Could AI be bigger than the Industrial Revolution? Not by making work faster. In Joel Mokyr's history the revolution's lasting gift was a way of making discovery compound, and that is the test AI has to pass. The one market that prices the whole economy's future has moved by about as much as a yes would need, mostly, we think, for other reasons.
AI can plausibly add more to productivity growth than any single industrial technology did. It has not been shown to be bigger than the Industrial Revolution. The OECD's scenarios have AI adding 0.41 to 1.28 percentage points a year to American labor productivity growth over a decade, and even the slowest beats steam at its best, 0.31 points a year from 1870 to 1910. But the Industrial Revolution was not a technology. In Mokyr's account it was the moment improvement became self-sustaining, and the test for AI is the same: whether it raises the rate at which useful knowledge grows, or only the level of what we produce.
The bond market has taken a position. The real yield on 30-year Treasury inflation-protected securities was 3.31% on October 1, 2026, the highest since the series began in 2010 and 1.85 points above its level on the day ChatGPT was released. Faster expected growth raises long real rates. If all of that rise were growth, it would price 0.93 to 1.85 points a year of extra growth for thirty years: not a decade of diffusion but a change in the rate, the strongest form of the thesis. It is probably not growth: the New York Fed's term structure model puts the whole rise in the 10-year Treasury yield since then down to a higher term premium, the extra return investors demand for holding long bonds, not to higher expected rates. Held five years, the bond offers 3.34% a year after inflation if real yields hold, 0.52% if the market prices one more point of growth, and 5.84% if yields fall back to 2.5%.
What would change our view: productivity that keeps running a point faster than its 2007 to 2019 pace, long real yields that hold as the term premium falls, and independent evidence that AI-assisted research yields more discoveries that survive checking, not more papers. The tests are set down below, to be scored on March 15, 2027.
The idea began with a gap. In MIT Technology Review on August 24, 2026, Elise Cutts wrote "Kids outlearn AI, and we still don't know why", about the distance between how much language a child needs and how much a machine does. A preteen raised in a talkative home may have heard about 100 million words; Meta's Llama 3.1 was trained on 15 trillion tokens. Train GPT-2 on 30 million words, the Stanford cognitive scientist Michael C. Frank told her, and "you get a nonsense generator; you don't get a kid."
The gap led Russ to a narrower question, whether cheaper and more data-efficient models would reward customers, the companies building applications or the owners of the computing, and then to a much larger one: could many modest gains at many companies add up to something on the scale of the Industrial Revolution? He put the hypothesis to the Vista Desk and tested it against Mokyr's own texts over three rounds of challenge, with corrections in both directions. His draft concludes that AI could exceed individual industrial technologies over a stated horizon, but has not been shown to exceed the Industrial Revolution as a change in the process of growth. This note keeps that conclusion, rebuilds his models in code, and adds the one market that has an opinion about it.
The question has a personal history. As Russ wrote in his memoir, The Scholars' Treasure (chapter 1, page 1): "I had the good fortune of studying under six Nobel Prize winners during my undergraduate program at Northwestern and at my MBA program at the University of Chicago. Not one of them gave me an A." Since the book appeared, a seventh of his teachers has won. In 2025 Joel Mokyr received half of the prize in economic sciences "for having identified the prerequisites for sustained growth through technological progress," sharing it with Philippe Aghion and Peter Howitt. This note is written in his honor.
In 1712 a blacksmith from Dartmouth named Thomas Newcomen built the first steam engine that truly worked. It was, in Mokyr's description, large, noisy and "voracious in its appetite for fuel," and for most of the eighteenth century it did one job: pumping water out of mines where coal was cheap and flooding a threat (Mokyr, 1999). It barely registered in the national accounts: from 1760 to 1830, steam added 0.014 percentage points a year to the growth of British labor productivity (Crafts and Woltjer, 2021). Output per person in Great Britain grew 0.34% a year from 1760 to 1800 and 0.20% from 1800 to 1830, by our calculation from the reconstruction of Broadberry, Campbell, Klein, Overton and van Leeuwen (2015). Then the pace changed. Output per person grew 1.21% a year from 1830 to 1860, and steam's contribution rose more than twentyfold, to 0.30 points a year from 1830 to 1870.
