Economic growth is a unique field, because it is relevant to both the global development side of EA and the AI side of EA. Global development policy can be informed by models that offer helpful diagnostics into the drivers of growth, while growth models can also inform us about how AI progress will affect society. My friend asked me to create a growth theory reading list for an average EA who is interested in applying growth theory to EA concerns. This is my list. (It’s shorter and more balanced between AI/GHD than this list) I hope it helps anyone who wants to dig into growth questions themselves.

These papers require a fair amount of mathematical maturity. If you don’t feel confident about your math, I encourage you to start with Jones 2016 to get a really strong grounding in the facts of growth, with some explanations in words for how growth economists think about fitting them into theories.

Basics of growth

These two papers cover the foundations of growth theory. They aren’t strictly essential for understanding the other papers, but they’re helpful and likely where you should start if you have no background in growth.

Jones 2016

Sociologically, growth theory is all about finding facts that beg to be explained. For half a century, growth theory was almost singularly oriented around explaining the “Kaldor facts” of growth. These facts organize what theories are entertained, even though they cannot actually validate a theory – after all, a totally incorrect theory could arrive at the right answer by chance. In this way, growth theorists are engaged in detective work; they try to piece together the stories that make sense given the facts, making leaps when they have to.

This places the facts of growth squarely in the center of theorizing, and Jones 2016 is the most comprehensive treatment of those facts, with accessible descriptions of how growth models try to represent those facts. You will notice that I recommend more than a few papers by Chad Jones in this list. That’s because he is by far the best writer in the growth literature. His exposition of complex ideas and coverage of the big picture is just not matched by any other growth economist.

Jones 2005

While Jones 2016 focuses on the facts of growth, Jones 2005 is an overview of the most common kind of long-run growth model – the idea-based model. Historically, economists used models like the Solow model, in which growth came from accumulating capital. But capital-based models are basically incapable of predicting sustained long-run growth. The now-canonical Romer model made a breakthrough by refocusing growth on ideas. Unlike machines, ideas are infinitely reusable. So if growth comes from creating new ideas, and ideas don’t get used up, we can generate sustained long-run growth. This is another demonstration of the aesthetic of growth theory: theories are celebrated if they can match facts that economists think are important.

The idea-based growth model is way too canonical to leave off this list. But I personally think it is overrated by EAs and not necessarily applicable to the issues we care about (AI or global development). I include it because working through it will help you understand the mechanics of growth models more generally, including the more complicated ones covered below.

Growth theory for AI progress

The growth theory that focuses on AI progress builds off the canonical growth work listed above, but with important advances from it. Few papers really apply their framework directly to AI, so it’s important to try and extrapolate what the model implies about AI. The three papers below are the ones whose frameworks I think are most applicable to thinking about AI impacts.

Aghion, Jones and Jones 2019

Aghion, Jones and Jones 2019 cover the most direct treatment of how AI could impact growth. They augment the standard idea-based growth model in a few ways, allowing AI to automate both production and research (i.e. generating more ideas that create growth) They argue that automation of production is consistent with recent short-run evidence in the US, but not dramatic in impact. AI only has dramatic impacts on growth when it can automate research. Since it’s pretty plausible that AI can have large impacts on the R&D process, and R&D is central to growth in the idea-based model, the simple conclusion is that AI could have dramatic impacts on growth.

I think this model has been quite influential for how EAs working on AI takeoff think about it – for example, Tom Davidson’s takeoff speeds model is quite heavily based on Aghion, Jones and Jones.

Jones and Liu 2024

It’s nice to talk about “ideas” being the engine of growth. But when we think about new technologies that have driven economic growth and quality-of-life improvements, they are material. We don’t fly on the idea of an airplane, we fly on an actual airplane. We can say that new ideas are embodied in capital goods like airplanes or tractors. Jones and Liu 2024 explore exactly this idea. They imagine a growth model in which there are two kinds of technological progress. The first is automation: more and more production tasks can be done by machines. This automation potential is what Aghion, Jones and Jones 2019 consider, among other papers. The second is capital-augmentation: making machines better at doing a task that they already do. With both of these forces, we can have growth that matches the key growth facts while being more representative of how we think of technological advances than idea-based models.

Capital-embodied growth models were historically banished by Uzawa’s theorem, which seemingly proved that the constant growth observed over the past 200 years in the US was impossible if technology augments capital. But capital-embodied models have been sneaking back into favor, mostly through empirical evidence that they can explain cross-country productivity differences and wage inequality trends in the US over the past few decades. Jones and Liu strengthen this return by showing that in their model, the two kinds of technological progress exactly cancel each other out so that their model is compatible with Uzawa’s theorem, and thus that it can be consistent with balanced growth. This is another demonstration of how growth theorists place a high premium on being consistent with the facts of growth.

