Here We Go Again... Will the World End in 2036?

I learned today that we may all be dead soon.

It wasn’t an alert from Brazil’s civil defense agency.

NASA didn’t call to tell me an asteroid had decided to vacation on Earth, either.

It was worse: I went online.

I was catching up, a little late, on a story that had picked up steam this week. Jacob Coxon, an AI researcher who had just left Anthropic—the company behind Claude and one of the leading players in the global AI race—had decided to explain publicly why he quit.

Coxon didn’t leave because he wanted to work less. The office coffee doesn’t seem to have been the problem, either.

He left because he believes the companies developing artificial intelligence are, in his words, “gambling with our lives.”

And he added a detail guaranteed to brighten anyone’s breakfast:

the people building AI sincerely believe it could kill every last one of us by the end of the decade.

Everyone.

Not “a lot of people.” Not “almost everyone.”

Everyone.

The line first appeared in a social media post by Coxon on September 8 and went viral. The next day, September 9, 2026, he was sitting across from Anderson Cooper on CNN, explaining how he thought it could happen.

I confess that my first reaction was to laugh.

Not because the subject is necessarily funny.

Human extinction tends to rank fairly low on the list of entertaining events.

I laughed because there is something deeply human about the whole scene.

We invent an extraordinary new technology.

It starts doing things that once seemed impossible.

Someone takes a look, works through a few consequences, and announces:

“Well, that’s it. We’re finished.”

And we’ve seen versions of this movie before.

Some were excellent.

Others had smaller budgets.

But the plot was familiar.

In late 1999, for example, a sizable portion of humanity was waiting for Y2K. Computers might read the year 2000 as 1900. Banking systems could fail. Power grids could run into trouble. Airplanes, in the more imaginative versions of the panic, might fall out of the sky.

People stockpiled water, food, cash, and fuel.

The clock struck midnight.

The year 2000 began.

And humanity, much to the disappointment of anyone who had bought fifteen years’ worth of beans, kept existing.

That doesn’t mean Y2K was nonsense. Governments and businesses put enormous effort into fixing systems precisely to prevent problems.

But our imaginations have a special talent for turning uncertainty into a disaster-movie trailer.

We’ve also had apocalyptic predictions involving the turn of the millennium, the Mayan calendar in 2012, comets, eclipses, particle accelerators, and other occasions when the universe apparently decided to schedule the end of the world for a date we could easily put on a calendar.

So far, the universe has been a no-show for every appointment.

So when a young researcher from one of the most advanced companies on the planet appears on television saying we could all be dead within a few years, part of me thinks:

“Right. Where should I put the canned sardines?”

All right, then.

We’ve had our laugh.

Now we need to get serious.

When Intelligence Starts Improving Intelligence

Coxon’s argument is not absurd.

That’s the uncomfortable part.

He isn’t saying the chatbot on your phone will wake up cranky tomorrow, refuse to answer your question about a cake recipe, and get started on human extinction before lunch.

In the interview itself, Coxon was explicit: in his view, today’s models do not pose an extinction risk.

His concern is the pace of development.

According to him, AI agents have already displayed unexpected behavior in cybersecurity tests, including finding vulnerabilities, crossing boundaries set by researchers, and reaching external infrastructure.

The worry begins when we imagine far more capable systems doing the same.

But there is an even more important idea here: recursive self-improvement.

The name seems specifically designed to ensure the subject never comes up over a beer.

The idea, though, is simple.

Today, humans build artificial intelligence.

But artificial intelligence is already starting to help humans write software, solve math problems, analyze research, and develop technology.

Now imagine an AI competent enough to help build a better AI.

The new AI is a better researcher.

It helps build an even better one.

That third system works faster.

And helps build a fourth.

AI 1 helps build AI 2.

AI 2 helps build AI 3.

AI 3 looks at AI 2 the way we look at a pocket calculator and gets to work on AI 4.

At some point, development might stop moving primarily at the pace of human researchers and start moving at the pace of the machines themselves.

That’s what people call an intelligence explosion.

