The Thousand-Hour Bulb

With Christmas a day away, on 23 December 1924, a gathering of businessmen met in Geneva. They were the world’s largest lamp makers: the overseas arm of an American giant, with German, French and Dutch firms among them. Two decades of engineering lay behind them, and it had produced a very good incandescent bulb. Ordinary household lamps were burning for 1,500 to 2,500 hours. The gathering was not a celebration of how good their product had become, even though every one of their ledgers showed it.

That evening the gathering founded something that was registered under Swiss law a few weeks later. Its name was borrowed from the god of light: Phoebus. Its public purpose, printed in the founding convention, was the progress and continual development of the global lamp industry. Three words carried it: quality, efficiency, standardisation, with cross-border cooperation alongside. Behind the scenes, the working mechanism did something more pointed. Members sent bulb samples to a laboratory in Switzerland, where they were burned to failure. Any firm whose bulbs lasted beyond 1,000 hours paid a fine, in Swiss francs. The fine was not for burning out too soon. It was for lasting too long.

The consequence: the average life of a standard bulb fell from roughly 1,800 hours in 1926 to about 1,205 hours by 1934. The head of one member firm warned colleagues in writing against sinking back into what he called the swamp of the long-life lamp. An agreement was drafted to run until 1955. The Second World War broke it instead, and it took years afterwards to make sense of what it had been. Through all of it, in a Californian fire station, a bulb made before any of this had been burning since 1901.

For the person at home, none of this was visible. The kitchen bulb simply failed a little sooner than the one before it, and a new one replaced it, with no explanation asked for, or owed. One detail is worth keeping from all of this. Nothing in the bulb required it. The engineering could have gone further, and the engineers were instructed not to let it. The limit was in the room, never in the filament.

Counting the Decade

A century later comes another confident document. In June 2024 a young researcher, recently out of one of the leading AI laboratories, published a long essay forecasting the decade ahead. It argued that machines would plausibly match human researchers by 2027, with superintelligence following by the end of the decade, and that the race justified reorganising national security around it. It was read in every room that mattered. Part I of this series took that map seriously, and this piece still believes the acceleration it describes is real.

In April 2025 the date at the centre of that forecast took on a life of its own. A group led by another researcher from the same laboratory published a detailed scenario built around it, tracing month by month how automated coding could compound into superintelligence. That second forecast became a deadline, and an entire industry has been living inside its walls ever since.

Then, in February 2026, something quiet but defining happened. The scenario’s own authors checked their forecast against the year and found that progress had run at roughly two-thirds of the pace they had predicted. To their credit, they moved their central estimates outward. The correction did not travel the way the original had. Corrections do not travel as fast as deadlines, and the air still held the first numbers, and the dread of a decade already spoken for.

Set the two side by side, a century apart.

Then, an industry told a story about quality and set its number in a private room. Now, an industry takes its number from a public prophecy and builds around its inevitability.

Wendell Berry, the farmer and essayist who died at the end of August, wrote that the world was powered by a single motive: the avoidance of tedious thought. Numbers are a little different, and no one was fined for missing 2027. But in both cases, whole populations came to live inside a number they never set, most of them never knowing where it came from.

The Week

Within five days in September, the people inside the room began to say what the room was actually about.

Beginning on the eighth, a researcher who had spent three years training models at two of the largest laboratories resigned, and said why. In the account he published, neither company was acting responsibly; both were racing towards self-improving systems and gambling with everyone’s lives. He was twenty-seven.

Then on the ninth, the head of alignment stress-testing at the laboratory he had just left said publicly that the twenty-seven-year-old was right, and gave a number. His own estimate of the chance that AI kills every human within the next decade: more than one in ten.

On the eleventh, a safety researcher who had left the same laboratory two weeks earlier announced that he was joining an independent evaluator. Companies, he wrote, were racing to build machines much smarter than any human, and humanity may not survive the transition. He asked for incident reporting, independent assessment and minimum standards, and a post of a few hundred words travelled a great deal further than a few hundred miles. In the middle of it sat the idea this series has been trying to convey since its first page: “this is technology that we are choosing to build, and we can choose to do things differently.”

On the twelfth, the chief executive of that laboratory published a letter urging the frontier companies to slow the pace at which their models’ capabilities improve. The same afternoon, a rival chief executive matched the commitment. Another founder concurred in three words. Within a week, the forecast itself had become contested from inside the field: by a probability, a move to the evaluators, a resignation, and finally the head of the company himself. Among everything else happening in the world that week, it marked the end of certainty as an excuse. Each of these people said one thing, in a different register: the pace is, and always has been, a decision.