This kind of capital-embodied growth model offers an interesting counterbalance to idea-based growth models in how to think about AI progress. If growth is based on ideas, then the impact of AI progress on growth is limited only by AI’s ability to generate new ideas. However, if growth is embodied in capital goods, then the impact of AI progress on growth is allso bottlenecked by physical manufacturing productivity. Intuitively, even if AI gives us a thousand new technologies to manufacture, if we cannot manufacture these technologies productively, then growth will remain slow. Alternatively, it points to a different set of tasks that it’s important to set AI towards. Whereas idea-based models imply that we should point AI towards generating as many ideas as possible, capital-embodied models imply that physical embodiments of AI (e.g. advanced robotics systems) could have larger effects on growth by increasing our ability to translate ideas into newer and better machines.

In general, I think that capital-embodied growth models tell a different story about AI impacts than the one that idea-based growth models tell, and this has important implications for EAs – but I’ll back up that assertion another day.

Klette and Kortum 2003

There is a notable lack of firms in growth models. Most growth models abstract away from the idea that businesses are real and that their activities matter in any way for growth. In particular, firms innovate and create new products, and they creatively destroy other firms. Klette and Kortum 2003 is the canonical example of this kind of Schumpeterian growth model. It focuses on how firms compete with each other to make new products and sell them, and how the process of R&D affects that competition and ultimately growth. It tries to microfound the overall rate of innovation in the rates of innovation in each firm, making it well-suited to thinking about policy questions about R&D in firms.

It is also an excellent example of a quality ladder growth model. In growth theory, there is a deep distinction between “expanding variety” models and “quality ladder” models. In the canonical Romer model, ideas create new goods. But these new goods are not better than old goods – the model has no conception of “better goods”. They are just different from previous goods, and the model assumes that expanding variety has value. In contrast, quality ladder models imagine a fixed set of goods. What innovation buys you is not new goods, but higher-quality versions of the same good. New goods are exactly the same as old goods, but better in every way. Klette and Kortum is a canonical example of this latter kind of growth model, since firms are competing to make the best version of a fixed product and sell it to the whole market.

If you take the Schumpeterian growth model as a lens to view AI progress, it tells you that AI-driven productivity improvements only impact the economy when they are embodied in firms’ new products. Thus, if there are barriers to firms adopting AI-driven productivity improvements (e.g. worker inertia or resistance to AI tools), then Klette and Kortum would tell you that the impact of AI on growth would be muted. In general, Schumpeterian models like Klette and Kortum focus on firms, and are most useful for analyzing questions about innovation and R&D that are specific to firms.

Growth theory for global development

The growth models covered above just don’t feel very applicable in developing countries. They mostly assume away problems that we know to be true in developing countries. But a large branch of growth theory has focused precisely on developing models that are tailored to developing countries. The three papers below are useful for learning about growth theory through the lens of development.

Banerjee and Duflo 2005 + Hsieh and Klenow 2009

Banerjee and Duflo 2005 deconstruct some problems with neoclassical growth theory using careful data work to show that the standard theory makes incorrect predictions about developing countries. Most importantly among these problems is that the neoclassical growth model assumes that all firms are identical. But Banerjee and Duflo argue that firms have different marginal returns to capital; the most productive firms have very high returns to capital, while many other firms have low returns. In an efficient market, this should be impossible; a bank should prefer to reallocate some of its loans towards these high-return firms, which would increase overall productivity (since capital goes to the firms for whom it’s most productive). So they conclude that there is misallocation of resources, that overall productivity is suppressed because the most productive firms don’t get the resources they need to expand.

Their attempt to model this misallocation is half-hearted, though. Hsieh and Klenow 2009 formalize it in the now-canonical way; they build a model where firms are different from each other, and some firms are implicitly “taxed” (e.g. subject to onerous regulations) while others are implicitly “subsidized” (e.g. given contracts because of political connections). This would reduce aggregate productivity, and thus incomes. This is the now-influential idea of misallocation. Aggregate productivity suffers when unproductive firms are artifically large, while productive firms are held back.

An older literature in growth and development emphasized the importance of “good institutions” for economic growth. Misallocation has inherited the mantle of this literature as The Explanation for why poor countries are poor, because it formalizes how institutions affect growth. Misallocation occurs when unproductive firms take resources that would have otherwise gone to productive firms. Why might that happen? Maybe the unproductive firm was politically connected. Maybe the productive firm was hamstrung by regulations that the unproductive firm was not subject to. These are the footprint of “bad institutions”, and now we can formalize exactly why they are bad for growth: because they distort the allocation of resources across firms.

