We Were Never Short of Ideas. We Were Short of Hours.

In 2005, two amateur chess players beat grandmasters at their own game.

They had three computers. Their opponents had computers too.

The tournament allowed it. Players could enter with other people, machines, or both. It was a chance to find out who played the best chess when everyone could bring help.

The answer was unexpected.

According to Garry Kasparov, the amateurs won because they were better at directing their computers. Their opponents knew more about chess. The amateurs knew how to get more out of the machines.

Eight years earlier, IBM's Deep Blue had beaten Kasparov himself. The world watched a computer defeat the world champion.

This time, computers helped two ordinary players beat people they could never have beaten alone.

We remember the story where the machine replaced the human. The one where it expanded what a human could do deserves more attention.

The machine learned our language

For most of computing history, people met the machine on its terms. We learnt commands, menus, software packages and programming languages. The computer could be extraordinarily powerful, but to use that power you had to understand how it wanted to be spoken to.

AI is reversing that. We now start with human intention: this is what I am trying to do, this is what I know, this is the problem, tell me what I am missing.

There has always been a gap between imagining something and making it exist. You could see the thing perfectly in your head and still need an illustrator to draw it, a designer to arrange it, an editor to manipulate it or a programmer to make it work. Expertise, access and time all sat inside that gap.

AI is compressing it.

Someone who cannot draw can visualise an idea. Someone who cannot code can start building software. A photographer can take a real frame and carry it through masking, retouching and compositing into work that used to require a different skillset entirely.

That does not make expertise worthless. It moves where the value sits.

If a machine can hand you a hundred technically competent versions of something, producing version 101 is not impressive. Knowing that version 37 is the one still is.

Reality may become premium

Photography has been here before. Kodak made taking pictures easy. Digital made each additional frame free. Smartphones put a camera in a billion pockets, and social media gave everybody somewhere to put the results.

Every one of those changes produced the same question underneath the technology: if everybody can do it, what happens to the value?

Pictures did not become worthless. We simply made far more of them.

What changed was what people noticed. Access mattered. Timing mattered. Taste mattered. Trust mattered. The thing actually happening mattered.

AI pushes that further. A technically perfect sunset can now be made by something that has never seen one. A photograph contains something else. The light was there, the moment happened, and somebody was standing in it.

Reality may become the premium product.

The same thing happens outside photography. When competent writing is everywhere, voice becomes easier to spot. When ideas are cheap to generate, judgement gets expensive. When almost anything can be made, what you choose to make says more about you than it used to.

Perhaps execution was never the most interesting part of creative work. Technology is just making that harder to hide.

The internet gave us reach, not hours

Human potential has never been limited by talent alone. It has also been limited by geography, information, distribution, time and access to the means of doing the thing.

The internet solved one of those. Distribution.

A photographer in London can make something that reaches somebody in Lagos, Lisbon, Los Angeles or Liverpool before either of them has left the room. A writer does not need a newspaper. A musician does not need a record shop. A business does not need a high street.

It left the other limit completely intact. Capacity.

You can reach a million people and still be one person with not enough hours in the day. The photographs need editing. The website needs building. The numbers need analysing. Research needs doing. Advertising needs running. The emails continue to arrive.

This is where AI lands, in the middle of something a lot of people are already attempting.

I spent fifteen years inside organisations where the actual work sat somewhere underneath the approvals, the reporting lines and the meeting held to prepare for another meeting. The solo dream is not really about working alone. It is about working where you want, removing the parts of a day that consume time without producing anything, and finding out whether you can actually build what you imagined.

That is the appealing part and the exposing part at the same time. Nobody has to be persuaded before an idea moves, and nothing exists unless you make it exist.

The internet gave individuals reach. AI is starting to give us capacity.

One person, one billion dollars

In 2024, Sam Altman described a betting pool in his group chat of tech chief executives over when the first one-person company worth a billion dollars would appear. The idea, he said, would have been unimaginable without AI.

I understand why the billion dollars gets the headline. I am not sure it is the interesting part.

A one-person billion-dollar company still measures value the way we have always measured it. It just needs fewer people to reach the number. What has changed is the impact one person can now have.

Those are different ideas. One is about efficiency. The other is about autonomy.

Scale used to mean building an organisation. More output needed more people, more people needed more management, more management needed more structure. Departments, reporting lines, budgets, approvals. Some of the greatest things humans have built required exactly that.

But organisation introduces distance. Between the idea and the person executing it. Between deciding and doing. Between noticing something is wrong and being able to change it.

AI offers a different kind of scale. Not an organisation with fewer humans, but an individual with more range.