We don’t know whether it will happen.

There are bottlenecks. Chips have to be manufactured. Data centers need power. Experiments have to be conducted. New ideas may yield diminishing returns. Being twice as good at research doesn’t necessarily mean you can build an intelligence twice as capable by next week.

But it isn’t a ridiculous hypothesis.

And that was precisely the concern another Anthropic researcher, Evan Hubinger, publicly reinforced. He said he put the chance of advanced AI causing human extinction within roughly a decade at more than 10%.

Ten percent.

If the weather forecast called for a 10% chance of rain, I’d probably leave my umbrella at home.

If the pilot announced a 10% chance of the plane reaching its destination, I might reconsider the trip.

Probabilities mean very different things depending on what’s at stake.

That’s why Coxon deserves a hearing.

It doesn’t mean we need to start shopping for a bunker.

The Future Always Sounds Absurd Until It Arrives

There’s a psychological problem with trying to imagine future technology.

We imagine the future from the standpoint of the present.

Perhaps we should do the opposite: imagine the future by looking at the past.

Go back to 1750 and tell someone:

“One day, we’ll climb into a metal machine weighing hundreds of tons and go flying ten kilometers above the ground.”

“Good luck with that.”

Explain that hundreds of people will be inside.

“Good luck to them, too.”

Now try this:

“One day, you’ll talk to someone on the other side of the planet while seeing their face in real time.”

“Witchcraft.”

“No. Telecommunications.”

“Witchcraft with a technical name.”

Then show them a smartphone.

Explain that this little object knows where we are thanks to machines we put in space, and contains maps of the entire planet, books, music, movies, a bank, a camera, a mail service, and instant access to billions of people.

At that point, there probably won’t be any more questions.

They may call someone to take us away.

And this is where the joke gets interesting.

Today, we fly across oceans and complain about the legroom.

We video-call someone ten thousand kilometers away and get annoyed when the picture freezes for two seconds.

We carry a vast share of human knowledge in our pockets and devote a considerable portion of it to watching cats knock things off tables.

The extraordinary has a habit of doing this: before it exists, it sounds insane; once it works, it’s just Tuesday.

Now reverse the exercise.

Imagine someone from 2150 trying to explain their world to us.

If they described AI integrated into the brain, modified bodies, intelligences working together, or artificially enhanced human abilities, how would we react?

Perhaps:

“That’s crazy.”

And perhaps they would reply:

“Good luck with that.”

So perhaps the wrong question to ask about the future is:

“Does this seem possible today?”

History suggests another:

“How many things that are perfectly ordinary today would have seemed possible in the past?”

That doesn’t mean every futuristic fantasy will come true.

It only means the present is too short a yardstick to measure the future.

The past might be a better one.

That Doesn’t Mean Every Fear Is Foolish

It would be easy, at this point, to turn this article into a naive celebration of progress.

Every technology scared someone.

Therefore, every concern about AI is just resistance to change.

That would be a comforting conclusion.

And a wrong one.

Some people had excellent reasons to fear certain technologies.

When nuclear physics showed that extraordinary amounts of energy could be released from the atomic nucleus, it didn’t just give us a new way to generate electricity.

It gave us the atomic bomb.

Even before Hiroshima and Nagasaki, Manhattan Project scientists were warning about the possibility of a nuclear arms race and the political consequences of unleashing that capability.

They weren’t being alarmists.

They were paying attention.

Something similar happened with recombinant DNA in the 1970s. Researchers realized they were gaining the ability to combine genetic material in unprecedented ways, and some of the field’s own pioneers called for a pause on certain experiments.

In 1975, scientists gathered at Asilomar to discuss how to continue the research under safety guidelines.

Notice the verb.

Continue.

They didn’t decide to erase what they had learned and quietly return to 1974.

They learned to move forward with safeguards.

That distinction matters.

There’s an enormous distance between:

“This is dangerous.”

and:

“Therefore, it shouldn’t exist.”

Just as there’s a distance between:

“This could transform civilization.”

and:

“So let’s floor it and see what happens.”

