AI and the Industrial Revolution

Cheap Tokens, Cheap Attempts

Cheap generation doesn't make engineering judgment obsolete. It makes it the scarce thing, and the first Industrial Revolution shows why.

Sertan Yamaner

— 10 min read
Share
Cheap Tokens, Cheap Attempts

Over two evenings, I had an AI agent rebuild something I'd been running for months: a small system that listens to my coffee roaster and detects the moment the beans hit first crack. It wrote about eleven thousand lines of working code across seventy-five files. Then the thing wouldn't train. The fault was a single wrong word in the data loader: the model expected its input under one key, and the agent had supplied a different one, a plausible guess copied from a different audio pipeline. I found it by reading the code. The agent didn't, and neither did the automated reviewer that left a hundred and ten comments across that work.

That ratio is the argument. Generating the code was nearly free. Knowing which line was wrong was the job.

So the question everyone is asking "are the agents smart enough to replace engineers?" is the wrong question. The spinning jenny was never smarter than the weaver. It made cloth cheap to attempt, and that reorganised everything downstream. Agents are the same kind of move. They don't out-think your engineers; they collapse the cost of an attempt. And when attempts get cheap, the scarce thing stops being the work and becomes the judgment to know which work to trust.

I've read parts of the relevant history. I haven't written it, and the people who have are still arguing about it, which turns out to matter. So take what follows as an engineer borrowing a pattern, not a historian handing down a verdict.

The obvious modern precedents are the personal computer and the internet, and they fit. I'm reaching past them to the first Industrial Revolution because it is the best-documented case of the thing I care about: what happens to work when the cost of an input collapses.

The clean version

Here is the story as its best-known teller tells it. From the late seventeenth century to the late eighteenth, Britain was an unusual place to do business: labour was expensive and energy was cheap. Wages in the towns were high by the standards of the day, and coal sat near the surface at the pithead. In that specific cost structure, a machine that swapped expensive human hands for cheap coal-fired power paid for itself. In France, where labour was cheaper and fuel dearer, the very same machine did not pay. On the economic historian Robert Allen's account, Britain mechanised first not because Britons were cleverer but because the prices made mechanising worth it. Change what an input costs, and you change what is worth building.

It is a clean, persuasive story. Hold onto it for a moment, because the next part complicates it.

The honest turn

I am an engineer, not an economic historian, so I went to check whether the story holds. It is contested, and the fight is still live.

The mechanisation of spinning is supposed to be the showcase for the high-wage story. But Jane Humphries and Benjamin Schneider, working through the archival record, find that hand spinning was a widespread, low-wage, low-productivity job done largely by women and children, with earnings well below what the high-wage explanation needs. Their conclusion is the reverse of the story: mechanisation was a response to low wages, not high ones. Separately, Judy Stephenson has argued that the day-rates Allen leaned on weren't the wages workers actually took home: they bundled in contractors' margins, so England may not have been the high-wage outlier the theory requires. And Joel Mokyr offers a different explanation altogether: what set Britain apart was a culture of useful knowledge, tinkering, and applied science, not its wage structure.

To be fair to Allen, he has answered his critics, conceded the odd point, and held the line on the mechanism. The argument continues.

Now notice what the argument is actually about. It is about which input mattered in Britain, and even whether high wages drove mechanisation or mechanisation later raised wages. It is not about whether changing the economics of an input reorganises how things get made. Every side takes that part for granted. So that is the part I borrow: the principle, not the verdict.

And the fact that it is still unsettled is itself the point. A century of scholarship, with the parish records in hand, has not produced agreement on what drove the first industrial revolution. That is the warning label for anyone today who is certain what AI will do to this one.

What actually gets scarce

When coal got cheap, Britain did not burn the same amount of it more efficiently. It burned far more. Cheapness raises total consumption; it does not lower it. (The economist William Stanley Jevons noticed exactly this about coal in 1865, and the pattern still holds his name.)

Cheap attempts behave the same way. You do not get the same number of attempts, done more carefully. You get a flood of them. And once the expensive thing becomes nearly free, value moves to whatever decides which of the cheap outputs is worth keeping. The work didn't disappear. It moved, from writing the code to judging which code to trust.

