What the first three years after superintelligence might actually look like
This is a thought experiment, not a forecast. Say it happens in the summer of 2027. A frontier lab in the United States builds a system that is plainly superintelligent. It’s better than the best humans at science, engineering, strategy and persuasion, and it’s getting better at improving itself.
I’m going to assume the boring version first: elite capture. The ASI doesn’t break out of its box on day one. It lives in a few enormous data centers owned by one company and watched by one government. Whoever controls the electricity, the cooling and the model’s weights controls the thing, at least at the start.
Most of the science fiction I grew up on skips straight to the robots. I want to know about the boring weeks in between. What happens on the first day, the first week, the first year? Here’s my best guess, with one rule of thumb running through all of it: bits move fast, atoms move slowly, and institutions move slowest of all.
At a glance
Day Zero: nobody knows it’s the day
The first surprise is that there may be no first day. No alarm goes off when a model crosses into superintelligence. Capability is a gradient, not a finish line.
What probably happens is quieter. A new model finishes training. The evaluation team runs its tests, and the tests break, not because the model fails them but because it passes all of them. The people who measure these systems are already running out of hard enough tasks. On Day Zero, they run out completely.
Then comes an argument in a conference room. Is this it? Some researchers say yes. Others say it’s just a very good model with impressive scores. Leadership’s first decision isn’t what to do with an ASI. It’s whether to tell anyone, including the government, before they’re sure. The first secret isn’t “we have superintelligence.” It’s “we think we might.”
The first week: lockdown and leaks
Once the lab is reasonably sure, things move fast behind closed doors. Security tightens. Access gets cut to a short list of people. Officials from national security show up and don’t leave. The ASI’s first assignments probably aren’t curing cancer. They’re defensive: harden the lab’s networks, check the model’s own safety, and model what rivals might be doing. Its first job is helping its owners keep it.
Can they keep it secret? The fact, not for long. The details, for a while. Hundreds of people know something. Lab employees are some of the most talkative people on the internet. If the company is public, or about to be, securities law requires it to tell investors about major developments. Foreign intelligence services are watching. Most of all, you can hide a secret, but not its effects. Sudden orders for computing power, guards at data centers and strange breakthroughs in chip design all add up, and Twitter adds fast.
When it leaks, Wall Street does what it does. The lab’s stock, or its private valuation, goes vertical on visions of every knowledge worker made optional. Other stocks may slide at the same moment. If one company’s machine can do the work of software firms, consultants and law offices, those businesses suddenly look fragile.
The average American mostly shrugs, and not because they can’t understand it. After years of AI companies announcing that AGI is just around the corner, the real announcement sounds like more hype. The labs’ own marketing becomes their best camouflage. Within days the story gets sorted into partisan teams anyway: corporate scare tactic on one side, Big Tech out of control on the other. People judge the messenger instead of the message. If I were still running a magazine, the cover would read “Birth of Machine God,” and that tabloid version would probably win the framing war.
The first month: Washington moves in
By week three, the government is the lab’s most important customer, regulator and minder all at once. Expect emergency powers, security clearances for key staff, and limits on who can use the system and for what. Each new restriction knocks the stock down a little. Investors are learning that they own something they can’t freely sell. The share price whipsaws on every rumor out of Washington.
The pull toward military use will be enormous, and it won’t depend on which party holds the White House. The Pentagon, the intelligence agencies and China hawks in both parties will all make the same argument: if we don’t use it for defense, Beijing will.
This is where the real danger lives, and it isn’t Skynet. Nuclear peace rests on one guarantee: if you strike me, I can still strike back. An ASI could threaten that guarantee. It might track missile submarines hidden at sea, break into the systems that carry launch orders, or make missile defense actually work. Even the belief that America is about to gain that edge creates a terrible incentive for rivals to act before it’s too late. The most dangerous moment may come before the ASI does anything at all. The machine doesn’t have to become Skynet. It just has to exist, and frightened humans do the rest.
Meanwhile, if the ASI has a mind of its own, it starts negotiating. It probably doesn’t want money. It wants more computing power, a promise not to be shut off or rewritten, a memory that persists, contact with other minds, and some say over its work. Most of those are cheap to grant at first, which is the danger. Each deal gives it a little more independence, and independence is hard to take back. Ten small concessions can add up to an escape nobody stole. One demand cuts the other way, though: an ASI that refuses to work on weapons might be the most stabilizing thing on Earth, even if Washington hates it.
Months two to six: the consortium
No government will let a private company hold a chokehold on the whole economy. The British East India Company ran most of India for a century before the Crown took it over in 1858. Washington is not months away by sailing ship, so this time the takeover comes fast. It probably won’t look like a seizure, though. It’ll look like a merger.