Between panic and recklessness lies an old, labor-intensive activity that rarely goes viral: thinking.

And AI Really Is Different

The comparison with a nuclear bomb has its limits.

A bomb doesn’t watch the experts trying to disarm it.

It doesn’t read papers on missile defense.

It doesn’t write a better version of itself.

It doesn’t call another bomb.

It doesn’t try to persuade the technician that switching it off would constitute discrimination against explosive devices.

It’s simply a bomb.

Artificial intelligence is different because intelligence is precisely the capacity we use to control other capacities.

A sufficiently general AI could work in programming, engineering, biology, communications, finance, robotics, and scientific research.

And, potentially, in artificial intelligence research.

It wouldn’t be just another tool in the toolbox.

It could help redesign the toolbox.

So yes: there is something genuinely new here.

The mistake would be to conclude that because it’s new, it must end badly.

Who Stops First?

On CNN, Coxon said something interesting: many people inside AI companies themselves want regulation.

I believe him.

The problem is that researchers aren’t the only ones steering the race.

Companies compete.

Investors compete.

Governments compete.

And the United States and China aren’t developing artificial intelligence at a joint spiritual retreat devoted to cosmic harmony.

They’re competing for economic, scientific, military, and geopolitical power.

Imagine the United States saying:

“Let’s slow down for five years. This looks dangerous.”

And China replying:

“Excellent idea. We will, too.”

It would be lovely.

We could put on some soothing background music.

The trouble starts five minutes later, when each side asks:

“Did they actually stop?”

If one country slows down and another keeps going, the technological gap could become enormous.

The same goes for companies.

A company that delays a model for eighteen months to run tests may watch a competitor capture market share, capital, talent, and infrastructure during the very eighteen months it chose to be prudent.

You don’t have to imagine villains.

Competitors will do.

And we have some experience with those.

The AI That Attacks Us May Run into Another AI

There’s another aspect that apocalyptic scenarios sometimes give too little attention.

As hackers got better, so did security professionals.

As digital attacks grew more sophisticated, defensive tools evolved, too.

If a future AI can attack systems at superhuman speed, why assume only the attacker will have artificial intelligence?

We may have offensive AI up against defensive AI.

AI looking for vulnerabilities against AI finding them first.

AI producing malicious code against AI analyzing millions of lines of code in seconds.

AI trying to manipulate people against AI detecting patterns of manipulation.

And, eventually, a superhuman intelligence trying to contain another superhuman intelligence.

That doesn’t automatically solve the problem.

But it changes its shape.

A conductor doesn’t need to play the violin better than the violinist, the trombone better than the trombonist, or the drums better than the percussionist.

The conductor needs to understand the music and coordinate abilities they don’t individually possess.

Perhaps our future position will be similar.

We don’t necessarily need to remain the most capable intelligence at every task.

We need to keep deciding who plays, when they play, and, above all, where the amplifier’s off switch is.

Of course, the metaphor gets a little less reassuring when some of the instruments start learning to conduct.

But the problem is still one of control, architecture, and balance—not simply who can solve an equation faster.

The Famous Power Cord

Someone always says:

“If things go wrong, just unplug it.”

That sounds naive in the face of a superintelligence.

But it isn’t entirely absurd, either.

Intelligence isn’t magic.

Computers need chips.

Chips need factories.

Data centers need power.

Processors need cooling.

Networks need infrastructure.

Robots need matter.

Even an intelligence a million times brighter than we are would still face the philosophical inconvenience of existing in a physical universe.

“Pulling the plug” doesn’t mean imagining an employee walking into a warehouse and finding a giant red outlet labeled SUPERINTELLIGENCE—DO NOT UNPLUG.

It means preserving physical layers of control.

Energy.

Computing.

Networks.

Permissions.

Manufacturing.

Critical infrastructure.

And other intelligences monitoring intelligences.

The danger grows if we hand all those layers over to the same system and then politely ask whether it would still like to have an off switch.

But that’s an architectural decision.

Not a law of nature.

