Lecture notes

Day 4

Lectures from Chicago, Thursday, July 30, 2026.

Chicago

Thursday, July 30, 2026

3 lectures

Open the full notes (PDF)

Lecture 1

AI and Economic Growth

A lecture by Ben Jones

Gordon and Llura Gund Family Professor of Entrepreneurship and Professor of Strategy, Kellogg School of Management, Northwestern University.

The bottom line

Any claim about what AI will do to growth and inequality is really a claim about three things: the share of tasks machines take over, how productive they are at those tasks, and how strong the bottlenecks are that tie tasks together. Because the economy is dominated by the things we cannot yet improve, bottlenecks make transformative, singularity-style acceleration unlikely, but they also mean that very capable AI can push value toward the workers who still do the hard, un-automated tasks.

A framework: three forces behind any AI claim

Economists model production as a set of tasks. To make anything (bread, a car, a cancer treatment, a unit of energy) you must complete many tasks, and each can be done by labour or handed to a machine. Total output is not a simple sum but a kind of average of what you produce at every task, and the way you average matters enormously. Written as a generalized (Hölder) mean, one parameter, $$, decides whether output behaves like an arithmetic mean (big wins at any task drag the whole economy up), a harmonic or minimum function (the weakest links dominate), or something in between.

That gives three forces to track. First, how many tasks AI can take over. Second, how good it is at the tasks it takes, like arithmetic, where a phone chip beats a human by hundreds of billions to one, or like driving, where it may only match us? Third, how strong the bottlenecks are, whether the tasks we remain bad at quietly hold the whole process back. Automation thus has a horizontal dimension (the share of tasks moved to machines) and a vertical one (how much more each machine produces); public debate fixates on the first, but the second and third do most of the real work.

Automation, abundance, and labour's share

Moving tasks to machines raises output per worker almost mechanically: as people are squeezed into a smaller set of tasks, more hands per task make each one abundant while machines cover the rest. The effect on inequality is subtler and cuts both ways. More automation shifts income from labour to capital, because a firm now pays machines for tasks labour used to do, the fear that runs from Marx to today's headlines. Yet Marx's prediction failed: despite two centuries of automation, labour's share stayed near two-thirds in the United States and France for a very long time.

The reason is the vertical dimension. When a machine becomes extremely productive, it floods the market and the price of what it makes collapses, so it captures little of the economy, think of search or email, given away for almost nothing because computing is so cheap. Value then flows to the tasks that stay expensive because we are still bad at them, and those are largely done by labour. So "a lot of automation" and "rising labour share" can coexist. The dangerous case is mediocre automation, AI only slightly better than a worker but far cheaper, which takes the job without making anything meaningfully cheaper; the recent decline in the U.S. labour share fits automation accelerating while machine quality has not risen much.

Evidence at a glance

Computers "crush" us where they win. A person needs about a minute to multiply two four-digit numbers; a phone chip does about 500 billion such operations in that minute, not a little better, but ridiculously so.
Bottlenecks are stronger than infinity. If output is a harmonic average, making half of all tasks infinitely productive and free raises total output only from 1 to 2, not to infinity, or even to 5.
Efficiency does not equal share of GDP. Computing power per dollar rose by a factor of about $10^17$ over roughly 50 years while overall productivity rose about threefold; information-technology equipment is still only about 4.5% of U.S. GDP.
Where the money actually goes. Meals away from home are about 12% of GDP and housing about 30%, sectors famously resistant to productivity gains (a clogged sink can cost $1,000 for a plumber).
The farm shows the pattern. A combine harvester once brought in about 6 million pounds of corn in 12 hours, yet corn near the farm costs a few cents a pound while corn chips in the store cost about $4.50, the un-automated steps set the price.

Implications for jobs, skills, and developing economies

Judge the forces, not the hype. Rather than predict a single number, ask of any sector: what share of tasks can AI do, how good is it at them, and where are the bottlenecks? That is what separates ordinary disruption from a genuine structural break.

Expect skill-biased pressure, with a twist. Automation tends to replace lower-skill and routine work first. But because AI is a cognitive technology, physical bottleneck jobs (the plumber, the electrician) look relatively safe, while routine white-collar work is more exposed.
Baumol's cost disease still rules. As agriculture and manufacturing were automated, their GDP shares fell and services rose; real incomes climb because everything else gets cheap, not because the service itself speeds up.
Speed brings its own risks. Near-2% growth is ordinary disruption the U.S. economy handles well; a jump to 5% would be ahistorical and could provoke resistance, especially where entrenched incumbents block change, a danger for emerging markets.
Two-sided prospects for poor countries. Automation is investment, so weak governance makes adoption hard and could widen divergence; but AI's latency means server farms need not be local, so (like mobile phones) cheap, even open-source AI could be accessed from afar as a broad enabler.
Innovation faces the same bottlenecks. Telescopes and regression software far out-produce human eyes or hand calculation, yet conceptual and data bottlenecks make an R&D "singularity" doubtful, even as useful acceleration in health and technology remains possible.

