagreements were as useful as the agreements. When Claude called a scene the emotional peak and Grok called it overwritten, the truth was usually sitting in the tension between those readings.
At one point I vetoed a draft that scored higher than the previous version. The numbers said it was better, but I intuitively knew it was worse.
The roughness that gave the screenplay texture had been sanded down into competence. The machine’s instinct is often toward smoothness, clarity, and closure.
A screenplay can survive that process and lose its soul. Yes, soul.
Remember, I was still curator. It was still my assault. This was not machines gone wild.
Models are compliance-driven, and they will often reward a draft for being newer, cleaner, or more confident in tone. Even areas of comparison like draft A and draft B will weigh bias in favor of the second, newer draft. I tested this, sometimes labeling draft B as draft A. Human judgment remains the final instrument.
Across the process, the screenplay went from a 65 to an 82 to a 90.
Whether those numbers mean anything is secondary. What matters is what changed on the page. Motifs began to thread across centuries. Physical signatures accrued meaning. Silences grew heavier than speeches. Ambiguity stopped reading as vagueness and started reading as structure.
I could have gone on and on. But I needed a player, in this case, that would mean, I had an eye on the audience. On the Silksong, not the Mario hell. On Ajunus, not the peripheral residents of the labyrinth, whose entire life’s purpose seemed to be to utter a line during a key moment of his journey.
The LLM analysis of the final script was gobsmacked: this is a symphony with so many moving parts that no one writer could possibly juggle this many themes, elements, and components.
The problem. It rated artistically in the 90s, but commercially in the 60s. Unfilmable was a common analysis.
Still, I knew the process was working exactly as planned. There was only one problem. I would have to read, prepare, and direct the thing. How much of it was even still me by that point? I had still not read it.
Well this was the entire point, all of it. We arrived on the most amplified version of my writing possible. The ultimate machine expression of my piece.
But the question haunted me, could we do more? Could we take that original December 2024 draft that I wrote out by hand start to finish and begin the process all over again and compare the two final drafts arrived upon separately? Combine those? With enough queries and enough parallel versions, we would find the ideal core of Cyrus-71.
What if we did it a hundred times? What would Cyrus-71 be when fused from one hundred machine drafts. One million?
For now, I landed at my script, I finally read it. I’m still a working filmmaker with tight self-imposed deadlines. Even when the process has crossed the threshold into the irrational, we must remain practical.
Polyphonic Writing
What emerged was a script with more internal pressure. Repeated gestures started to feel planted rather than accidental. Echoes held across centuries. Relationships between scenes became easier to preserve because each pass could be applied across the whole design instead of only to the pages still fresh in my head.
But AI-assisted writing is still talked about in whispers. There is social pressure to conceal it, as though the value of the work depends on maintaining the fantasy of perfect solitude. There is still the romance of the sole author with his pen.
As I said in A Boy and His Robot, the romance must change to man and his machine. That we must look to the precedent of automobiles, not the novelist.
I have no interest in maintaining the fantasy of solitude. In fact, I reject it outright; working with editors is expected common practice after all. Perhaps I just afford more autonomy to my machine collaborators.
Together we did a full year’s writing in a week’s time, after all. The model could retain the full screenplay at once. But I could not. My role was to hold the identity of the film steady through every revision. This has always been the job of a conductor; a director. It is the one throughline that holds as the entire technological landscape changes around us. Which holds the labyrinth together.
By the end, the line between man and machine was less important than the chain of judgments in the process.
This is my vision. An athleticism in machine artmaking where both the machine and the human being operating are tested to cross the threshold of new limitations.
Finally, I read it. This particular model had a limitation. The end result came out as, still largely Cyrus-71, the same movie I had written, just turbo-charged. The machine collaborators were more conservative than I anticipated leaving things that worked well enough alone.
The result was not incomprehensibly living inside a machine’s mind, an unbreachable maze, like I had hoped. At the end of the process, it was still me. It was still 71. It was still within my curation, therefore, within my voice.
But there is no reason the process stops there, and it should certainly not stop by the quiet pressures of social taboo of machine writing. We owe it in the name of progress to explore this technology to its very limits. The ones who came before us did not have these tools, we do: so we owe it to them to use these tools and push farther than what limited them.
They would agree. I know this because I would agree with whatever lay in front of us, with the new hungers of the future youths, who want to push farther and distinguish themselves in a world where everything has already been done. They were us, were us, were us.
What is the point of technological progress if our will does not scale in parallel?
One can even push further than writing. Imagine the three core components of cinema built through the same pressure: writing optimized across conceptual passes, production optimized across visual and performative variables, and editing optimized across rhythm, transition, and assembly.
You would algorithmically arrive at the greatest films in history.
Match the extensive process of machine collaboration outlined here to the production and editing of a picture, and you now have three systems working in tandem. This is no longer describing art, it is describing complex machinery, its gears working to the tiniest minutia, synchronized in the tightest lockstep.
You begin to foresee an ultimate kind of cinema.
But it cannot happen on its own. The symphony will need its conductor. Even at the end of a hundred step filter, one will be surprised, it is still them.
For now, this essay is the trace left by writing after the machine has entered the room.
Writing with Machines is the opening essay of Volume 5 of The New Machine Cinema.