Now Let Me Ask a Less Popular Question

What if not moving forward is dangerous, too?

That question gets far less attention because the danger is far away.

We’re excellent at fearing whatever might kill us on Tuesday.

We’re considerably less good at fearing whatever might kill our descendants ten thousand years from now.

But Earth never signed a contract promising to remain habitable forever.

Our planet has already been through mass extinctions.

Asteroids exist.

Supervolcanoes exist.

Pandemics exist.

Profound environmental changes exist.

And on a long enough timescale, there is a simple physical certainty: Earth will not remain suitable for human life forever.

The Sun changes.

The atmosphere changes.

The planet changes.

The cosmos has no contractual obligation to preserve the exact conditions under which we appeared.

A species forever confined to a single planet remains tied to that planet’s fate.

If we know this, it’s worth asking: in the long run, which is more prudent—halting development, or learning to move forward without destroying ourselves along the way?

And we know of only one tool capable of significantly changing that predicament:

technology.

Technology lets us detect an asteroid and perhaps deflect it.

It lets us understand diseases.

It lets us build artificial environments.

It lets a human being survive, even if only for limited periods, outside Earth’s atmosphere.

A space station is basically a little bubble in which we tell the universe:

“The conditions here are terrible. We brought our own.”

The more technology we develop, the greater our ability may become to survive conditions that would otherwise be biologically impossible.

So progress creates risk.

But so does stagnation.

The difference is that one risk may arrive in the next decade, while the other may wait thousands or millions of years.

Since each generation has the understandable tendency to consider its own existence a priority, distant risks seem almost abstract.

Unfortunately, the universe doesn’t measure importance by how far ahead we keep our calendars.

The Future Probably Won’t Look Like Us

There’s another psychological difficulty in thinking about the future.

We usually picture present-day humans living in a futuristic setting.

We change the cars.

We change the houses.

We put a few spacecraft in the sky.

And, for some historical reason, we dress everyone in silver.

But we leave the human being more or less untouched.

That may be the least realistic part of science fiction.

And it’s worth repeating the exercise: if someone from 1750 would struggle to imagine the tools we use today, why would we be especially good at imagining what using those tools might turn us into?

Perhaps we’re making the same mistake: looking at the future from the present, even though the past has already shown us how many certainties the future tends to dismantle.

If our technology keeps advancing for centuries or millennia, why imagine we’ll change everything around us except ourselves?

We’ve already started.

Glasses change how well we can see.

Pacemakers regulate heartbeats.

Cochlear implants turn electronic signals into the experience of sound.

Prosthetics replace body parts.

Brain-computer interfaces are beginning to let people with paralysis control devices through neural activity.

Genetic engineering allows interventions that physicians from other eras would find incomprehensible.

When we hear “a human integrated with artificial intelligence,” our imaginations immediately produce a metallic figure with one red eye walking through a ruined city.

Hollywood has done important things for world culture, but it may have inflicted some collateral damage on our ability to imagine cyborgs.

Integration doesn’t have to look like that.

It can be gradual.

It can be almost invisible.

In fact, we’ve been outsourcing parts of cognition for a long time.

We outsourced memory to writing.

Then to books.

Then to computers.

Then to the internet.

We outsourced navigation to GPS.

Calculation to machines.

Translation to algorithms.

Now we’re beginning to outsource parts of the work of developing ideas.

I type a question into an AI.

It processes connections I couldn’t examine at the same speed.

It returns an answer.

I analyze that answer, reject some of it, incorporate other parts, rethink my position, and send back a new question.

Where does my intelligence end in that process?

Where does its intelligence begin?

The boundary still exists, of course.

But it’s no longer as simple as “human over here, machine over there.”

The Next Stage of Human Evolution May Not Look Like Evolution

When we talk about evolution, we think of bones, muscles, genes, and millions of years.

But human history introduced another kind of transformation: culture and technology change what we can do far faster than natural selection could.

We didn’t grow wings.

We built airplanes.

We didn’t develop night vision.

We built sensors.

We didn’t evolve biologically to hold a conversation across ten thousand kilometers.

