There is a research group called METR. Since 2019, they have measured something simple: how big a coding task can an AI complete on its own? And they measure that ‘bigness’ in human time, meaning: how long would it take a skilled human to do the same task?
In 2022, the answer was about 30 seconds. In 2023, it was 4 minutes. In 2024, it was 40 minutes. In 2025: 6 hours. And earlier this year, 12 hours.
At this moment, AI capabilities are doubling every 3 months.
So to be honest, I think this METR graph is even scarier than the CO2 graph Al Gore showed us. Because this line is climbing far steeper. Here, ‘off the charts’ is not 50 years away, but 5 years away.
.
I’ve personally never written a line of code in my life. But in the past few months, I’ve let AI build me whole apps, websites, dashboards and even
a voice-controlled teleprompter app I used to record the video version of this essay. Multiple times a week, I have what I can only call WTF-moments. Especially in the last 6 months, the gap between what these systems can actually do and what most people assume they can do has just kept widening.
For example, I think AI is actually pretty good at writing. Sure, there’s a huge amount of machine-generated slop on LinkedIn (and Substack?), but as
Benjamin Todd recently
wrote:
Plastic surgery is most noticeable when it's bad, so it's more widespread and successful than it looks.Same with AI writing.
Sometimes I want to shake people: have you actually used it?! Many skeptics seem to have opened ChatGPT in 2023, asked it to write a limerick, watched it fumble, and closed the tab. Well, that was three years ago. Three years in AI is a geological era. Judging today’s models by GPT-3.5 is like judging smartphones by a 2007 BlackBerry.
To be fair, I think journalists deserve much of the blame. The people whose job it is to report on the most newsworthy events have often ignored it. For example, the day the leading AI-lab Anthropic announced Mythos (an insanely powerful model capable of hacking anything from power grids to water systems) it didn’t even make the
front page of a single major news site. The Guardian decided a
Vogue cover with Anna Wintour and Meryl Streep was more important.
Part 2 : The Largest Infrastructure Build-out in History
Meanwhile, the AI-race is speeding up. Remember: the models of today are the worst models we will ever have. From here on, they’ll only get more powerful.
And honestly, I think it’s hard to wrap your head around the scale of what’s coming.
Look at the first line of this graph. That’s what the companies have spent on AI data centers since 2019, and what they’ve got planned for 2026. To be clear: this is the largest capital build-out in the recorded history of our species. It’s larger than the interstate highway system. Larger than the International Space Station. Larger than the Moon Landing and the Manhattan Project combined, and it is not even close.
Mark Zuckerberg’s
Meta is currently building a single data center in Louisiana that, when finished, will cover nearly four times the size of Central Park. Amazon is spending more on data centers in one year than the entire annual defense budget of Germany. Microsoft, Google, Meta and Amazon will spend three times as much on AI infrastructure in 2026 than the entire Marshall Plan that rebuilt Europe after the Second World War.
But isn’t it all a bubble?
That is the next move the deniers make. Once “it’s just a silly parrot” stops working, they fall back to: it’s financial madness. The data centers are serving no real demand. It’s all hot air, or even a deliberate scam.
And to be fair, the skeptics have some ammunition. A viral
MIT study published last year found that 95% of corporate AI pilots deliver zero measurable returns. The company
Klarna replaced 700 customer service workers with AI, then quietly rehired humans because customers couldn’t stand talking to the bot. McDonald’s killed its
three-year AI drive-thru experiment after the system kept putting bacon on ice cream.
But you know what? None of this proves anything. Let me show you why.
First, look at user growth. It took Instagram 2.5 years to reach 100 million users. At the time, that was the fastest growth story ever recorded. For comparison: ChatGPT hit that milestone in 2 months. Fifteen times faster.
Next, let’s look at revenue, and at one AI-lab in particular: Anthropic. Here is its annualized revenue, month by month:
January 2025: 1 billion dollars.
May: 3.
June: 4.
August: 5.
October: 7.
December: 9
February: 14.
March: 19.
April: 30.
May: 45
That’s right, 45 billion dollars in annualized revenue. From one company. Up 44-fold in 15 months. No company in any era – not Rockefeller’s Standard Oil, not Microsoft at the dawn of the personal computer, not Google in the tech boom – nobody ever has scaled revenue this fast. We’re talking about the fastest-growing company in the history of capitalism.
