In 1865 an English economist warned the country that better steam engines would burn more coal, not less. He was right, and he has been right about every cheap resource since. Intelligence just became one.
William Stanley Jevons published The Coal Question in 1865, and the question was whether Britain would run out of coal. The obvious answer was that James Watt had fixed it: the new engines got several times more work out of a ton than the old Newcomen pumps, so surely the country would burn less. Jevons said no, and said it in a sentence that has aged better than almost anything else written that year.
A furnace that makes iron with less coal makes iron cheaper. Cheaper iron sells more iron, so more furnaces get built, and "the greater number of furnaces will more than make up for the diminished consumption of each." Efficiency did not shrink the coal industry. It built the railways, the steamships and the factories that ate coal by the mountain.
Economists now call this the rebound effect, or when the rebound is bigger than the saving, the Jevons paradox. It is not a law of nature. It happens when demand for the thing was being held back by its price, which is a polite way of saying when people wanted far more of it than they could afford. Keep that condition in mind. It is the whole argument.
The Nobel economist William Nordhaus once worked out what a thousand lumen-hours of light cost, from tallow candles forward. In 1800 it was about $785 in today's money, or five and a half hours of an average worker's labour. By 1992 it was 23 cents, or a little over four ten-thousandths of an hour. Light got roughly three thousand times cheaper and Britain's total consumption of it rose tens of thousands of times. Nobody responded to cheap light by using the 1800 amount and pocketing the difference. They lit the streets, the factories, the night shift and the fridge.
David Shapiro made the bandwidth version of this point last week. His first multiplayer game patch was 8.3 megabytes and took an hour to download. He can now stream a 32 gigabyte film in 4K without thinking about it, roughly four thousand times the payload, and we all use the internet thousands of times more than we did during the bubble that was supposed to kill it. The bubble popped. The demand did not.
And then the tellers, the case about people. ATMs were supposed to end the bank teller. Instead, according to economist James Bessen's count, US teller employment went from about 500,000 in 1980 to about 550,000 in 2010, while the number of machines quadrupled. The machine made a branch cheaper to run, so banks opened more branches, and each branch needed a smaller number of tellers doing more valuable work. Bessen is also honest about the ending: after 2010 the smartphone did what the ATM never did, because it removed the reason to visit a branch at all. The lesson is not that automation never takes jobs. It is that automation takes jobs when it removes the demand, not when it removes the task.
| the cheap thing | what the forecast said | what happened |
|---|---|---|
| Coal, after Watt | Britain burns less | Britain burns mountains; railways, ships, factories |
| Light, 1800 to 1992 | People save on candles | Price down three thousand times, use up tens of thousands of times |
| Bandwidth | The dot-com bubble ends the internet | Four thousand times the payload, thousands of times the use |
| Teller time, after ATMs | Tellers vanish | More branches, more tellers for thirty years, then phones removed the visit |
| Intelligence, now | Jobs vanish | See below. The curve has only just started. |
Andrew Ng, who taught something like eight million people how machine learning works, gave a long interview this month that the AI Corner newsletter boiled down for the rest of us. Three of his points are the x, y and z of this section, and they line up with Jevons exactly.
Economists Erik Brynjolfsson and Andrew McAfee stopped counting job titles and started counting tasks. Ng's summary of where that lands: AI can do 30 to 40 percent of many jobs. Not 100, not zero. The part that matters is what happens to the other 60. When one input to a job gets cheap, the input it depends on gets scarce and valuable. Economists call that a complement. Your judgment, your context, the thing you know about the customer that never made it into a document, just went up in price.
This is the teller mechanism again. The ATM did not reduce banking. It reduced the cost of a branch, and the industry spent the saving on more branches. When a first draft, a research pass or a working prototype costs an afternoon instead of a month, the response is not to do the 2022 amount of drafting and go home. It is to draft ten things, test three, and ship one that would never have been attempted. Ng builds something new most weekends. So do I. Neither of us was doing that in 2022, and none of it shows up in a job-loss headline.
Ng's blunter line: AI is not positioned to replace the person, but people who use AI will likely replace people who don't. That is a reshuffle inside the profession, not a removal of it. His example is software engineering, the field AI has hit hardest, where he says the good engineers he knows are busier than ever and the openings are up. The liability is not the job. The liability is doing the job the way you did it in 2022, which is the part that Jevons would recognise as refusing to buy the cheaper coal.