Note that even though misallocation is probably the most important topic in growth and development in the past 20 years, its importance is not empirically settled. For technical reasons, the canonical method laid out by Hsieh and Klenow could overestimate the extent of misallocation (see this, or this) and we could thus be assigning it undue importance as an explanation for cross-country income differences.

Herrendorf, Rogerson and Valentinyi 2014

It is probably painfully obvious to some readers that thinking about “growth” in the aggregate economy masks very important differences between sectors within an economy. The sectoral transformation has been one of the most important features of growth. Every rich country has industrialized, and every poor country is agrarian. This is a fact that demands an explanation, and growth theorists have duly obliged. Herrendorf, Rogerson and Valentinyi 2014 is a review article that summarizes both empirical and theoretical work on structural transformation. This article is extremely dense but it’s a must-read if you want to think about structural transformation, or agriculture in economic growth.

Jones 2011

One of the conclusions from Hsieh and Klenow 2009 is that if we eliminated all misallocation from India and China, it would increase their productivity by 2-3x. This is certainly large, but it pales in comparison to the income gaps between them and the US (6-30x). This is a common theme with growth theories that seek to explain cross-country income differences. Income differences are so vast compared to other differences between countries that it’s hard to see how these smaller differences could be amplified into the massive income differences we observe. Almost all theories predict smaller income gaps than we actually observe. How do we fix that?

Jones 2011 attempts this amplification using two very old ideas in growth and development. The first idea is inter-sectoral linkages: rather than there just being one sector as in neoclassical growth models, we can imagine there are many sectors, whose outputs are used as the inputs in other sectors. This creates a loop where a sector’s output indirectly determines its own inputs, which magnifies problems in each sector. The second idea is complementarities: we need a lot of inputs to produce each good, so having less of even one input is enough to reduce overall output. When both of these effects are present, even small differences in productivity between countries can be magnified into large differences in income.

Rather than being a fundamental explanation for why poor countries are poor, this paper focuses on the mechanics of how even modest differences between countries’ productivities can be magnified into large differences in their income levels. These discussions of how to kick the tires on a growth model’s numbers are more abstract and difficult to follow than the fundamental economic intuition behind a model. But they are actually quite representative of how growth models are used in practice. In aiming to match models to facts about the growth, the most important one is just how large cross-country income differences are. Most features of countries covary with income, but not enough to fully explain these differences. So plugging in numbers based on macro statistics is the way you actually figure out whether a model does a good job at explaining the data. Banerjee and Duflo 2005 is another great example of this kind of tire-kicking, albeit on a simpler neoclassical growth model.

Reading advice

I’ll conclude with some advice on reading dense economics papers like these ones.

Know your goals; the returns to understanding growth theory are nonlinear. It’s great to have a basic understanding of it, and it’s great to understand it well enough that you can think really critically about applying growth models to any situation. But the middle valley of understanding is basically extra effort for no benefit. If you aim for that basic understanding, even just reading the first two papers is far more than enough. If you aim for that deep understanding, then I encourage you to read and re-read a paper until you feel like you understand it very deeply. This can mean spending a week or two on each paper, just letting the logic sink in. Once you recognize the logic in one paper, it will become quick to understand it in other papers.
The introduction is often sufficient to understand the paper. If you want to deeply understand a growth model to the level where you could write a variant of it yourself, then you need to read the theory section of the paper in great depth. But if you only want to understand it at a high level, the introductions of economics papers are designed for exactly that. They explain the key insight behind the model, or the key facts, at a level that is sufficient for 99% of readers. (Review articles and handbook chapters are not like this, unfortunately.)
Think about what qualitative story a model is telling. Growth theory is mathematical storytelling. Writing a growth model never proves anything to be true – it only conveys a story in precise language. Most of the value of reading a growth model is in grasping this story. Don’t throw yourself at understanding every equation of a growth model until you understand what story it is trying to tell. This will be harder than you think.
Treat model assumptions as ways to conjure up a desired outcome, rather than literal statements about the world. For example, in many growth models, we create “output” by aggregating labor and capital in an arbitrary-looking Cobb-Douglas function, $Y = K^{α} \cdot L^{1 - α}$ . What the hell does this even mean? Do we actually think that inputs are combined in some multiplier function with weird-looking weights? No. This function just has the property that if you work it out, capital makes up a constant share $α$ of costs, which used to be a well-established historical fact. In other words, it was an assumption made to match some desired outcome, rather than a real statement of what we think production looks like. Many model assumptions are like this. They seem inscrutable at first, until you work through the model and see how they make everything else work out smoothly. Recognize those instrumental reasons to make assumptions, rather than spending a lot of time constructing a story behind them when the story isn’t important.

Growth theory for EAs – reading list and summary