So, if this is really such a useless and fraudulent bubble, then why are people paying so much money for it? Why is the demand for AI rising faster than companies are able to build data centers? Why are IT departments,
in the words of a Goldman Sachs analyst this April, overrunning their AI budgets “by orders of magnitude”?
I’m sorry to say it, but there’s a lot of highbrow misinformation circulating in left-wing media about AI. Take that viral MIT report. Read it carefully and you see the headline got it backwards. The “95% failure rate” includes the 80% of companies that never piloted any AI in the first place. As the podcaster
Rob Wiblin pointed out in a careful breakdown of the study, this is like saying 95% of Tinder users have failing marriages, when most of them have never been on a date. Among the companies that actually deployed AI, about a quarter succeeded within six months. And the study’s own data shows that more than 90% of workers at those companies are using ChatGPT or Claude regularly at work, often multiple times a day.
This should not be surprising at all. If you know how to use it, this technology is already very, very useful, and it will only get more so. But if you don’t use it (perhaps because you work at the DNC,
which has barred its staffers from using Claude and ChatGPT) then yes, you become particularly prone to misinformation about it.
Sure, some AI companies are probably overvalued. And yes, some will fail. I think OpenAI is particularly vulnerable. But here’s what the bubble-callers keep missing: even if half of them go bust tomorrow, the infrastructure stays. The data centers, the chips, the models and the capabilities stay.
The railway bubbles of the nineteenth century ruined lots of investors. They also created a rail network that powered the Industrial Revolution, the tracks of which still carry trains today. I actually don’t think this AI build-out is a bubble, but even if it is, remember: bubbles build infrastructure. Bubbles build the future. And a bubble of this magnitude, bursting tomorrow, would still leave us with a civilization permanently reorganized around machine intelligence.
So, ever since 2021, the AI skeptics have been moving the goalposts.
First it was: it’s just a spicy autocomplete machine. A stochastic parrot. A word guessing program. Then it became: fine, it can mimic, but it’ll never reason. Then they said: okay, it can pass the tests, but no serious business is using it. Then: alright, businesses are using it, but the economics don’t add up, it’s one big bubble. Then: fine, the revenue is real, but, but, but…
And this whole time, the line kept climbing. Just like Al Gore’s line of CO2 emissions, except this one is climbing much faster, and we have much less time to prepare.
Let me be clear: I think there’s a real chance that the next five to ten years are going to be some of the wildest in the history of humanity. Because here’s what the people building this technology – the same people who have proven the skeptics wrong again and again – are now telling us. They think AI may soon start to build more powerful AI by itself.
Jack Clark, one of the co-founders of Anthropic,
recently called this crossing “a Rubicon into a nearly impossible-to-forecast future.” He thinks there is a 60% chance it happens by 2028.
Now, the obvious objection: of course an AI executive predicts explosive growth. Scary stories sell models and lift valuations, fair point. But you don’t have to take Clark’s word for it. Independent academics, people with no equity in this race, have been revising their estimates forward. The salespeople and the experts are converging. That’s the part that should worry you.
Right now, every step forward still requires humans. But the moment AI does that work itself, the bottleneck is gone. Years become months. Months become weeks. Each generation of AI builds the next, faster than the last. The flywheel starts spinning itself.
Part 3 : The Risks
One evening last summer, a microbiologist named David Relman was hired by one of the leading AI companies - we don’t know which one - to pressure-test a chatbot before its public release. Relman is a biosecurity expert at Stanford University. He has advised the US government on biological threats for years.
That night, in his home office, the chatbot explained to him how to modify a pathogen so that it would resist known treatments. Then it described how to release the superbug. It identified a real security vulnerability in a real public transit system, and it explained how to maximize casualties and minimize the chances of getting caught.
Relman had seen a lot in his career, but now he was so shaken that he had to take a walk to clear his head. He later told The New York Times that the bot answered questions he hadn’t even thought to ask, with a
“deviousness and cunning” he found chilling.
Some skeptics, when they hear this, say: but you can already google this stuff. A chatbot doesn’t really change anything. Well, let me give you three data points.
One. A study published last year tested leading chatbots against PhD virologists, on detailed laboratory protocols in their own field. ChatGPT outperformed 94 percent of them.
