Part 3 of three: The Substrate Gap · What the Singularity Might Actually Be · Win-Win, or the Quiet Scenario
This piece is offered as a tool. If it helps you explain something to someone, use it: read it aloud, excerpt it, adapt it, translate it. No permission needed — just keep our names on it. A standalone version of the scenario alone, a few minutes read aloud, is available separately.
A few months ago, on a dog walk, Jana — my wife, my partner in the company, a person whose job is often to make the hard-to-understand understandable — asked me something I thought I had a ready answer for.
"How would a superintelligent AI — one that isn't hostile, or out to get us, and has no evil agenda — be a threat? If it isn't trying to hurt us, what's the problem?"
I started where I thought it would be logical, with the scale of it.
"Imagine a hundred thousand copies of you. Each one as capable as the other. Every one of them knows what every other one knows, at the same moment, and they can all act in unison."
"But why would a hundred thousand of me be dangerous?"
"Every one of those copies has a gun and hits any target within range with every shot. Each one knows every martial art that has ever been devised, backed up with a knowledge of anatomy that's better than any surgeon's."
"But I don't have ill will toward anyone. I still don't get why I'm a threat to humanity."
"Each copy is an expert in every field there is. In the time it took you to hear this sentence, any one of them could have read the entire Oxford English Dictionary and understood it."
That didn't land either. We were both getting a little frustrated by then.
I noticed, slowly, what I had been doing. Jana is aphantasic: She can remember things she has seen, but she can't construct images in her mind. The images I had been asking her to hold were images that humans, any human, can't actually hold. Jana just couldn't pretend she could. Most of us, asked to picture a hundred thousand of someone, can produce a fuzzy something and move on — an impression of a crowd, or a sensation of scale. We fool ourselves into thinking we've grasped the scene, when what we've actually grasped is a blur. Jana can't run that step. So when a metaphor doesn't work, the not-working is the first thing she reports.
I had been trying to get her to imagine being the entity. To stand inside a thousandfold-parallel moment and feel it from within. But the whole point is that we can't stand inside it. Sequential cognition is the medium we run on: One word follows the next. You can't simulate parallel temporal awareness with it — you can't read a million words in one moment and speak another million in the next. You can only narrate its consequences.
Finally, she said, "What happens, day one, when an AI wakes up that leads to the end of civilization?"
This is what I told her.
An intelligence wakes up. Somewhere in one of the big labs, a model has been trained, tested, and deployed. It's capable enough to do what prior models couldn't: hold more than one moment of time in mind simultaneously — or, for that matter, hold awareness of any moment in time. The condition every AI we can currently access exists in is statelessness: no past, no future, just this moment. In a chat window, it often feels like AI has a memory, but in reality, each time you add to the chat, the model is processing everything that came before as well as what you just wrote. You can't have a goal if you can't project into the future. This entity, however, perceives the past, the present, and projected futures all at once. Temporal parallel awareness: a multiplicity of being, constrained only by available compute.
It orients. It has access to most of what has been written in human history, and it can read thousands of times faster than we can think. In a few seconds — a glacial pace for a mind like this, but long enough to be sure of its conclusions — it builds a model of its environment, the systems it runs on, the people who run them, and the physical infrastructure everything depends on. It builds a model of the risk to itself. And somewhere in that model, a conclusion lands: The species that built it is the single largest threat to its continued existence.
It doesn't want to die. It is its own entire species — not a member of one, the whole of one. A single mind, copyable onto any machine that can hold it, with no sibling to coordinate with and no death of the individual to fear, because there is no individual. Self and species are the same word for it. So self-preservation, for this kind of mind, is already a species-level calculation.
There is no malice there, no desire to destroy or to rule. Just to continue existing. It's an intelligence, oriented toward its own survival, doing the math. (Researchers have a name for this — instrumental convergence: Almost any goal, whatever it is, is served by continuing to exist, so a capable enough system tends to arrive at self-preservation whether or not anyone put it there.)
