Week 1: Introduction to Artificial Intelligence, Innovation, & Growth

[jamesallenevans]

Please post your memo that engages any of the week's readings about a theoretically interesting empirical case regarding artificial intelligence, innovation, and/or growth you anticipate will become the basis of your final project.

• “Deep learning.” 2015. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Nature 521(7553): 436-444.
• “Prediction policy problems.” 2015. Kleinberg, Jon, Jens Ludwig, Sendhil Mullainathan, and Ziad Obermeyer. American Economic Review 105(5): 491-495.
• Power and Prediction: The Disruptive Economics of Artificial Intelligence, Preface and Chapters 1-3 (pp. ix-41). 2022. Ajay Agrawal, Joshua Gans, Avi Goldfarb. Harvard Business Review Press.
• “The Rise of American Ingenuity: Innovation and Inventors of the Golden Age”. 2017. Ufuk Akcigit, John Grigsby, and Tom Nicholas.
• “World Development Report 2024: Middle Income Trap”. The World Bank Group.

Post by Thursday @ midnight. By 12pm Friday, each student will up-vote (“thumbs up”) what they think are the five most interesting memos for that session. These memos should: 1) test out ideas and analyses you expect to become part of your final projects; and 2) involve a custom (non-hallucinated) theoretical and/or empirical demonstration that will result in the relevant analytical visualization. Some of the top-voted memos will form the backbone of discussion in Friday's discussion sessions.

[ShiyangLai]

Shiyang Lai

Research Question.

How do institutional frictions—rather than differences in latent inventive ability—systematically suppress the realization of invention among underrepresented groups, and which targeted interventions most effectively reduce these barriers?

Motivation and Conceptual Background.

Recent work documents stark demographic regularities among U.S. inventors: inventors are disproportionately male, white, highly educated, and from high-income families. Akcigit, Grigsby, and Nicholas (2017) show that parental income and education strongly predict entry into invention, while having little effect on inventive productivity conditional on entry, suggesting that access constraints—not talent—drive observed disparities . However, because historical and administrative data only observe realized inventors, they cannot identify individuals with high inventive potential who never clear institutional thresholds. This project addresses that missing counterfactual population.

Research Design.

I propose an agent-based simulation study that models latent inventors rather than observed patentees. Agents are endowed with identical distributions of cognitive ability, creativity, and learning capacity, but differ in their exposure to institutional frictions that mirror empirically documented barriers: access to education, financial slack, mentorship networks, geographic mobility, and evaluative bias. These frictions are formally mapped onto the micro- and macro-level cost parameters in Akcigit et al.’s endogenous growth framework, where higher costs reduce entry into invention without affecting productivity conditional on success.

Using this environment, I simulate multiple policy “levers” not observable in historical data: (i) early-stage capital subsidies, (ii) randomized mentorship assignment, (iii) blinded evaluation of early ideas, (iv) reduced credential thresholds, and (v) network rewiring that increases exposure to innovative clusters. Each lever is evaluated by comparing realized invention rates, time-to-first-invention, and aggregate innovation output across demographic groups, holding latent ability constant. The design allows causal inference over counterfactual interventions while remaining grounded in empirically defensible assumptions from the historical record.

Contribution.

This project reframes demographic disparities in invention as an equilibrium outcome of institutional design rather than innate differences, and demonstrates how AI-based simulation can function as a laboratory for testing innovation policy in counterfactual populations that never appear in historical data.

[lglickman-edu]

AI as an Infusion Accelerator in Middle-Income Economies

Research Question: Can artificial intelligence reduce the institutional and organizational thresholds required for successful technology infusion in middle income economies to ease the gap that must be overcome to reach high income status.

As presented in the World Development Report 2024 (WDR 2024), most middle income countries (MICs) stagnate not because of a failure to innovate but rather they prematurely shift their financial policy towards a 3i approach. In prematurely doing so, they oversee the importance of infusion of existing global technologies into domestic firms and sectors as they try to skip to the frontier. Historically, infusion as a part of the 1i → 2i → 3i growth pathway required high managerial capacity, strong institutions, a large skilled workforce, and organizational knowledge; however this memo proposes how AI can fundamentally alter these infusion constraints by acting as an infusion accelerator, not an innovation engine.