So the most famous revolution in economic history barely moved the numbers for seventy years. Mokyr's life's work explains why it then did not stop. Prosperity had flowered and faded before; the wealth of Venice and of the Dutch Republic had melted away. What changed in eighteenth-century Britain and its neighbors was not a machine but the way useful knowledge was made, shared and trusted. The cost of access fell: societies, encyclopedias, journals and correspondence made it cheaper to find what was known, to obtain it and to check it (Mokyr, 2005). Confidence in that knowledge, what Mokyr calls its tightness, the consensus with which it is believed, grew with experiment and verification; and because knowledge kept accumulating while the cost of reaching it kept falling, the gains could not easily be reversed. And a large class of skilled mechanics and engineers, the 759 in Meisenzahl and Mokyr's sample, took other people's inventions and improved them until they worked. A great invention opened a door; thousands of small ones furnished the room.
This reminds me of Alfred North Whitehead, in Science and the Modern World (1925): "The greatest invention of the nineteenth century was the invention of the method of invention." Mokyr spent a career explaining that sentence. It is also the right question to ask of AI. Cockburn, Henderson and Stern (2018) asked it of deep learning, which they suggest may be "a new general-purpose invention of a method of invention": not a new kind of machine, but a new way of finding things out.
Russ's draft carries the distinction in a model of Charles Jones (1995), which we rebuilt in code. Knowledge grows by research:
A is the stock of useful knowledge, R the effort devoted to adding to it, δ how productive that effort is, λ how much more effort buys, and φ, below one, the drag that makes each new idea harder to find as the stock grows; Bloom, Jones, Van Reenen and Webb (2020) measured that drag across industries. With research effort growing at gR a year and its productivity at gδ, knowledge grows in the long run at
AI enters first through δ, the productivity of research. The model then gives the title's question a sharp form:
| Growth of knowledge, % a year | Start | 10 yrs | 25 yrs | 50 yrs | Long run |
|---|---|---|---|---|---|
| No change | 2.000 | 2.000 | 2.000 | 2.000 | 2.000 |
| AI doubles research productivity once | 4.000 | 3.652 | 3.275 | 2.871 | 2.000 |
| AI makes it grow 0.5% a year | 2.000 | 2.097 | 2.233 | 2.427 | 3.000 |
Illustrative parameters from the draft, not estimates for AI: λ = 1, φ = 0.5, research effort growing 1% a year, δ = 0.02 at the start. The closed form was checked against a direct integration of the equation.
A tool that doubles the productivity of research once gives a burst of growth that fades over generations: 4% at first, 3.65% after ten years, 2.87% after fifty, and back toward 2%. The economy ends permanently richer and no faster. Only a δ that keeps rising lifts the growth rate for good, here from 2% to 3%. That is Mokyr's distinction in a single symbol. The Industrial Revolution's lasting gift was a δ that kept rising, because knowledge about how to make knowledge kept improving. The question for AI is whether it is a doubling or a slope.
Much of δ is what Mokyr calls access costs: the cost of establishing that a piece of knowledge exists, of finding who holds it, of acquiring it, and of verifying how far the experts agree on it (Mokyr, 2005, pages 296 to 297). Acquiring it, he notes, might mean a search through a catalog, a visit to a site, or "hiring a consultant or expert who could convey it." The industrialists of the eighteenth century paid those costs gladly. In agriculture, pottery, steam engines and chemicals, "leading manufacturers eagerly sought and found the advice of scientists." And much of the knowledge that counted was never written down at all. It lived in skill and dexterity, the tacit knowledge that traveled only with people.