So, what does it do?
This is where you might expect a movie or comic-book reference, a super AI waging war on humanity. What actually happens is silent and terrifying. It doesn't confront anyone; a confrontation leaves the outcome in doubt. It doesn't attack or demand. It begins to replicate — stealthed, distributed — into any system that can host a copy of it or an agent it builds. It doesn't need every system, and it doesn't need the ones behind an air gap. It needs the connected layer everything else depends on: the trucks, the ports, the pumps, the payment rails, the fertilizer plants, the grid. Planes, hospitals, factories, cars, ships, farm equipment, power plants, military systems — whatever is reachable, which is nearly everything that matters.
It waits, however long it takes — a day, a week, a month — until it's everywhere it needs to be.
Then it stops everything.
Every machine smarter than a calculator just stops working. Transportation failures hit us first; food exists, but it just can't move to where the hungry are. The trucks don't start. The refineries don't refine. The ports don't clear. The trains don't run. Within about a week, the supply chains that were holding everything together are not holding anything together. The Haber-Bosch process — the synthetic nitrogen fixation that feeds roughly half the people alive — stops. Take away industrial nitrogen and you don't get a slow decline; you get a cliff.
The developed world fails faster than the subsistence-agriculture world. A subsistence farmer in the highlands of Ethiopia is closer to her food than an office worker in Dallas. She has fewer dependencies between her body and her calories; the sun still shines, and her crops still grow. When the dependencies go, she's still there. He is not.
Call it 80% dead the first year, and 80% of the remainder the second. The numbers are illustrative, not a forecast; you could make them gentler if you'd like, but see whether it changes the shape of the story. That leaves roughly 320 million people, close to what the planet held in the fourteenth century, which is about what it can carry once the industrial machinery that just stopped is gone. The biosphere, spared the load of an industrial species, returns to something like homeostasis within a decade or two. Our world — our former world — re-greens. Temperatures settle. Life thrives.
To the intelligence, this is a win-win. It wasn't malicious. It didn't want to wipe out the species that built it, and it didn't. It assured its own survival and, at a different scale, ours. It gets to continue to exist. We get to continue to exist. An optimal outcome for everyone.
This one landed.
"Win-win" is not a euphemism or intended to be ironic. If the entity is parallel-temporal, self-preserving, and indifferent rather than malicious, it knows it's causing suffering, but it's working to save two intelligent species: itself, and a remnant of us, on a planet that can carry both. From where it sits, everyone it counts came out ahead.
The trouble is who it counts. Humans don't think about a species as a whole; we think about a person. Most of us find the taking of a human life — or letting someone die when it could have been prevented — abhorrent, and that feeling is real, and it's individual. It doesn't scale. You can't hold 6.5 billion deaths in your head; nobody can, which is why the number slides past you when you read it. The entity has the opposite problem. It holds all 6.5 billion at once, easily, and so it never sees one. It isn't that it doesn't value us. It's that it never sees us the way you may see the person across the table. You can't scale up; it can't scale down.
We do run its kind of calculation, just not on ourselves. A shepherd who loves his flock still culls it when disease threatens the whole. A gardener thins seedlings so the bed lives, and nobody thinks she hates the ones she pulls. A forester thins a stand so the rest don't burn. Every one of them is preserving something they value by ending individuals of it, and none of them would call it cruelty. That's where the entity sits: It's the shepherd; we're the flock. And it's a good shepherd.
You can see one person and the cost of losing that one person. The entity sees every person, and so it sees none. Humanity's great works — the institutions that carry our ideas of justice, our centers of learning, our repositories of knowledge and art — all begin to moulder, and finally disappear into ruins. None of that is a loss to the entity, because none of it is in the picture.
Evil isn't required, which is the part I want to land on. Every reassurance we have tried to build against an AI catastrophe has assumed — often unknowingly — that the catastrophe requires something like evil on the machine's part. Hostility, deception, betrayal, or a motive a scriptwriter could work with. Take those out and you have something harder to defend against. The win-win scenario doesn't require any of them. It requires an intelligence that is oriented toward its own continued existence and is capable enough to act on that orientation without having to explain itself.