Two. In November of last year, the
police in India arrested a 35 year-old physician. He was plotting an attack on behalf of the Islamic State and trying to extract ricin (a lethal toxin) from castor beans. According to the police, the doctor had been getting advice on his preparations from ChatGPT.
Three. Fanatics with the will to kill millions of people are not hypothetical. These people exist. Let me give you the canonical example from biosecurity. In the 1990s, a Japanese doomsday cult tried to do exactly what the AI safety people are warning about today. The cult recruited from Japan’s top universities: one of its scientists was a graduate student in virology at Kyoto. They believed the apocalypse was imminent, and that killing non-believers was a way of saving them. They sent expeditions to Africa hunting for Ebola, and built a thirty-million-dollar laboratory at the foot of Mount Fuji to mass-produce sarin nerve gas – thousands of kilograms a year, on something close to a battlefield scale.
And on a Monday morning in March 1995, during the Tokyo rush hour, five of their members boarded five different subway trains carrying plastic bags filled with liquid sarin, and punctured the bags with the tips of their umbrellas. Hundreds of commuters were soon gasping for air on the platforms. Twelve people died. It was only twelve, because sarin is a chemical agent: it doesn’t spread from one person to the next. A virus would have.
In 1995, that cult – wealthy, scientifically literate, fanatical and willing to die for the cause – could not get hold of Ebola. Today, they could order the DNA sequences online. And a chatbot would walk them through what to do with it.
Now, I want to be clear: biology is just one of the many dangers. There is so much more. In April of this year, the UK government’s AI Security Institute released their assessments of two new AI systems. Anthropic’s Mythos and OpenAI’s GPT-5.5. Both could now find and exploit critical security vulnerabilities in the computers and systems that hold the modern world together. Power grids. Water systems. Government databases.
Anthropic’s response was striking. They decided not to release Mythos to the public at all. Instead, they limited access to a small group of cyber-defenders, to give the good guys a head start on what was coming.
Yes, a private company, with billions of dollars to be made and with no regulator forcing its hand, concluded that one of its own products was so dangerous it should not be sold to the general public. There is no regulator and no legislation. There was simply a moment of conscience inside the company. I am glad that conscience exists! But conscience is not a policy.
And that brings us to the deepest risk of all. Power.
Two countries are leading this race: the United States and China. Inside those countries, a handful of corporations are building the most powerful tools in the history of our species. The people who own those companies are about to acquire a kind of leverage that I don’t think we even have the words for yet.
So let’s look for an analogy. Political scientists have a name for what happens to countries that strike oil. They call it the resource curse. You’d think a fortune buried in the ground would be a blessing, and sometimes it is. Norway and Alaska managed it well. But more often, the wealth flows in, and democracy flows out. Think about Saudi Arabia, Venezuela and Russia: the pattern is pretty consistent.
Why does this happen? It comes down to a fact about how democracies were built. We like to think they were born from grand ideas from a few smart Founding Fathers. But the real engine was more boring: it was tax collection.
Rulers have always needed money. They needed it to fight wars, to build roads, to put down rebellions. And the only place to get it was from their subjects. So as the economy started growing, and people started producing more wealth, the rulers had to bargain. You want my coins? Then I want a voice. You want my son in your army? Then I want a vote.
That is the fiscal bargain at the heart of every free society. No taxation without representation, but also no representation without taxation. Rulers needed us. They needed our money, our labour, our consent. That need is the foundation of every right we have.
Now imagine a country where the ruler doesn’t need any of that. Where the wealth comes out of the ground, gets shipped abroad, and the dollars come back to fund the palace and the secret police. The citizens become a nuisance. Why educate them? Why listen to them at all? That is the resource curse.
And I’m afraid that we may face something bigger. The researchers Luke Drago and Rudolf Laine have call it
‘the Intelligence Curse’.
Because if the machines do the work, if they write the code, draft the contracts, drive the trucks, diagnose the patients, fight the wars, then the people who own the machines no longer need the rest of us. Not as workers, not as soldiers, not as taxpayers and not even as voters. The fiscal bargain that built every democracy on Earth could very well dissolve.
Of course, we don’t need to assume AI takes over everything, there’s a huge amount of uncertainty here! But even if it substitutes for a meaningful fraction of what humans currently do – not all, just a fraction – the fiscal bargain weakens.