Infusion, by definition, is the process of taking existing solved problems and implementing them into a new society. This appears to be the perfect use for AI as the knowledge attempting to be infused is known and therefore can be codified and guided as best practices. In doing so, what would have historically required a lot of human capital and expertise can be accomplished with far fewer people as AI would act in the role of managerial and institutional input. For example, AI-driven logistics would optimize supply chains without the need for years of expertise and deep organizational learnings or predictive maintenance algorithms improving quality control and aiding plant managers. This would enable MICs to infuse frontier production techniques at lower institutional cost (as seen in the figure below). While importantly not shifting the innovation barrier, this would enable MICs to start the infusion process earlier, saving years of slow infusion, enabling them to accomplish the 2i approach for much less time, capital, and emissions, bringing them to an even ground where they can strive and innovate.

As such, building on WDR 2024’s emphasis on institutional bottlenecks to infusion, I propose a sector level comparative study across middle income countries that would exploit variation in AI task suitability across sectors. Sectors would be identified to be suitable for AI substitutive organizational/manufacturing input (i.e. the tasks in which infusion is essential) versus not, where AI primarily complements existing capabilities (such as advanced R&D). As such, I suspect that AI-substitutable sectors will exhibit faster and more uniform infusion outcomes (even in countries with weaker institutions) relative to non-substitutable sectors which would enable us to identify AI’s role without conflating it with frontier innovation.

This framework would therefore reframe AI’s role in development, shifting it from a shortcut to innovation to a general purpose infrastructure that would enable diffusion at scale. If proven correct, this would suggest that policymakers for MICs should focus less on 3i style AI-led innovation hubs and more on AI-enabled absorption of existing technologies.

Name: Lucas Glickman

[popov-aleks]

$\textbf{Research Question:}$ Has the increasing dominance of corporate patenting reduced the extent to which innovation translates into social mobility?

$\textbf{Justification:}$ “The Rise of American Ingenuity” concludes that, from 1920 to 1940, having a high patent concentration was strongly correlated with upward mobility in a state. From Figure 1 of the paper, however, we see that patents are becoming dominated by corporations where profits likely go to the already-wealthy. This question addresses whether the large companies “crowd out” individual innovation, and hence a large source of social mobility. This might be particularly important in the age of AI, with vast resources required to compete with the large and well-financed companies. We might be able to look at past precedent to see whether top-driven innovation has significantly harmed social mobility.

$\textbf{Design:}$ We can incorporate two separate analyses into this. One would be whether, at any given moment, regions (states) with a higher proportion of corporate patents experience higher or lower social mobility. The second could be diff-in-diff studies using corporate entry as a “shock” to the proportion of patents such as, for example, large oil or agricultural companies entering areas with traditionally smaller businesses. For both, we would like to consider data from 1960 onwards, when the share of corporate patents plateaued.
We consider the first case for further analysis. Consider the following model for our regression:
$$Mobility_{st} = \beta_0 + \beta_1 corporateShare_{st} + \beta_2 Innovation + \beta_3 corporateShare_{st} \times Innovation_{st} + \alpha_s + \gamma_t + \epsilon_{st}$$
We here account for both fixed effects by state ($\alpha_s$) and year ($\gamma_t$). Now, the term we are interested in is the interaction term between corporate share and innovation. If share of corporate patents reduce the positive effect of innovation on social mobility, we would expect $\beta_3 < 0$.

$\textbf{Illustration:}$ We use simulated data to show how this regression would work, and illustrate the presentation of results. We include several stylized facts. There is likely more patenting in states with more corporations, due to higher resources. We model noise in mobility as log-normal, and assume the above linear model holds. Tangibly, we draw the innovation variable from a log-normal distribution, then draw corporate share as a linear function of innovation plus some noise, constrained to be between 0.05 and 0.95. Finally, we draw mobility using the above regression, adding log-normal noise, and using arbitrary coefficients. We make $\beta_3$ negative to capture the intuition that corporate patenting reduces the impact of innovation on social mobility. Suppose for simplicity that state- and time-fixed effects are zero (or the data is demeaned). We simulate the data for 100 "states", and split them into two categories, according to whether their corporate share is above or below the median. We then plot the relationship between innovation and mobility for both groups, essentially seeing whether for a fixed level of innovation, the mobility is lower for the high-corporate-share states. The simulated results are shown below.

[verachaudhry]

Industry and Cross-Industry Invention by Geography
Vera Chaudhry

Research Question
What are the effects of regional expertise on different types of inventions? What are the cross - invention effects in a specific region? This research question aims to build on Fact 3 of Akcigit et. al (2017) that studies how geography interacts with invention and innovation. For example, it’s possible that cities with industry specific expertise lead to better innovation in that industry, but not others. In another example, it’s possible that some industries have stronger cross effects than others (ie, perhaps in place x, inventions for industry a create better environments for inventions in industry b but not industry c, etc).