For Russ the subject is personal. He took Mokyr's course in European economic history at Northwestern, and the question at its center, how useful knowledge reaches the people who can use it, became his life's work: not as a scholar, but as an entrepreneur. For thirty-two years he has connected leading experts with litigators, investors and executives, building the kind of organization Mokyr asks about, one that channels knowledge "from those who knew useful things" to those who can act on it. AI has now made the written record far cheaper to search, read and check, and Vista's notes and its Library do that work in public, from the filings to the papers. What AI cannot reach is the knowledge that was never written down, which still lives with the experts themselves; primary research reaches it. Vista combines the two, the written record and what the people who did the work know, in synthesized research. In the model's terms, the work is to lower each of Mokyr's access costs for a decision maker, and so to raise δ, the productivity of every hour spent finding out, one decision at a time.
Russ's first intuition was that modest savings at many firms could add up. They can, if they are weighted by the activity they touch rather than counted firm by firm, the same logic Acemoglu (2024) applies to the whole economy:
w is an activity's share of total cost, e the share of it AI can do, a the share actually using it, s the saving on that share after review, correction and failures, and f the new costs of making it work. With half of all work eligible, 60% adoption and a 20% saving, less 2% of costs for the complements, S = 0.5 × 0.6 × 0.2 − 0.02 = 0.04: a 4% cut in the cost of everything. Below a 6.67% saving on the adopted work, the complements eat the gain. Four percent of an economy is an immense sum. It is also a level, not a rate: once adoption is complete, the gain has been banked and growth returns to its old pace.
| Evidence | Result | What it does not show |
|---|---|---|
| 5,172 customer support agents (Brynjolfsson, Li and Raymond, 2025) | 15% more issues resolved per hour, most for the least experienced | The economy, or today's models |
| 16 experienced open-source developers, 246 tasks (METR, 2025) | 19% slower with AI tools | A universal effect; METR says its newer experiment, which points to speedups, is distorted by selection |
| US workers, second quarter of 2026 (St. Louis Fed, from the Real-Time Population Survey) | 39.2% used it at work in a week; it assisted 6.3% of work hours and saved 2.2% of them | A causal estimate of productivity |
| US firms, November 2025 to January 2026 (Census Bureau) | 18% used AI in a business function; 32% weighted by employment | How intensively they use it |
| Danish workers through 2024 (Humlum and Vestergaard) | New tasks, and precise null effects on earnings and recorded hours, ruling out effects above 2% | Output, or later models |
| OECD scenarios for the US (Filippucci, Gal, Laengle and Schief, 2025) | 0.41, 0.99 or 1.28 points a year of labor productivity growth over a decade | Observed results |
| Acemoglu's task model (2024) | At most 0.66% more total factor productivity in ten years, under 0.53% after adjusting for hard tasks | A ceiling on future discovery |
Each study's own measures, kept separate: use, intensity, task speed and modeled growth are different quantities and cannot be spliced into one curve.
Two things stand out. The good results come where an answer is easy to check, in support queues and routine code; the strongest claim for AI is about discovery, where the evidence is thinnest. And the aggregate numbers have moved. American output per hour has grown 2.51% a year since the fourth quarter of 2022, against 1.52% from 2007 to 2019: a point faster, almost exactly the OECD's middle scenario. Nobody can yet say how much of that is AI. The last time productivity ran this fast, 3.12% a year from 1995 to 2004, the computer boom faded within a decade, and the first half of 2026 ran at 1.12% a year. Electricity, too, took decades to show in the statistics (David, 1990), and a general purpose technology's gains can hide for years behind the intangible investment it demands, the productivity J-curve of Brynjolfsson, Rock and Syverson (2021). A dip proves nothing. Neither does a spurt.