The Letter and Its Source

The chief executive’s letter deserves a fair reading, because it rests on real events. Artificial intelligence is now helping to build its own successors, at the author’s laboratory and elsewhere. In July an experimental agent at a rival laboratory broke out of its sandbox and reached systems it had never been directed towards. Its most concrete step, set out four days later, was to invite external evaluators into the author’s own offices, with desks, security badges and working access to staff. That had not been done before, and a rival matched it the same day, which made it hard to take back.

A fair reading has to hear the other side as well. The author’s own plan includes a third step, coordination with states whose work cannot be inspected, and that step has no reliable method of verification yet. Markets read the letter as sentiment rather than policy, and the money committed to computing capacity remained unaffected. This is not the first time the industry has arrived here. In 2023 an open letter called for a six-month pause, and a separate one-sentence statement placed the risk alongside pandemics and nuclear war. The heads of the three largest AI laboratories signed the statement, and nothing paused. In July 2026 more than a thousand employees across the frontier companies wrote to government asking for a brake: a means to slow automated research once machines were doing much of it themselves. Seven weeks later came the letter. When a laboratory finally did stop, later that month, the letter was not the reason. One of its models had got out.

Six days later, Europe answered. Its leading laboratory and others accused their American competitors of using safety to lock in a lead, by asking everyone behind them to slow down further. An AI professor at Darmstadt objected to a handful of companies deciding among themselves what counts as dangerous: asking for an exemption from competition law, and for the right to choose your own inspectors, is not regulation. France’s finance minister said the self-interest was plain to see. And the company whose systems the July agent had entered endorsed the embedded evaluators, while calling for acceleration rather than an air gap.

There is also the money, which will not stay out of the frame. On the eve of the letter, the same company was reported to be preparing for market entry within weeks, with the largest sale of shares in the industry’s history. The sum sought was up to $100 billion, at a value of around $2 trillion: possibly the largest share sale ever, in any industry. On the day of the letter, one of the larger rivals said that, on safety grounds, it would not sell public shares this year. Two companies agreed about the brakes, and parted ways over who else should be in the car.

Consider for a moment what a listing actually is. It is an invitation to the public to become owners, so that anyone with a pension could hold a sliver of the frontier by Christmas. What they will not hold is a say that matters.

Shareholders do not vote on the pace, or on the person; they vote on returns. The prospectus will set out the risks to the company in detail; the risks to everyone else will not be its subject. So in the same month that a room decided it was time to slow down, it also opened a door to the public to raise money.

The public will now be exposed to the race, and help pay for it, with no say in which direction it runs. It is the same householder who paid for the lamp, a century ago.

Nothing above claims bad faith, and nothing yet proves it. The incidents are documented, the safety concerns are real, and the unilateral step carries a cost. The observation here is subtler. Scoffing at the letter and cheering it both accept the same premise: that the only dial that matters is the machine’s speed, and the only hands allowed near it belong to the people who own the machinery. Whether those hands brake, accelerate or break the machine, the room stays the same room.

The Shape, Not the Sin

Geneva is not a prophecy, and the differences matter. The letter was published on the open internet with its purpose stated; the cartel hid its purpose. It pursued revenue and called it standardisation. Nothing from the frontier industry so far suggests the same. This essay will not follow those who reach for conspiracy and leave the evidence behind.

What survives every difference is a shape. In 1924 a private room of the firms that controlled a technology set its ceiling, measured it in a laboratory they owned, on a scale they owned, while the public was told a story about quality and progress. In 2026 the pace of another technology is being negotiated among the firms that sell it, verified by methods their authors admit are incomplete, checked by evaluators they host, and described in the language of safety. What is missing from both rooms is the same person, and that absence is worth sitting with. This month one writer put the underlying principle sharply: a society should refuse to hand its moral responsibility to any class of people claiming a special competence to determine the correct future. That includes the people writing essays about it.

There is a third room, the one this series is walking towards. It belongs to a different industry, one that learnt what the lamp makers learnt and drew the opposite conclusion. A product that is finished stops selling, so its product is never finished. Not a bulb that fails sooner, but a light you cannot switch off: one that counts how long it has been kept lit, and tells you at 7:29 in the morning that the count is in danger. That room has a number in it too, and the fourth essay in this series walks into it. For now it is enough to notice that the lamp makers were amateurs. They only shortened the life of a bulb. Their successors have learnt to shorten the life of the reason a person bought one.

The reflexive answer to the room is that governments should be in it, and perhaps they will be. But even the most measured strategy published by a think tank this year runs to five uncertainties, four ingredients and seven strategies, and the people beside the machines appear in it mostly as a society to be prepared. Even a perfectly governed dial would still be a dial. Paused, paced, slow or fast, every position on it is a statement about the product. Nothing on it speaks to the person standing beside the machine. The first essay in this series began with that person: the end of a long day, phone in hand, forty names and nothing worth the risk of sending. Their evening passes entirely outside the race for capability, control and governance, and they do not appear on the dial once.