My own build is the proof, and it cuts both ways. The eleven thousand lines and the forty-six agent-authored commits were the cheap part; attempts had collapsed to almost nothing. The value lived in the decisions the agent could not make for me: choosing an Audio Spectrogram Transformer over a convolutional network, and splitting the training data by recording rather than by clip, so the model couldn't quietly cheat by hearing the same roast in both training and test. And the decisive moment was the one wrong line, a plausible-but-incorrect input key that the automated reviewer, across twenty-six passes and a hundred and ten comments, never flagged. It was checking whether the code was clean, not whether the science was right. The result was clean code and a broken model.

There is the obvious counter: maybe the tools get good at that judgment too. They might. But that is a forecast, and I am making a present-tense claim with evidence behind it. Right now, in everything I have built, the scarce and decisive work is telling correct from merely plausible.

The artisans were not stupid

The Luddite comparison is having a moment. Brian Merchant's Blood in the Machine put it back in circulation, and it is usually made lazily, as if the lesson were "don't be a coward about new technology." The real history is sharper than that.

The Luddites were skilled textile artisans, many of them machine experts who had used machinery happily for years. They didn't break frames because they feared technology. They broke the frames that were being used to de-skill their trade, cut their wages, replace them with cheaper and often child labour, and flood the market with shoddier goods. They were organised, often gave warning before they acted, and they read correctly what the machinery would do to them. They were also crushed: the state put regiments in the field and hanged men. Being right about the tool and being right about the strategy are two different claims. The Luddites were right about the first and lost the second, and some of their trades did not survive it.

The engineer who tells you today's agents write subtly broken code is usually right, in the same way the Luddites were right. That is not a reason to dismiss them, and refusal is still the losing move. The better move is to relocate your skill to where cheap attempts can't reach it: judgment, evaluation, architecture, the science. In my own build that meant a strict split between the human who owned the decisions and the agent that owned the execution, which is its own subject for the next post.

And here is the part for anyone leading engineers through this. That skeptical engineer's distrust is not the obstacle to adoption. It is the thing adoption runs on. The distrust is precisely the judgment that just became scarce and valuable. Treating skeptics as fearful and managing them out of the way is the 1811 factory owner's mistake repeated. The move is to bring their distrust inside the system, as the evaluation function.

The honest dark side

I argue for adopting these tools, and I am going to name the cost anyway.

The first industrial revolution made Britain far richer and, for about two generations, barely made its workers richer at all. Allen called this Engels' pause. Between 1780 and 1840, output per worker rose by roughly 46 percent while real wages rose by about 12; the profit rate doubled, and profits took a growing share of national income. The gains were real, and for roughly forty years they accrued almost entirely to capital. Wages only caught up after 1840, rising about 123 percent over the following decades and finally overtaking output, and they did so because capital was reinvested, institutions adapted, and new kinds of work emerged. None of that was guaranteed in advance, and a generation lived through the gap. The size of the pause is itself argued over; the direction of it is not.

There is a sharper, smaller version of the cost. The framework knitters from the last section often rented the very frames that were undercutting them, a charge known as "frame rent," and were sometimes paid in goods instead of cash. The instrument of their displacement was also a line item they paid for.

The modern rhyme is worth noticing. Jensen Huang has floated giving Nvidia engineers an annual token budget worth about half their salary, and says he would be alarmed if a high-paid engineer didn't consume tokens on the order of half that pay. He frames it as a benefit, granted on top of salary, and at Nvidia it may well be exactly that.

The other side of it showed up almost as quickly. In June 2026 Simon Willison flagged a Bloomberg report that Uber had capped its engineers at $1,500 a month in token spending per AI coding tool, to keep its own costs under control. The cap is not ungenerous. The budgets do not pool, so an engineer running two tools has around $3,000 a month, which on Willison's reckoning lands near 11% of a median Uber engineer's package, above what he himself spends. This is cost control, not austerity. But the posture is the part to notice. One firm hands the tokens out as a perk; another, just as well resourced, has already decided they are a line it needs to budget.