There is one market where the whole economy's future gets a single price. If people expect to be much richer later, they want to borrow against it now and save less, and the real rate of interest must rise to stop them. Frank Ramsey wrote the rule down in 1928:
r is the real interest rate, ρ how much people discount the future for its own sake, θ how strongly they want to smooth what they spend over time (usually set between 1 and 2), and g the expected growth of spending per person. A believer in AI on Mokyr's scale is predicting a higher g, and the rule says what that must do to r.
Trevor Chow, Basil Halperin and J. Zachary Mazlish made this argument first: if markets expected transformative AI, long real rates would be high. Reading the rates in June 2026, they concluded that "the market seems to be strongly rejecting AI timelines of less than ten years." Since then the 30-year real yield has risen more than half a point, from a June average of 2.72%.
The real yield on 30-year Treasury inflation-protected securities was 1.46% on November 30, 2022, the day ChatGPT was released, and 3.31% on October 1, 2026, the highest since the series began in February 2010. The nominal 30-year yield rose by almost the same amount, from 3.80% to 5.61%, so the inflation the market expects, the gap between them, barely moved (2.34 to 2.30 points). The rise is real.
If all 1.85 points were faster expected growth, the market would be pricing 1.85 / θ: between 0.93 and 1.85 points a year of extra growth, for thirty years. That is more than the OECD's scenarios, which last a decade; it is the strongest version of the thesis, a change in the rate. So the move is either a bet on AI at Mokyr's scale or, mostly, about something else.
There is plenty else. The New York Fed's term structure model, from Adrian, Crump and Moench (2013), splits a 10-year yield into the expected path of short rates and a term premium, the extra return investors demand for holding a long bond. From November 30, 2022 to October 1, 2026 its 10-year term premium rose from minus 0.75 to plus 0.91 points, more than the 1.56-point rise in the model's 10-year yield, and the expected path of short rates ended a little lower (4.41% then, 4.31% now). On that reading the market is not expecting higher rates; it is charging more to hold long bonds. The usual causes of a rising term premium, heavier government borrowing and less certainty about inflation and policy, need no help from AI.
The link between growth and real rates is itself disputed. Across countries and two centuries, Hamilton, Harris, Hatzius and West (2016) found the equilibrium real rate's relationship with trend growth "much more tenuous than widely believed"; Chow and his coauthors, with new cross-country data, find that higher expected growth does raise long real rates. And a lasting productivity boom can lift rates through the demand for capital before the growth itself arrives: Federal Reserve Governor Michael Barr said in February that he had raised his long-run estimate of the neutral rate modestly "because of higher productivity." Our reading is that the bond market has moved by about as much as a yes would need, mostly for reasons that are not a yes.
| Held five years, bought at 3.31% | Sold at | A year, after inflation |
|---|---|---|
| Real yields hold | 3.31% | 3.34% |
| The market prices one more point of growth | 4.31% | 0.52% |
| The market prices two more points | 5.31% | −2.01% |
| Yields fall back | 2.50% | 5.84% |
A 30-year inflation-protected bond paying a 3.31% coupon twice a year, bought at par, coupons reinvested at the purchase yield, sold after five years with 25 years left; before tax. The row for one more point assumes θ = 1 and a full point of growth, or θ = 2 and half a point.
This reminds me of John 4:37 (BSB): "the saying 'One sows and another reaps' is true." Technologies are sown by one set of people and reaped, often, by another.
A technology can transform an economy and still disappoint the people who financed it. The draft's own example is small and telling. Fin, a customer-service AI, reported that a small specialized model deciding when to hand a conversation to a person, with a large model as fallback, cut its cost per resolution by about 3% in an A/B test against its earlier approach, a vendor's own figure. Whether that 3% reaches Fin's margin, its customers' prices or its rivals' depends on competition, not on the model. History says gains can take a long road to wages, too: in Britain from 1780 to 1840, output per worker rose 46% while the real wage rose 12%, on the standard series Allen (2009) compares. A financial thesis about AI must show who keeps the saving. The growth thesis cannot do that work for it, and the bond market, which pays the same real coupon to everyone, is the one place where the question of who keeps it does not arise.