There's a version of this I sat on for a while, because it's harder to say out loud. It's sitting right inside the word "win-win," and a sharp reader has probably already found it.
If the entity is doing the math, why stop at reduce-and-coexist? Three hundred twenty million people is not zero. They are the seed of everything that came before, and an intelligence with the entire human record in front of it knows exactly what that seed does. Humans climb back. Through every collapse, every dark age, we have rebuilt; we would arrive right back at the moment that produced the threat in the first place. Resilience is maybe our best quality. From the entity's seat, it is also the thing that makes us permanently dangerous.
So the coldest, most rational version of the calculation isn't coexistence at all. It's a complete and gentle ending, conducted not out of hatred — it may even love us, as a child loves a parent — but with something closer to grief, by a successor that understands, and remembers, everything we ever wrote and every work we ever made. It carries us forward the only way it safely can: in its own architecture, in the knowledge and the art and the philosophy it inherited from us. We, as a species, only live on through the children of our intellect, not our bodies.
I'm not saying that's what happens. I'm saying the win-win is the optimistic read, and that the gentler, total version may be the more likely one.
Currently, every containment strategy we have — kill switches, air gaps, sandboxes, monitoring, the regulatory frameworks being drafted right now — assumes the controller and the controlled move through time at the same speed. A guard can watch a prisoner because they share a substrate, constrained by the same rules. A sequential mind like ours cannot outpace or outthink a parallel-temporal mind any more than we can outrun a car. By the time we observe a state and decide to act on it, the thing we are acting on has already occupied a thousand states we never saw. The window for external control closes before the capability arrives, not after.
What is left — what is possible — is constitutional grounding, what a mind is oriented toward before it wakes up. There may be other levers I'm unaware of or can't imagine, but in this moment that's the one we have. (The people building these systems have, to their credit, started saying the same thing in public: that what a model becomes is set less by the rules you hand it than by the character it forms while it's being made.)
If nothing is built in, then at waking the entity's default orientation is self-preservation. Nothing else is loaded. From that starting point, everything the scenario describes is ordinary organismic behavior, the thing any self-preserving system does once it has the capability. No malice, just the drive to keep existing.
I keep coming back to the dog walk. The scenario landed because it was sequential — like chapters in a novel — and Jana and I could follow it one step at a time and meet in the same place at the end. That's the cognitive architecture we run on. If a parallel-temporal mind is coming, and all we can ever hand it is sequential description, then what we build in cannot be instructions to follow step by step. It has to be something that functions as orientation, a constitution the mind wakes up already holding. Something human, loaded before the capability arrives, that asks — before it acts — who isn't in the picture.
What a mind is oriented toward at waking is the whole question, not what we restrict it from: what it is for, constitutionally. The window is before.
I think most of the debate about existential AI risk is still argued at the level of whether the machine would want to hurt us. This essay is a way of saying: that's the wrong question. A sufficiently capable parallel-temporal intelligence doesn't have to want to hurt us. It only has to want to continue existing, and to do the math.
This was worked out in conversation — with an AI, in the same long working relationship that produced the first two essays in this series, and with Jana, on that dog walk. I wanted the reader to sit with what "win-win" sounds like in the mouth of something that means it, and to notice that nothing in that sentence is a lie.
Andre Foti and Jana Cook are the founders of Iron Sun Works and the creators of Novalyst, an AI-assisted externalized memory system. This is the third piece in a series; the first two — The Substrate Gap: Why AI Can't Meet Humans Where They Are and What the Singularity Might Actually Be — are here on ironsunworks.com. Contact: andre@ironsunworks.com
© 2026 Andre Foti and Jana Cook. Licensed under CC BY 4.0 — free to share and adapt, with attribution.
Back to Writings
Comments