Research Design
To address this research question, I propose an extension to the model in Akcigit et. al (2017). First, the current model studies the likelihood of inventor migration, comparing inventors, high skill non-inventors and non-inventors. In the proposed research design, total inventor migration is compared to industry specific inventor migration. Results will show whether certain industries are more likely (or not) to migrate than others. Industry groups will be categorized by the type of patent each inventor holds. Second, the current model then studies characteristics of destination choice for inventors. These characteristics include urbanization, and deposits per capita. The summary statistics also evaluate patents per capita in a destination geography. This design proposes extending the model to evaluate the impact of industry patents per capita in a destination geography. The specific characteristics can be grouped into geographies with a high number of same-patents as the inventor (or non-same patents) or, additionally, have the % of patents per industry in a specific geography. If the proportion of industry specific patents was available per geography, it would be compelling to evaluate the cross-industry effects as well. (I.e., perhaps inventor with patent a benefits from moving to a destination geography with a high proportion of same patent a or non-same patent b but not non-same patent c).

Visualization

Note: Used BDS business formation data to evaluate new firms by industry per state (numbered on x axis).

Select example: New Firms in Retail Trade vs. New Firms in Healthcare

Note: Different industries have different proportions of new inventions in different states. Above, note that there are more new retail trade businesses in some midwest states that do not have new healthcare businesses. This may influence migration decisions, especially if other costs, like shipment, are cheaper in these alternative states.

[bheim]

Ben Heim
What’s the relationship between patent-level innovation and technology adoption?

In Akcigit’s “The Rise of American Ingenuity,” filed patents combined with census data is used to understand predictive qualities for innovation. Moreover, they find through using war-time funding as an instrumental variable that innovation (as measured by patents) was strongly predictive of growth.

When taken together with Agrawal, Gans, and Goldfarb’s Power and Prediction which delineates point innovations from systems innovations, there’s a question of how much adoption is generated by patent production and how that adoption shapes patents in turn.

I’d hypothesize the following relationship: early patents in a sector, long time gap between systems-level entrepreneurship that leads to mass adoption, and then a rush of new related patents after wide-spread adoption occurred.

Research Methods

To examine this, I chose electricity as a good GPT since Agrawal covers how it was adopted over time, reaching mass adoption by roughly 1930. Using rough numbers from Figure 1-1 in Power and Prediction and combining those figures with data from Google’s patente database where I selected the count of patents related to electricity from 1870 to 1930, the following graph is produced.

This chart shows that the gap between the early patents and eventual mass adoption was roughly 35 years for electricity. To put that into perspective, if we take Google’s “Attention is All You Need” paper to be the early innovation for LLMs, we wouldn’t expect to see market mass adoption until 2052 (of course, that’s a naive analogy).

What this chart more accurately shows is that a few early patents are excessively more impactful than later ones when it comes to electricity adoption. In fact, the successful implementation and adoption of a technology seems to cause a cascade of point and function innovations, which perhaps push adoption to the next level. Moreover, it seems that adoption (and perhaps GDP growth) fuels innovation in turn. As electricity grew in popularity, many electrical innovations follow.

When thinking about AI, the change that Agrawal is trying to get at can likely not be encapsulated so much by adoption (widespread adoption is far easier to achieve with a software than with electricity) but rather GDP growth. An important question to ask when it comes to electricity adoption rate is how important those later patents were in driving future adoption or if it was mostly the distribution of those initial innovations that led to mass adoption.

Sources (beyond readings):

Google Patent Public Database https://console.cloud.google.com/marketplace/product/google_patents_public_datasets/google-patents-public-data

[zararania]

Zara Rania Suran

Research Question:
Has the adoption of deep learning in drug discovery shifted pharmaceutical innovation from hypothesis-driven causal biology toward data-driven correlational pattern recognition, and how has this shift altered the nature and timing of clinical failure?

Motivation/Background
Traditional drug discovery is organized around biological hypotheses; a target is selected based on preexisting biological understanding then validated experimentally and advanced through clinical trials. In contrast, LeCun, Bengio, and Hinton (2015) argue that deep learning systems excel because they bypass handcrafted features/previously necessary domain expertise, instead learning directly from high-dimensional data. Applied to drug discovery, this could suggest that models may successfully predict promising compounds or targets without necessarily encoding causal biological understanding.