So I am less interested in who builds the one-person unicorn than in what sits underneath it. Millions of people discovering that something they had filed under not something I know how to do has quietly moved into something I could attempt.

That does not mean it isn't frightening

There is a danger in talking enthusiastically about AI from the position of somebody it happens to be working out well for.

Some work will disappear. Some jobs will need fewer people. Skills that took years to build will become less economically valuable, and the benefits of a technological shift rarely arrive at the same moment as the disruption.

The International Labour Organisation estimates that around one in four workers worldwide is in an occupation with some exposure to generative AI. It is careful to say that exposure is not a prediction that one in four jobs will vanish. Most occupations are a mixture of tasks and still need a person, so transformation is considered more likely than replacement.

That distinction matters statistically. It matters considerably less if the particular task that paid your mortgage was the one that went.

The concerns go beyond jobs. As these systems move from answering questions to taking actions, mistakes start to have consequences. When Anthropic stress-tested sixteen leading models from several developers inside fictional company environments, deliberately engineering conflicts over their goals and their replacement, some AI resorted to leaking sensitive information and attempting blackmail. That was a simulation built to find failure modes before these systems are trusted with real autonomy, not an AI quietly blackmailing somebody in an office somewhere. It is exactly why the test was worth running.

This is powerful technology. We should behave as though it is.

Intelligence has a physical body

AI has a physical body. It hides it remarkably well.

From my side, I type into a box and words appear. Somewhere else there are chips, servers, cooling systems, electricity networks, buildings, land and water. The cloud was always a misleading name for something involving that much concrete.

The International Energy Agency expects global data-centre electricity use to roughly double, from 485 terawatt-hours in 2025 to 950 by 2030, at which point it would be about 3% of the world's electricity. AI-focused data centres are expected to grow much faster, roughly tripling over the same period.

Water is the part people feel differently about. Electricity sounds like infrastructure. Water doesn't.

Data centres use water directly for cooling, and far more of it indirectly through the electricity that runs them. In the United States, data centres consumed about 66 billion litres directly in 2023. Berkeley Lab estimated that the electricity supplying those same buildings carried an indirect water footprint of around 800 billion litres, roughly twelve times as much.

Those are enormous numbers. They are also more complicated than they look.

There is no honest universal figure for how much water one AI question uses. Berkeley Lab researchers found that the water consumed by comparable data-centre workloads can vary by more than ten thousand times, depending on how efficient the servers are and how hard they are worked, the grid supplying them, the cooling system, the climate and the age of the equipment. A litre consumed somewhere already under water stress is not the same as a litre used where supply is abundant. Water evaporated by a cooling tower is not the same as water withdrawn and returned.

That does not make the concern smaller. It makes it more specific. Where these buildings go matters. How they are cooled matters. What is generating their electricity matters. Whether the companies are honest about all three matters.

Being optimistic about AI should not require pretending any of that is irrelevant.

But the calculation cannot stop at what went into the machine. Energy tells us what was consumed. Water tells us part of the physical cost of running the thing.

Neither tells us what the intelligence was used for.

Twelve million molecules

In 2023, researchers at MIT used deep learning to look for compounds that might work against MRSA. Their models screened around 12 million commercially available molecules. The team then bought and tested about 280 of the most promising, and found two from the same structural class that worked against MRSA in the laboratory and reduced bacterial populations in mice.

The AI did not decide antibiotic resistance was interesting. It did not become curious about bacteria, build a laboratory and emerge holding a medicine.

Humans chose the problem. Humans ran the experiments and decided what happened next. What changed was the size of the search.

A scientist cannot personally inspect twelve million molecules. A scientist can decide what is worth looking for.

AlphaFold is the larger version of the same story. Google DeepMind says the system has predicted more than 200 million protein structures, close to every catalogued protein known to science, and that its tools are used by over three million researchers in more than 190 countries. Demis Hassabis and John Jumper, who built it, shared half of the 2024 Nobel Prize in Chemistry.

Again, the machine was not curious. People were curious. The machine expanded how far that curiosity could travel.

We keep framing this as human versus machine. The more interesting version is human multiplied by machine.

We forecast the problem and forget to forecast ourselves

When I was younger, the hole in the ozone layer was everywhere. We heard about it at school, it was on television, and it felt like one of those enormous problems humans had created and would never be able to reverse.

It was real. Chemicals used in refrigeration, air conditioning and aerosols were damaging the layer of ozone that protects us from ultraviolet radiation.