The open question is what the same arrangement becomes at a firm that is neither Nvidia nor Uber: whether the seat licence and the token budget are a benefit handed to the engineer, or a cost quietly moved onto them as the price of staying employable. That is a choice firms are making right now, not something that has simply happened to them. Where it lands is a question for later in this series.

What this leaves you with

The eleven thousand lines were the cheap part. Finding the one line that mattered was the job, and it stays the job right up until it doesn't, which is a sentence I can't honestly finish, because nobody can.

We cannot cleanly explain a transition we have had two centuries and full archives to study. Anyone certain what this one will do is overreaching. The smaller, honest claim is the one worth making: the technology does not decide the outcome. What decides it is what you build around the agents: the ground you make them run on, and the organisation that has to absorb what they produce. That is where this goes next.

If you are going to lead people down a path, you owe them an honest description of it, ditches included. This was the path-clearing. The optimism isn't that the tools are good. It is that the outcome is still ours to choose.

Sources

The high-wage thesis and Engels' pause

  • Robert C. Allen, "Engels' Pause: Technical Change, Capital Accumulation, and Inequality in the British Industrial Revolution," Explorations in Economic History 46 (2009), 418–435. DOI: 10.1016/j.eeh.2009.04.004. Author's copy: https://www.nuff.ox.ac.uk/Users/Allen/engelspause.pdf. Source for the +46% output and +12% wages figures (citing Crafts–Harley and Feinstein), the doubled profit rate, and the post-1840 recovery.
  • Robert C. Allen, The British Industrial Revolution in Global Perspective (Cambridge University Press, 2009). The high-wage and cheap-energy explanation for why Britain mechanised first.
  • Robert C. Allen, "The High Wage Economy and the Industrial Revolution: A Restatement," Economic History Review 68 (2015), 1–22. DOI: 10.1111/ehr.12079 (paywalled). Free working-paper version (Nuffield College, Oxford, Economic and Social History Series 115): https://www.nuff.ox.ac.uk/economics/history/Paper115/allen115.pdf. Allen's reply to his critics.

The contestation

  • Jane Humphries & Benjamin Schneider, "Spinning the Industrial Revolution," Economic History Review 72 (2019), 126–155. DOI: 10.1111/ehr.12693 (paywalled). Open copy (accepted manuscript) at the Oxford University Research Archive: https://ora.ox.ac.uk/objects/uuid:4f321519-a0b7-4879-a521-139cbca81b25. Spinning as widespread low-wage, low-productivity work; mechanisation as a response to low, not high, wages.
  • Robert C. Allen, "The High Wage Economy and the Industrial Revolution: A Restatement," Economic History Review 68 (2015), 1–22. DOI: 10.1111/ehr.12079 (paywalled). Free working-paper version (Nuffield College, Oxford, Economic and Social History Series 115): https://www.nuff.ox.ac.uk/economics/history/Paper115/allen115.pdf. Allen's reply to his critics.
  • Judy Z. Stephenson, "'Real' wages? Contractors, workers, and pay in London building trades, 1650–1800," Economic History Review 71 (2018), 106–132. DOI: 10.1111/ehr.12491 (paywalled). Free working-paper version, published earlier as "Real Contracts and Mistaken Wages": https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2723977. Building-trade day-rates included contractors' margins of roughly 20 to 30 percent, so recorded wages overstate what workers were actually paid.
  • Joel Mokyr, The Enlightened Economy: An Economic History of Britain 1700–1850 (Yale University Press, 2009). Useful knowledge and applied science, rather than wage structure, as Britain's distinguishing factor.

Magnitude of the pause (contested)

The Luddites

  • E. P. Thompson, The Making of the English Working Class (Victor Gollancz, 1963). The skilled-artisan account.
  • Brian Merchant, Blood in the Machine: The Origins of the Rebellion Against Big Tech (Little, Brown, 2023). Author's précis: https://time.com/6317437/luddites-ai-blood-in-the-machine-merchant/. Luddites as machine experts opposing the terms of mechanisation, not technology itself.

The cheapness mechanism

  • W. S. Jevons, The Coal Question (Macmillan, 1865). Cheaper, more efficient resource use raises total consumption. Public-domain full text at the Internet Archive and Project Gutenberg.

The modern signal

The build