Lecture 2

A Master Class on Business Dynamism

A lecture by Ufuk Akcigit

Co-Director, Growth Academy and Arnold C. Harberger Professor of Economics, University of Chicago (on leave), Deputy Chief Economist, World Bank Group; with Furkan Kilic, Postdoctoral Scholar, Growth Academy, University of Chicago.

The bottom line

Economies cannot be run as physical experiments, so the way to design credible policy is to build a quantitative "laboratory": high-quality microdata, disciplined by theory, calibrated to match a real economy, and then used to run counterfactual policy experiments. Because any single fact can support many competing stories, only a model that confronts all the symptoms at once can identify the true cause, and get the policy right.

The economist's laboratory

Chemists have a lab and physicists have their accelerators; economists have neither. We cannot spin up ten copies of Germany to try ten different policies. The substitute is economic theory: build a model whose firms and inventors behave like the real ones, and it becomes a laboratory in which counterfactual experiments can be run many times over, as one would replay a city-simulation game. The workflow runs from microdata to structural modelling to computational policy analysis, and every assumption bears directly on the conclusion, so the theory must be built on reliable data.

Microdata are essential because aggregates hide what matters. Two countries can post identical headline growth yet differ completely underneath: for the same innovation, the United States generates about twice as many jobs as Europe, and roughly half of all growth comes not from innovation itself but from reallocation (resources moving from less to more productive firms) a process far slower in some economies. Frictions live at the micro level, so policy must be designed there too.

Symptoms of declining dynamism

Read together, U.S. microdata since 1980 describe an economy losing dynamism. Market concentration has risen, average markups (price over marginal cost) have climbed, the profit share of GDP has roughly tripled, and labour's share has fallen. These are connected: within essentially every sector, higher concentration is associated with a lower labour share, so weaker competition leaves labour on the losing side. Meanwhile frontier firms keep advancing while laggards fail to keep up, the firm entry rate has trended down, the employment share of young firms (under five years old) has shrunk, job reallocation has slowed, and the dispersion of firm growth rates (a clean summary measure of risk-taking) has narrowed.

The danger is misdiagnosis. Any one symptom is consistent with many explanations, and a single data point can be fitted by many lines. Competing single-cause stories, lower corporate taxes, incumbent subsidies, changing entry costs, weaker knowledge diffusion, low interest rates keeping "zombie" firms alive, ideas getting harder to find, declining union power, each imply a different remedy, and treating the wrong cause is costly. Like a doctor, one must weigh all the vital signs before prescribing.

Evidence at a glance

We know how many people, not how many firms. A quick search agrees the world holds about 8.2 billion people, but there is no agreed count of firms, a gap the World Bank is closing with a firm-level database spanning 64 countries over 1995–2025, including panels for 17 of them.
Micro firms are everywhere; the giants once were small. Firms with fewer than five workers dominate the count, yet 30–40% of today's large firms (over 100 workers) were smaller a decade earlier, and were already more productive years before they grew.
Young firms create the jobs. Most net job creation comes from young firms, and roughly 5–10% of firms ("high-growth" firms) create 50–60% of jobs, in all sectors, from textiles in Ethiopia to food and beverages in Hungary, not only in technology.
Jobs are often not durable. In one Middle-Eastern economy, 55% of new matches dissolve within a year and 80% within five; of those who leave in year one, 34% exit the formal labour market, a "slippery jobs ladder" where churn does not lift wages as it does in advanced economies.
Estimates you cannot guess. Calibrated to data, "expansion efficiency" is about 0.19 in the United States versus about 0.5 in India, a fourfold gap only structural estimation reveals.

From diagnosis to policy

Test every story in the model. Feeding each candidate cause into a calibrated model shows that most fit only one or two facts: declining corporate taxes explain the growth-rate dispersion alone, and R&D subsidies little more. The channel that matches all the moments, qualitatively and quantitatively, is weaker knowledge diffusion.

Diffusion, not just invention. Creating an idea is one thing; spreading it through the economy is another. Killer acquisitions, patent barriers, and incumbents that deter entrants all reduce the diffusion of knowledge, which the model identifies as the core problem.
Put talent first. With the same budget, subsidizing R&D, subsidizing schooling, or opening more university slots allocate talent very differently. R&D subsidies look effective at first (firms simply buy more equipment), but beyond about year eight the strongest policy is to strengthen the talent pool through education, especially by giving talented but poor children access to school.
Capability determines where subsidies work. Estimated for Germany, the best R&D policy for Eastern citizens is to subsidize the more capable Western firms and let knowledge diffuse, a euro given to a low-capability firm is largely wasted.
Retire legacy policies. The first federal U.S. R&D tax credit dates to 1981 (Minnesota followed in 1982); such temporary measures are renewed indefinitely for political reasons. Shutting down what no longer works frees resources for what does.