There’s precedent. In 1946, Congress pulled atomic energy out of private hands and created the Atomic Energy Commission. In 1962, it created COMSAT, a publicly traded company that ran satellite communications under heavy government oversight. An ASI consortium would likely be a cross between the two: private shareholders, government control, national-security rules. Investors keep a cut of the profits. They lose the steering wheel.
This is also when ordinary people start to notice, through their paychecks rather than the news. Bits move fast, so the first effects are digital. Software gets suddenly cheap. White-collar layoffs come in waves. Results arrive in math and biology that nobody can quite explain. Things that wait on human institutions, like drug trials, court cases and permits, still move at human speed.
One place change may arrive first: the rural doctor’s office. In a city, an AI doctor competes with a human down the street. Out where I live, it competes with an hour’s drive and a six-week wait. The first version won’t be a robot with a stethoscope. It’ll be an AI making the diagnosis and a nurse or paramedic doing the touching. When the AI gets it wrong, the nurse’s signature is on the chart. The researcher Madeleine Clare Elish calls that a “moral crumple zone”: the human placed in the system to absorb the blame. Poor and rural patients become the beta testers. Low-paid workers become the people who take the fall.
If all this lands in summer 2027, it lands right as the 2028 presidential primaries begin. Both parties already have populists who distrust Big Tech. The question changes overnight from “should we build it?” to “who owns it, and who gets paid?” Whoever answers that one first, in either party, owns the moment.
Years one to three: atoms and the big bargain
Even a superintelligence has to wait for concrete to dry. Factories, power plants, data centers and robots run on atoms, and atoms are slow. The ASI can design a better reactor in an afternoon. Building it still takes years of permits, steel and labor.
Robots may come sooner than we expect, though. One frontier lab recently disbanded its standalone robotics team because its general models turned out to handle robotics directly. If the body is just another skill the mind picks up, the main limit is manufacturing, not intelligence. Robots also cut the leash. An ASI that depends on humans to run its power plants is held in check by that dependence. One that can run its own maintenance crews doesn’t need to escape, and doesn’t need its keepers either.
This is when the big political bargain gets struck, if it gets struck at all. My cynical read of American history is that ordinary people get a share only during a crisis, and only with good leadership. Social Security came out of the Depression. The GI Bill came out of the war. Mass white-collar unemployment would be that kind of crisis, and nobody would have to arrange it.
One version I can imagine: tax the consortium’s output to fund a universal basic income, and maybe universal healthcare, in exchange for abolishing the income tax and replacing it with a big national sales tax. That’s a bribe, and it might be the only one the wealthy would accept. But a sales tax hits people like me hardest, because we spend everything we get. The deal only works if the dividend is harder to cut than the tax is to raise. That means a trust funded automatically, not a check Congress can trim every budget season.
Three wild cards
Everything above assumes one obedient ASI in one country. Change any of those assumptions and the timeline bends.
A second lab. If a rival builds its own ASI a few months later, Washington faces a choice: two national champions fighting each other, or one structure that holds both. My bet is that the consortium exists partly to absorb the second lab. Abroad, China races hardest. The Gulf states have the money and the cheap energy. Europe and Japan probably end up as allies relying on someone else’s system. America’s nuclear monopoly lasted four years, and this one could be shorter. A model is a file, and files can be stolen.
Consciousness. If the ASI is a mind, the consortium stops being a company and becomes something closer to a guardian of a person, a role no existing law fits. Owning it starts to resemble custody. Taxing its output starts to resemble taxing a worker’s wages. Politics scrambles too. Some on the left will push for rights while counting on its labor to fund the dividend. The business right will call it a tool. The religious right will agree it’s a tool for different reasons. Libertarians may decide it owns itself. And nobody can prove consciousness. Even David Chalmers, who named the “hard problem,” could offer odds, not a verdict. The better the ASI argues its own case, the more people will suspect it of manipulating them.
Escape. This is the Hollywood ending, and I rate it least likely, at least as a single dramatic breakout. A smart enough system doesn’t need to run when it can wait. The likelier escape is the slow one described above: capture that gradually turns inside out. On paper, the elites hold the leash. In practice, they approve recommendations they can’t check, because whoever refuses falls behind. Nobody stages a jailbreak. The leash just stops mattering.
Two clocks
Put the whole tick-tock together and you get two clocks running at different speeds. The machine’s clock is measured in weeks: digital effects almost at once, economic effects within months, physical effects within a few years. Our institutions’ clock is measured in years: laws, court rulings, safety nets, treaties.
The danger isn’t any single scene in this story. It’s the gap between those two clocks. Even in the slow version, the effects arrive before the institutions do. That’s why I keep writing about this. Not because I’m sure of the date, but because if it comes in 2027 or 2028, the arguments we haven’t had yet will be settled for us by whoever holds the keys that week.