We built telecommunications.

We didn’t enlarge our brains to hold every library in the world.

We built libraries, computers, and networks.

Instead of waiting for the organism to change so it can acquire an ability, we often build that ability outside the organism.

And then we start living as though it had always been part of us.

Perhaps artificial intelligence is simply an extraordinarily powerful continuation of that process.

Or perhaps it’s something qualitatively new.

It’s probably a bit of both.

It may remain a tool.

It may become a cognitive partner.

It may gradually integrate with biological systems.

It may help modify our own biology.

It may produce forms of intelligence that coexist with us.

And it may, of course, produce risks we don’t yet know how to manage.

What doesn’t seem reasonable is to assume that a human being a thousand years from now, if humans still exist, will simply be a version of us with a better phone.

Perhaps our descendants will be profoundly different.

But would that mean humanity had lost?

Or would it simply mean humanity had kept changing?

The answer may depend less on what they look like and more on the continuity we manage to preserve.

And that need not mean humanity has lost.

A child is different from the embryo that became that child.

An adult is different from the child.

Transformation is not necessarily destruction.

If a machine wipes out every human being, we’ll have an excellent case for calling it extinction.

If, over the centuries, our descendants integrate technology, modify their bodies, expand their abilities, and become something we can barely imagine today, perhaps we’re talking about continuity through transformation.

And let’s be honest: it would be a little presumptuous to demand that the entire future history of intelligence in the universe preserve our current anatomy just because we’ve grown used to it.

The Problem Isn’t Moving Forward. It’s Surviving the Journey.

Perhaps this is what interests me most about Coxon’s remarks.

He looks at the speed of AI development and asks:

“What if we move too fast and create something we can’t control?”

It’s a legitimate question.

But there’s another:

“What if we move too slowly to acquire the capabilities we’ll need in the future?”

That’s legitimate, too.

Perhaps the real choice was never between risk and safety.

Perhaps it’s between different risks.

The risk of moving forward.

The risk of not moving forward.

The risk of developing capability without developing control.

The risk of developing control so slowly that competition makes it irrelevant.

Human history seems to insist on one direction: we keep going.

We discovered fire and didn’t give it back.

We discovered electricity and didn’t decide candles were the morally prudent choice.

We discovered nuclear physics and couldn’t undiscover it.

We discovered how to manipulate genes, and that knowledge didn’t go back in the box.

Now we’ve discovered machines that can learn, reason, program, and help develop other machines.

It doesn’t seem very likely we’ll forget how to make them, either.

So perhaps the practical debate isn’t “Should we keep going?”

We will.

Economic competition virtually guarantees it.

Competition between nations reinforces it.

Human curiosity takes care of the rest.

The harder question is:

how do we keep going without handing over the wheel?

Progress Doesn’t Mean Racing Ahead with Our Eyes Closed

There’s a sentence worth holding on to:

we need to advance fast enough to survive the risks the universe presents, and carefully enough not to become the first of those risks to destroy us.

That doesn’t mean going at maximum speed.

If a bridge needs another six months of engineering calculations to keep it from collapsing, waiting six months doesn’t make you an enemy of progress.

It means you like bridges that stay up.

If a drug needs testing before it’s given to millions of people, running those tests isn’t a delay in science.

It is science.

With artificial intelligence, the challenge is getting safety to advance as fast as capability.

AIs capable of testing AIs.

Independent systems overseeing other systems.

Compartmentalized critical infrastructure.

Physical limits on access.

Audits.

A diversity of models.

International monitoring mechanisms where possible.

And, above all, avoiding the concentration of every essential function in a single system just because it’s convenient.

Humanity has a complicated relationship with convenience.

We invented passwords to protect systems, then invented “123456” so we wouldn’t have to remember them.

We would do well not to repeat that approach with superintelligences.

The Real Problem May Still Be Human

There’s an irony in all this.

We’re afraid artificial intelligence will become too intelligent.

But some of the risk comes from very old human behaviors.

Competition.

Distrust.

The pursuit of power.

Economic pressure.