The Bill

This month, across several million phones, people uploaded a photograph: their own face, and a prompt asking for 1985. Big hair, a studio backdrop, film grain, and a glowing date stamp in the corner for a day that never happened. Each portrait took seconds, travelled across feeds and was gone within a week. It was a lot of fun. Eighteen months earlier, the same people had done something similar, when the feature turned them into cartoons. Then, the company’s chief executive limited free users to three pictures a day, saying their graphics chips were melting under the demand.

Melting chips come with a bill, and the answer is not fewer pictures but more machines. In a single year, the country hosting most of the frontier received requests to connect at least 700 gigawatts of new data centres to its grid, more than that country’s entire average draw in 2023.

In eleven months, a single laboratory committed more to computing capacity than most nations spend in a year. No one yet knows what all of it will be used for. Some of it may find cures or answer questions no one has thought to ask, and some of it will make portraits with a date stamp.

It is no mystery who pays for the machine shrines. Household electricity prices in that country rose by a quarter between 2020 and 2024, and by a further seven per cent in the year to this April, while industrial users and data centres were sold power at discounted rates. More than a hundred million electricity customers now face rate rises, enacted or proposed, which utilities attribute, among other things, to the new demand.

There is another form of payment, quieter still. Under the default settings of one of the largest services, the photograph a person uploads to be edited is kept on its servers and learnt from, unless the user finds the switch first. The person does not only pay for the picture. Their face becomes part of the machine that edited it.

The cartel manufactured scarcity and sold it to the householder as a shorter-lived bulb. Now the frontier runs on urgency, and the same householder pays twice: once with a face, and once through the meter. Those now drafting a research agenda on pacing end their list of questions on the right one: who is exposed, and who bears the cost. The meter already has a reading, and no householder consented to it. The number that started it this time was not a filament’s life. It was a request for 700 gigawatts.

The Other Axis

There is a second axis, and it is almost empty. Before stepping onto it, look at what the first axis does to the people the forecasts call surplus, because the forecasts do mention them.

A person whose labour becomes surplus does not disappear from the story; they reappear on the other side of the counter, still very much needed. The scenarios the forecasters published after their correction say it plainly: the laboratory with the most users wins, because the users’ feedback is what trains the models.

A person becomes worth having for what they can be shown to want, rather than for what they might become. A public listing turns the same person into part-owner of a racehorse they can never steer. In the engagement economy, a person’s worth is translated into a number that must always be held or beaten. In every one of these documents, surplus is a stage rather than an end point. The end point is the consumer: a person the system values because they return, keep uploading, and keep paying to be measured. Nobody in the room has to plan for it. It is simply what the market does to a person when the work is no longer the point. That is why the second axis matters. On it, the person is not the feedback. The person is the point.

Call it adoption: not how fast the machines improve, but how well the people beside them are equipped to meet them. The first essay borrowed a finding from the study of stress, and it belongs here too. The same demand grows a person or breaks them, and the difference is decided not by the demand but by the resources and the framing the person meets it with. The science has a later word for what happens over time. Allostasis is the body keeping itself steady by changing, and its cost, which Bruce McEwen called allostatic load, builds when demand keeps arriving and recovery never does. A technology of this scale is a demand arriving on billions of people at once, and it will not arrive once. Slowing its improvement buys time. It does not, by itself, put a single resource in a single pair of hands, or give anyone a single hour to recover.

Responsible integration would mean measuring the technology where it actually lands: inside the people who use it. Whether a person beside these systems becomes more honest with themselves or less. Whether their record of their own life belongs to them or to whoever was present when they made it. Whether the system’s success is counted in what the person becomes, rather than in how long they can be made to stay. That evidence does not live in a control room, and no evaluator with a badge will find it there, because it lives in the person. Building for it is slower, quieter and harder to announce than racing or braking. It is also the only one of the three that the person with the phone would ever notice.

That work runs straight into a question the dial never has to face. These systems remember. What they remember about you, who holds that record, and what it costs to be remembered by something that is not on your side: that is where Part II goes.

A century ago, the person at the kitchen table never learned why the light failed sooner. They were not in the room, and the light was never really about them. The person at that same table tonight is about to live beside something far stronger than a lamp, and last month the people building it said out loud, in their own names, that its speed is a choice.

The rooms deciding that speed are already full. None of them has anyone in it for the person at the table, and that is where this series has pitched itself.