Lecture 3

Skipping the Factory: Service-Led Growth and Structural Transformation in the Developing World

A lecture by Fabrizio Zilibotti

Tuntex Professor of International and Development Economics, Yale University.

The bottom line

Today's developing economies are leaving agriculture for services, largely skipping the factory stage that defined earlier transformations. This is not automatically a symptom of failure: careful measurement shows genuine productivity growth inside local consumer services. But the gains are unequally shared, they favour richer, urban households, while poorer, rural households gain most from productivity in tradable goods such as agriculture.

A different road out of agriculture

The canonical story of development is a sequence: labour leaves agriculture for industry, and only later moves into services. Italy ran this course, about half its workforce farmed in 1948, industry then expanded, and by 2019 services dominated, and the United States traced the same path a lifetime earlier. Plotted against income per capita, industrial employment follows a clear hump: it rises, then falls. Crucially, this decline is not mainly about China; Italy's industry shrank on the same trajectory the United States followed between 1880 and 1975, long before Chinese trade mattered.

The developing world today is taking a different route. As agriculture contracts, workers move overwhelmingly into services rather than manufacturing, "skipping the factory," or premature deindustrialization. Across countries poorer than China, the pattern is remarkably tight: for every 10-percentage-point fall in agriculture's employment share, manufacturing rises about 2.4 points while services rise about 6.4. India shows it, as do fast growers such as Ethiopia and Tanzania. The relationship holds for fast- and slow-growing countries alike, so it is not a mark of weak performance.

Do services deserve their bad reputation?

Services have long been treated as second-class. Goods are tangible and tradable; services are intangible and typically require producer and user to interact on site, which makes many of them local. A hierarchy runs from Marx's distinction between productive and unproductive labour, to the Soviet "material product" accounts that excluded health, education, finance, and personal services, to the modern view that manufacturing (with its scale economies, tradability, and clustering) is the true engine of productivity, so that industrialization was development.

The current transformation is dominated not by high-skill producer services (ICT, accountants, lawyers, designers) but by local consumer services, retail, personal services, health. Do they carry productivity growth, or signal its absence? Prices cannot settle the question, because observed service prices confound quality with efficiency and are often unavailable locally. The lecture's method sidesteps prices: since consumer services must be produced where they are consumed, household spending (demand) can be matched against local employment (supply), and the gap between demand-driven and actual employment growth reveals local service productivity. Applied through non-homothetic preferences to 380 districts in India and to Sub-Saharan Africa, it yields a striking result, productivity growth in local consumer services is real, and often higher than in tradables.

Evidence at a glance

The tight reallocation. For every 10-point decline in agriculture's employment share across developing countries (1991–2019), manufacturing rose about 2.4 points and services about 6.4, most of the exit went straight to services.
Services out-grew tradables in productivity. Estimated productivity growth in non-tradable consumer services exceeded that in tradables in both urban and rural Sub-Saharan Africa, a pattern that also holds in India and, recently, China.
These are not merely reserve jobs. Controlling for location and education, consumer-service workers have access to electricity and piped water similar to or slightly better than manufacturing workers.
Gains favour the better-off. On a willingness-to-pay measure, rural poor households value tradable-goods productivity at about 16% of income (versus about 11% for the rich), while consumer-service gains are valued mostly by richer, urban households.
Place matters enormously. Estimated consumer-service productivity growth reached about 11% a year in Bangalore versus about 2.4% in a poor rural district, so the rich in dynamic cities gain most, the rural poor from agriculture.

Policy implications

Treat services as productive assets, not consolation prizes. Service-led growth can be genuine growth. The task is not to force a return to the factory, but to raise the efficiency of services and connect them to the rest of the economy.

Understand where the productivity comes from. Three channels appear to drive it: agglomeration externalities (denser consumer cities support more and more varied providers), the decline of informality (which lifts barriers to better jobs), and complementary technology and infrastructure, mobile phones, local banks, and property registries that support inventory, payments, and credit.
Do not abandon industrialization, but keep it grounded. The benefits of industry remain real; the mistake is overambitious plans detached from a country's comparative advantage.
Build on revealed strengths and linkages. Target clusters where an economy is already strong (for example, industries connected to tourism) and use linkage logic beyond the manufacturing input–output table.
Raise scale and efficiency in services. Encourage the entry of more formal, larger providers rather than over-protecting small incumbents, as early Indian policy did.
Read the shift correctly. What looks like deindustrialization is better understood as slow industrialization alongside a real service transformation, to be strengthened, not treated as failure.