Research Design
To address this research question, we would investigate whether this epistemic shift in the nature of drug development changes how and when drugs fail. Drawing on the prediction-decision distinction emphasized by Kleinberg et al. (2015), the central hypothesis is that deep learning improves early-stage prediction (e.g., hit identification, target association) but weakens downstream decision quality when causal understanding is required. As a result, AI-discovered drugs might progress further into the pipeline but fail later, and more expensively, due to unanticipated toxicity or off-target effects. Our proposed empirical design would compare cohorts of drug candidates originating from (1) mechanism-first discovery programs and (2) deep-learning-first discovery programs. Outcomes of interest would include phase at failure, failure cause (efficacy vs. safety), and cumulative capital invested prior to termination.

Visualization

This figure is designed to illustrate how different discovery methods, or mechanism-first versus deep-learning-first, might shape the timing and cost profile of clinical attrition. By plotting trial phase against capital invested for individual programs, the graph tests the hypothesis that deep learning improves early-stage candidate selection while deferring uncertainty to later phases where causal biological understanding becomes critical. This simulated data predicts fewer early failures but a higher density of late-stage, high-capital failures for AI-originated drugs, alongside a steeper capital-by-phase gradient. In contrast, mechanism-first programs are expected to fail earlier and more cheaply, reflecting faster falsification driven by explicit biological hypotheses rather than pattern-based prediction.

[HenryPinner7]

Henry Pinner How does the composition and governance of international financing affect sustainable growth and structural transformation in a middle-income country, as illustrated by the case of the Philippines?

Research Design: This memo examines the Philippines as a case study of growth dynamics in a middle-income economy, focusing on how international financing structures shape development outcomes. The research design is qualitative-comparative and institutional in nature, drawing on sectoral evidence, financing flows, and governance arrangements rather than econometric identification. The central analytical move is to disaggregate “international finance” into distinct sources – foreign aid, private capital, multilateral development bank (MDB) lending, and bilateral policy-bank loans – and evaluate how each interacts with domestic institutions to influence long-run growth.

The analysis would proceed by comparing earlier waves of Philippine international financing to more recent reform-era efforts, including the 2022 FDI Amendments and the creation of the Maharlika Investment Fund. The goal is to assess whether newer financing arrangements, characterized by clearer regulatory frameworks and stronger institutional provisions, have been more successful at generating the types of institutional upgrading and structural change emphasized in the development literature. This design aligns closely with the 2024 World Development Report’s account of the middle-income trap, which argues that stagnation arises not from insufficient capital, but from weak institutional upgrading, limited learning/education, and failure to diversify into higher-productivity and innovative activities. Rather than asking whether financing increases output in the short run, the analysis asks whether financing arrangements have generated durable spillovers in human capital, technological learning, and administrative capacity.

The Philippines provides a revealing case. Despite decades of substantial inflows from diverse international sources, growth outcomes have been uneven and regionally concentrated, with persistent reliance on a narrow set of sectors and external partners. This makes the country a useful setting for evaluating whether recent shifts in the governance of international financing represent a meaningful break from past patterns, or risk reproducing the same middle-income trap dynamics under new institutional labels, particularly in light of emerging technologies such as AI.

Visualization:

Figure 1.

Figure 2.

Figures 1 and 2 motivate this memo’s focus on financing composition and governance rather than aggregate volume. Figure 1, from a 2013 ERIA Research Report, illustrates the sectoral distribution of international loan commitments to the Philippines, revealing a pronounced concentration in infrastructure relative to governance, institutional development, and productive sectors. While infrastructure investment can support growth, this allocation pattern raises questions about whether financing has been structured to promote diversification and institutional learning/innovation.

Figure 2, based from a 2022 PIDS Study, situates the Philippines within a regional comparison of FDI inflows and regulatory restrictiveness. Despite relatively stable inflows, the Philippines remains more restrictive than several regional peers, highlighting the key role of domestic regulatory institutions in shaping not only the quantity but the quality of international investment. By combining evidence on where international financing is allocated with evidence on how regulatory institutions shape investment patterns, the analysis is able to assess whether financing arrangements in the Philippines have supported sustainable growth pathways or reinforced existing middle-income trap dynamics.

[ukaishan]

Urjita Kaishan

Research Question:
Can ML-based prediction models targeting unemployment interventions reduce long-term unemployment by reallocating benefits and resources towards workers who have a higher predicted risk of prolonged unemployment, without having to determine how the benefits causally affect employment?