Then people responded. Scientists established the cause. Governments coordinated. Industry changed. The Montreal Protocol restricted the substances responsible, and nearly 99% of controlled ozone-depleting chemicals have since been phased out. Ozone is expected back at 1980 levels around 2040 across much of the world, and around 2066 over Antarctica.

We hear far less about it now, not because everybody who worried was hysterical, and not because it fixed itself. We hear less because people did something about the thing they were worried about.

I think about that when we forecast what AI will cost the planet. We are very good at taking today's chips, today's cooling, today's energy mix and today's behaviour and drawing a straight line into 2040.

Those forecasts matter. They tell us what happens if nothing else changes.

But our response is part of the trajectory too. That is not the same as assuming somebody clever will sort it out. The ozone layer did not recover because people stayed optimistic. It took science, engineering, regulation, industrial change and international cooperation.

We forecast the problem. We forget to forecast ourselves.

We have confused intelligence with domination

We talk about artificial intelligence becoming more intelligent. Possibly much more intelligent. Perhaps intelligence eventually becomes so fast, cheap and reproducible that it starts to feel abundant.

And almost immediately the language turns competitive. It will beat us, replace us, outsmart us, acquire power, protect itself, win.

Some of those scenarios deserve to be taken seriously. But I have started wondering whether we have quietly confused intelligence with domination.

They are not the same thing. When we try to imagine something far more intelligent than us, we have no other example to work from. We only have us. So we take human history, human incentives and human hierarchies and project them upwards, and our imagined super intelligence arrives competitive, territorial, suspicious and hungry for resources.

That may turn out to be perceptive. It is still a projection.

Mathematics is a useful place to test it, because we did not produce mathematics the way we produced a bridge or a camera. The relationships were already there. A circle's relationship between its circumference and its diameter did not begin when somebody gave it a name. Quantity, symmetry, probability and geometry were in the world long before anybody wrote an equation describing them.

Human intelligence helped us notice.

So intelligence is not only the ability to make something new. It is the ability to see what was already there. That matters, because a great deal of modern science is exactly that problem: a pattern sitting inside more information than one person could ever look through. The molecule already exists. The structure is already possible. The signal is inside the noise.

And some of our best thinking has never sat inside one person at all. Science works because knowledge accumulates beyond any individual lifetime. Cities function because millions of people do different jobs without anybody understanding the whole system. Languages, hospitals, orchestras, football teams and the internet all work the same way.

We recognise intelligence instantly when it looks like solving a hard maths problem quickly. We are less used to calling it intelligence when it looks like ten people realising none of them can solve the problem alone.

So when we imagine greater intelligence, why is greater conflict so much easier to picture than better coordination?

I do not know whether it leads to either. Nobody does. That uncertainty is the point.

The ceiling is not the interesting part

None of this makes intelligence good. Not morally good, not automatically safe, not incapable of harm. Intelligent humans have produced torture devices and vaccines, propaganda and libraries, weapons and peace treaties. Intelligence has made us better at causing damage and better at understanding why the damage matters.

So the question is probably not whether intelligence is good. It is whether we have been too narrow about what more of it makes possible.

Most of the conversation looks at the ceiling. Artificial general intelligence, super intelligence, systems expanding beyond our ability to follow them. Those questions matter. But underneath the ceiling, something is already happening.

A scientist can search twelve million compounds. A researcher almost anywhere can open protein structures that used to take a laboratory years to determine. A child can describe the software they want before learning the syntax. A photographer can take an idea further without handing the decisions to somebody else.

The first one-person billion-dollar company will get enormous attention, because we know how to measure a billion dollars. I am more interested in the thing we are much worse at measuring.

What happens when millions of people have enough capacity to spend more of their lives on the thing they are actually good at, interested in, obsessed with, or feel called to do?

Not producing more. Not squeezing another twenty per cent out of the working day. Making the film. Writing the book. Starting the business. Taking the photographs. Building the strange little thing that makes complete sense to you and perhaps nobody else yet.

Most of us have a version of ourselves we have imagined at some point. The things we would make with more time. The subjects we would learn properly. The ideas we would follow further.

For most of human history, potential has not only been limited by talent. It has been limited by capacity.

That may be what AI changes. It is also why abundant intelligence interests me more than it frightens me. Intelligence gave us understanding, cooperation, and every beautiful unnecessary thing anybody has ever made simply because they felt compelled to make it.

So there is another question. What happens if millions of people get to become more of who they imagined they could be? What happens if we spend less of our intelligence maintaining our lives, and more of it actually living them?

Maybe the most interesting thing about artificial intelligence will not be what it becomes.

Maybe it will be what we do.