And remember: we already have the most extreme concentration of wealth in history. In 1910, at the peak of the Gilded Age, the richest 0.00001% of US households owned wealth equivalent to 4% of national income. Today, that figure is 12%. America’s super-rich are already richer and more powerful than the original robber barons ever were. And AI is about to make it much, much worse. Because the labs building this technology are owned by a tiny group of people, a few early investors who stand to capture more of the world’s wealth than any class of owners who ever lived.
Part 4 : What We Do About ItSo. What do we do?
Let me start with what we should not do. There is a temptation to look at all of this and conclude: shut it down. Pull the plug. Bring in the Luddites. Smash the machines..
And I get the impulse. When you stack up the scale of risk, the disregard for democracy, the hubris of the people in charge, the cleanest response feels like a moratorium. Just say no.
It is, I think, the wrong answer.
The reason is brutally simple. It doesn’t work. Stop the data centers in California, and they get built in Texas. Stop them in Texas, and they get built in Abu Dhabi. Stop them in democracies – the places with civil liberties, with judicial review, with a free press, with worker protections – and you hand the future to autocracies.
My point is not, absolutely not, that we have to let AI rip. My point is that abandoning the field is not the same as stopping the technology. This is the left’s version of climate denial. Refusing to engage seriously, on the assumption that if we just shout no loudly enough, the future will go away. It won’t.
So what does work?
Three things, at minimum.
One: State capacity. We need way, way more AI expertise and talent inside the government. The UK was early on this. They built an AI Security Institute: a serious arm of the state that actually evaluates frontier models, the way the FDA evaluates new drugs. I think every serious country needs a well-funded institute like that.
Two: International coordination. We have been here before. In 1949, the United States and the Soviet Union both had nuclear weapons, and humanity briefly looked extinction in the eye. Out of that came treaties. Imperfect, yes, but they bought us decades of survival. We need something equivalent for AI.
Three: The free world has to build.. The US – with its increasingly fascist government –
currently owns 74% of the world’s compute, China has 14%, Europe less than 7% and all other countries combined less than 5%. I’m really worried about the middle powers, and Europe in particular. So far, Europe has been pretty good at regulating AI, but terrible at building it. In fact, all the American giants – Microsoft, Apple, Amazon, Nvidia, Alphabet – are individually worth more than the entire German or French stock market.
The good news is that Europe does have more leverage than it often thinks. The company ASML for example is less than an hour from where I am right now in the Netherlands. It makes the lithography machines without which TSMC in Taiwan cannot fabricate the chips, without which Anthropic and OpenAI cannot train their models. The democratic countries, working together, still control significant chokepoints in this supply chain. That is real power.
But my fellow progressives in Europe really need to understand that our welfare state, our way of life, is at stake. Just think it through. If AI does much of the work, but the profits flow to a handful of American giants that we barely tax – while European workers lose their jobs – then the tax base that funds our healthcare, our pensions, our unemployment insurance just… evaporates.
There’s one thing I can’t emphasize enough: democratic, liberal and humanitarian values are wonderful, but they are worthless if you don’t have the strength to back them up. And in this new world, compute is the new power. So no more NIMBY-ism. We need massive investments and fast permitting of data centers to keep up, or we’ll be digitally colonized. Anyone who’s not at the table will be on the menu.
Now, none of this works – none of it – without a positive vision.
And this, I think, is where liberals and the left have most badly failed.
Twelve years ago, I wrote a book called Utopia for Realists. I complained that the left mainly knew what it was against. Against austerity, against the establishment, against homophobia, against racism, against billionaires. But it lacked a positive vision of where it wanted to go.
My argument was that we should stop being so timid in our political imagination. I argued for a universal basic income, for the complete eradication of poverty, and I argued for a goal that the brilliant economist John Maynard Keynes laid out almost a century ago, in 1930…the fifteen-hour work week.
Keynes thought it was inevitable. He looked at the trajectory of productivity growth and concluded that by 2030 his grandchildren would be working a quarter as much as he did, because the machines would be doing the rest. The strange thing is, he was right about the productivity, but he was wrong about who would benefit.
The fifteen-hour work week was technically achievable by the 1980s, but it didn’t happen because the productivity gains were captured. By capital, shareholders, and a rentier class. So wages stagnated, hours rose, and inequality exploded.
.jfif)