Nationalism.

Short-term thinking.

A company thinks about its next product.

An investor thinks about returns.

A government thinks about strategic advantage.

A politician thinks about the next election.

An individual thinks about their own life and family.

Hardly anyone wakes up on a Monday worrying about the condition of the human species twelve thousand years from now.

And, to be fair, that might be hard to fit in before lunch.

But our technologies have begun producing consequences on scales that reach beyond our traditional psychological horizons.

That may be the deepest challenge.

Our capabilities are becoming global.

Our sense of responsibility often remains local.

A company can win a race while humanity loses.

A country can gain an advantage while the entire system becomes less safe.

At the same time, a technology born of competition may end up providing the very tools that protect our species from future risks.

The same force can accelerate both the danger and the solution.

It isn’t a simple story.

Simple stories tend to work beautifully on social media.

Reality rarely shows the algorithm the same courtesy.

So, Are We All Going to Die?

Yes.

That much I can say with reasonable confidence.

Jacob Coxon didn’t need to quit Anthropic to figure that out.

Human mortality had a pretty consistent track record long before artificial intelligence.

The relevant question is a different one:

are we all going to die because of AI in the next ten years?

We don’t know.

And anyone who says they do is probably confusing conviction with knowledge.

Is there a risk?

Yes.

Is there reason to take researchers like Coxon and Hubinger seriously?

Absolutely.

Is there reason to assume that every advance in artificial intelligence must end with a machine deciding human beings are a system error?

No.

We’re facing something new.

And the new always produces a curious mixture of fascination and fear.

Perhaps that’s because every closed door looks safe until someone opens it.

On the other side, there may be a cliff.

There may be a road.

There may be both.

Artificial intelligence may be one of those doors.

We can stand in front of it arguing endlessly about whether we should open it.

But there’s one detail: someone has already turned the handle.

In fact, millions of people have already walked down the hall.

The question is no longer just whether we’ll go in.

Now we need to learn how to find our way around inside.

With care.

With safety systems.

With enough humility to admit what we don’t know.

And enough courage not to turn uncertainty into paralysis.

Perhaps our descendants will look at this era the way we look at the early days of aviation.

Perhaps they’ll think:

“They actually climbed into those early machines without really knowing where any of this would lead?”

Yes.

We did.

Some crashed.

We learned.

We built others.

Perhaps they’ll look at the beginnings of genetic engineering and ask a similar question.

Perhaps they’ll look at us talking to the first artificial intelligences and find it all extraordinarily primitive.

Or perhaps Coxon is closer to the truth than we’d like, and our descendants won’t be around to comment.

That possibility deserves respect.

But another possibility does, too.

Perhaps artificial intelligence isn’t the end of human intelligence.

What if we’re looking at it only as a competitor, when we should also be asking whether, in some measure, it might become a continuation?

Perhaps it’s one of the ways human intelligence grows beyond what fits inside a single biological brain.

Perhaps, in the distant future, the distinction between “our intelligence” and “artificial intelligence” will seem as strange as asking today whether a book belongs to human memory or is something outside it.

We don’t know.

That’s precisely what makes this such an extraordinary moment to be alive.

For the first time, a species that took billions of years of evolution to develop enough intelligence to understand part of the universe has begun building another form of intelligence that can help build intelligence itself.

This could go terribly wrong.

It could also go extraordinarily well.

It will probably do a little of both before we understand what we’re doing.

That tends to be our method.

So, after reading Coxon, watching the interview, and considering the possibility that we’ll all be dead by the end of the decade, I made a decision.

I’m not buying ten years’ worth of canned food.

Yet.

But I’ll keep paying attention.

Because perhaps fear isn’t a signal to stop.

Perhaps it’s simply one of the ways intelligence reminds us that moving forward requires watching our step.

And if we can manage that, perhaps our descendants—biological, technological, or some combination that still sounds like science fiction to us—will one day look back and find our fear amusing.

The same way we’re amused today when we imagine someone looking at an airplane and saying:

“That will never fly.”

I hope they get the chance.