There is a fourth reason that is less flattering to everybody, me included. Ng says the panic itself has a funding source. Frontier models cost billions to train, open models are catching up for a fraction of that, and a frightened public writes the kind of rules that protect whoever spent the billions first. He calls the nuclear-weapons comparison a mood, not an argument. I would only add that the job-apocalypse story has the same shape: it assumes AI does all of a job or none of it, which is not how any of the last two hundred years of cheap things worked.
Here is where I move from the record to my own opinion, and I'll label it. Shapiro is right that the cost per token of intelligence is falling toward zero the way the cost per bit of bandwidth did. The AI skeptics are right that a company can lose money on every token and not make it up in volume. Both things are true and neither one ends in the collapse that the second group keeps predicting.
The reason is that nobody selling a scarce thing prices it at the cost of the cheap thing underneath it. Bandwidth got free and Netflix still charges you, because what you are paying for is not the bits. The frontier labs are already pricing the frontier, not the token: the model that can do the thing this month that no other model can, plus the agents, the harnesses, the enterprise contracts and the memory that make switching painful. The commodity tier races to zero and they let it, because the commodity tier is where the open models live.
My call, for the record: a few bankruptcies among the second-tier labs, a handful of mergers where a lab with a model but no distribution meets a company with distribution but no model, and the four biggest labs still standing at the end of it. Some of the trillions being spent on data centres will be lost. The internet buildout lost fortunes too, and the fibre it laid is what you are reading this on.
Shapiro's phrase is that humanity is getting less than one percent of the intelligence it actually wants, and I think he is being generous. Consider what the current uses look like: writing emails, summarising meetings, drafting code, and a lot of people asking a chatbot the kind of question they used to ask a search engine. That is the 1800 level of light. It is candles, but cheaper.
The Jevons condition is that demand was being held back by price. Was it? Every small business that never hired an analyst because analysts cost $90,000. Every idea that died because building it needed a team. Every question a patient did not ask because the appointment was twelve minutes long. Every game a fifty-nine-year-old in Peterborough never made because he cannot draw. The unmet demand for intelligence is not a market. It is everything that did not happen because thinking was expensive. Ng's marketing team writes code now. His CFO writes scripts that check documents. His recruiters embed engineers. None of that was a job opening in 2022.
Picture him. He has seen a steam engine. It pumps water out of a mine, and a clever one might soon pull a cart along a rail at a walking pace. Now ask him what steam will do in a hundred years.
He will tell you, confidently, that there will be more pumps and faster carts. He will not tell you about the commuter, because there are no commuters, because nobody can live twenty miles from work. He will not tell you about suburbs, or the weekend, or the factory town, or the time zone, or the seaside holiday, or the standardised clock, or the idea that a fresh strawberry in London could have been picked in Kent that morning. He certainly will not tell you about the rocket, even though the chain that runs from a mine pump to a moon landing is real and every link in it made sense at the time. He cannot see past link two, and it is not because he is stupid. It is because links three through forty do not exist yet and he is a sane person.
That is where we are. The emergent explosion is the part nobody can forecast, and I mean nobody, including the labs. Most of the AI predictions you read, hopeful or doomy, are pumps and faster carts: today's uses, done more. It takes a slightly unhinged imagination to hold twenty links of the chain in your head at once, the kind of imagination that in my case came with a diagnosis and a prescription. I am not claiming it is a superior instrument. I am saying that the 99 percent thinking in terms of today are not wrong about today. They are answering the wrong question, and the honest answer to the right one is that we do not know what the suburbs of intelligence look like, only that there will be suburbs.
Jevons was not a futurist and did not try to be one. He did not predict the railway boom or the steamship. He predicted the direction and got out of the way of the details. That is the only forecast on offer here too: cheaper intelligence gets used more, the total goes up, and the shape of the increase is the thing nobody can see yet.
Ng's advice for anyone who works is three words: learn, build, ask. Mine is the same, with the Jevons framing bolted on.
The man in 1800 could not have planned for suburbs. But the man who bought land along the rail line without knowing what would be built on it did fine. You do not need to see link forty. You need to be standing near the line.