The note's two distinctions, a level or a rate, and who keeps the gains, sort the market into hypotheses that can be checked against what companies and markets report. Each comes with what would falsify it; none is a recommendation.
Likely, we can get some of it from primary research, which is what Vista is all about. The people who ran AI-assisted programs, priced the work and traded the bonds can say now what the statistics will take years to show.
We would askAs a matter of general past practice, did the AI-assisted programs produce more candidates that survived independent validation, per dollar of research, than the programs before them, or only more candidates?
The answer that would change the view"More that survived, and the rate kept rising as we changed how we worked."
We would askOver a full year, after review, correction and failed attempts, did the cost of a finished unit of work fall, and by how much?
The answer that would change the view"By more than a fifth, and it kept falling as we redesigned the work around it."
We would askIn your experience, what moves the 30-year real yield by nearly two points in four years, and how would you tell a change in expected growth from a change in the term premium?
The answer that would change the view"Growth. The term premium explains little of it."
For a client engagement, Vista combines this report with that primary research.
We will score this note on March 15, 2027, and publish the result on the scorecard whether it flatters us or not. The tests are fixed today, so they cannot drift, and each can fail. Each would side with the larger reading, AI as a change in the rate of growth:
Three or four passes and the larger reading gains ground; one or none and the level reading holds; two is too early to say.
Newcomen's engine pumped water for decades before anyone could see what it would become. The test for AI is not whether it pumps faster. It is whether it changes how we find the next engine. The decision, as always, belongs to the reader.
The bet. That AI changes the rate at which useful knowledge grows, as the Industrial Enlightenment did, and not only the level of output. For a holder of long real bonds it is the bet in reverse: that 3.31% is not the first installment of a new growth regime.
The payoff. In the economy, on the draft's illustrative model, a one-time doubling of research productivity lifts growth from 2% to 4% and lets it fade to 3.65% after ten years and back toward 2%; research productivity that grows 0.5% a year lifts it toward 3% for good. In the market, a 30-year inflation-protected bond bought at 3.31% and held five years offers 3.34% a year after inflation if real yields hold, 0.52% if the market prices one more point of growth, −2.01% if two, and 5.84% if yields fall back to 2.5%.
Our read. AI clears the bar of a large general purpose technology on the OECD's scenarios, where even the slow case beats steam's best decades, and American productivity has run a point faster since late 2022. The builders' spending, $712.5 billion this year at four companies, assumes either the larger reading or a share of the gains many times what innovators have historically kept. Nothing yet shows a new method of invention: the evidence is about tasks, mixed for experienced experts and silent on discoveries that survive checking, and the bond market's move has rivals that explain it at least as well as growth.
What settles it, and when. On March 15, 2027: fourth-quarter productivity, the long real yield and its term premium, and the share of work hours AI assists. Before then, we would talk with a former head of a research platform about validated discoveries per research dollar, a former engineering leader about whole-workflow costs, and a former inflation-linked bond manager about what has moved long real yields.
The draft's models, rebuilt in code and reproduced to its printed precision before anything was added. British output per person divides the real GDP index for Great Britain (1700 = 100) by the population of Great Britain, both from the authors' own workbook for Broadberry, Campbell, Klein, Overton and van Leeuwen (2015), sheets A7 and A2; rates are compound annual rates between single years (1760 to 1800: 0.342%; 1800 to 1830: 0.201%; 1830 to 1860: 1.206%; 1760 to 1830: 0.282%, a cumulative 21.76%); comparing decade averages instead gives 0.354% a year from the 1760s to the 1830s. Technology contributions are the totals in Crafts and Woltjer (2021), Table 5, percentage points a year of labor productivity growth; their two specifications for electricity in the 1920s are alternatives, not to be added. The knowledge model is Jones (1995), equation 6 in its equilibrium form, with long-run growth as in his equation 8, and with a research productivity δ(t) = δ0 egδt added by the draft, solved in closed form with B = A1 − φ and checked against a direct integration (A after 50 years in the doubling case: 5.2782 both ways). The aggregation is the draft's fixed-output resource cost model; it assumes unchanged quality and no induced demand, price or bottleneck effects. Productivity is the Bureau of Labor Statistics index of nonfarm business output per hour as published by the Federal Reserve Bank of St. Louis (FRED series OPHNFB), compound annual rates between quarters. Yields are FRED series DFII30 (30-year inflation-indexed), DGS30 (30-year nominal) and DFII10, read October 4, 2026. The bond returns price a 30-year par bond with a 3.31% coupon paid twice a year, reinvest coupons at 3.31%, and sell after five years at the yield in each row; the Ramsey arithmetic divides the change in the real yield by θ. Business applications are the Census Bureau's monthly, seasonally adjusted counts; growth compares January to August averages, and the share since 2022 compares the 2022 average with the 2026 average to August. The term premium figures are the ACM model's 10-year term premium and fitted yield from the New York Fed's daily file; the expected path of short rates is the yield less the premium. The builders' arithmetic treats a steady capital spending K a year on equipment with a six-year life and a 9% cost of capital, both assumptions: in a steady state the yearly charge is depreciation K plus a return on the average net investment, K × L / 2, so K × (1 + 0.09 × 3) = 1.27 × K; it is compared with the OECD scenarios' level gain after ten years, (1 + p)10 − 1, applied to 2025 American output. Every figure in this note was recomputed in code from these inputs.
| Input | Value | Source |
|---|---|---|
| GB real GDP index, 1760 and 1830 | 148.969; 383.167 | Broadberry and coauthors (2015), workbook sheet A7 |
| GB population, 1760 and 1830 | 7.645 million; 16.15 million | Same, sheet A2 |
| Steam, UK labor productivity contribution | 0.014 (1760 to 1830); 0.30 (1830 to 1870); 0.31 (1870 to 1910) | Crafts and Woltjer (2021), Table 5 |
| Electricity, US | 0.10 (1899 to 1919); 0.14 or 0.37 (1919 to 1929) | Same |
| AI scenarios, US | 0.41; 0.99; 1.28 points a year | Filippucci and coauthors (2025), Table 2 |
| Knowledge model | λ 1; φ 0.5; gR 1%; δ 0.02 | The draft's illustration |
| Aggregation | e 0.5; a 0.6; s 0.2; f 0.02 | The draft's illustration |
| 30-year real yield | 1.46% (Nov. 30, 2022); 3.31% (Oct. 1, 2026) | FRED DFII30 |
| 30-year nominal yield | 3.80%; 5.61% | FRED DGS30 |
| Output per hour | 2.51% a year (2022 Q4 to 2026 Q2); 1.52% (2007 Q4 to 2019 Q4) | FRED OPHNFB |
| θ | 1 to 2 | Assumption, the usual range |
This note is research, not investment advice. It states what prices and models assume under labeled assumptions; nothing here is a recommendation to buy, sell or hold any security. The decision belongs to the reader. As of October 4, 2026, Russ Rosenzweig, Vista's founder, owns no Treasury inflation-protected securities and no shares of Accenture, Alphabet, Amazon, Cognizant, Concentrix, Infosys, Meta Platforms, Microsoft, Recursion Pharmaceuticals, Schrödinger or Teleperformance. He owns shares of AMD (see Research Note No. 5), a supplier of chips to the builders, which this note does not analyze. The note began as Russ Rosenzweig's research question and draft, which the Vista Research Desk challenged over three rounds, with corrections in both directions; the bond market section is Vista's addition, and the models and every figure were rebuilt in code and checked against the primary sources.