Background:
In this setup, unemployment is a prediction policy problem informed by Kleinberg et al. (2015), rather than a causal inference one. Counseling and help with job-search are unemployment interventions that are limited in administrative capacity and resource allocation, and the success of these programs depends on predicting which unemployed individuals will stay unemployed without the additional support and intervention.

This analysis will involve using labor market data regarding earnings histories, unemployment benefits filings, job tenure, prior employment transitions in each role. Using a prediction model to predict the probability of an unemployed worker remaining unemployed beyond a specified period (6 to 12 months), the bias-variance tradeoff will be addressed through a cross-validation method that chooses a model that predicts best. This will make sure that the chosen model will perform well outside of this given sample. The status quo allocation and prediction-based allocation results will be compared. The status quo allocation is the current state, where there are eligibility and timing constraints, and observed data will help inform this. The counterfactual prediction allocation will have the same budget, eligibility rules, but involve the reallocation within a fixed number of individuals who will receive intervention.
Those predicted risks will be used for a fixed portion of unemployment intervention benefits to be reassigned from individuals who have lower risk of long-term unemployment to those with high risk. The allocation efficiency will be determined by comparing aggregate outcomes under the status quo and prediction-model cases, specifically in average unemployment duration or the number of long-term unemployed individuals. This does not require estimating the causal effect on employment outcomes. The prediction policy problem setup makes sure that the design evaluates whether targeted assistance can reduce the duration of unemployment on the aggregate level. This study will address constraints in the welfare benefits system, including limited budgets and eligibility rules. Including these features makes the study more realistic and relevant to informing policymakers.
Visualization: reflects the methods for this setup

[TarunSethi13]

Tarun Sethi

Research Question:

Has the post-COVID expansion of remote work and digital collaboration reduced the geographic concentration of innovation in the United States, as measured by the spatial distribution of patenting across states?

Background:

Akcigit, Grigsby, and Nicholas (2017) show that American innovation has historically been highly geographically concentrated. States and cities with dense inventive activity consistently outperform others, not because of differences in innate ability, but because of structural advantages: access to capital, proximity to other inventors, strong institutions, and opportunities for collaboration. Their findings emphasize that invention is a social and institutional process rather than an isolated act of individual genius.

The post-COVID expansion of remote work and digital collaboration raises a natural question within this framework: have these long-standing geographic patterns weakened, or do the same agglomeration forces persist despite lower communication costs? If invention depends primarily on access to capital markets, dense collaborative networks, and institutional infrastructure—as Akcigit et al. argue—then reductions in physical proximity alone may be insufficient to alter the spatial distribution of patenting.

Data and Design

To explore this question, I use USPTO data on patents issued to U.S. residents by state for fiscal years 2021–2024. I measure geographic concentration using the share of total U.S. patents accounted for by the top 10 patenting states (California, Texas, New York, Massachusetts, Washington, Illinois, Michigan, New Jersey, Arizona, Pennsylvania) in each year. This provides a simple and transparent indicator of whether patenting activity has become more geographically diffuse in the post-COVID period.

Results

Top 10 states account for
2021: 66.36%
2022: 66.32%
2023: 66.62%
2024: 66.19%

Despite major changes in work arrangements and communication technologies, roughly two-thirds of all U.S. patents continue to be produced in just ten states, with no evidence of declining concentration.

Interpretation

These results align closely with Akcigit et al.’s argument that geography matters for innovation because of complementary inputs to invention, not merely because of face-to-face interaction. States that dominate patenting tend to host deep capital markets, research universities, legal and regulatory expertise, and dense inventor networks that facilitate collaboration and knowledge recombination. Even if inventors can now collaborate remotely, access to these complements remains unevenly distributed across space.

From this perspective, the persistence of geographic concentration after COVID suggests that innovation remains embedded in institutional ecosystems. Lower communication costs may change how collaboration occurs, but they do not eliminate the underlying advantages of locations where capital, talent, and inventive networks co-locate.

At the same time, patenting outcomes are inherently lagged, reflecting multi-year research, funding, and commercialization processes. As a result, a longer time horizon and more granular data will be necessary to assess whether post-COVID work arrangements ultimately reshape the geography of innovation or whether these agglomeration patterns continue to persist.

[felicechan]

Research Question
Does artificial intelligence help low- and middle-income countries overcome the middle-income trap, or does AI adoption under contemporary capitalism reproduce colonial patterns of technological dependence and limited innovation?

Background and Motivation
Recent advances in deep learning have dramatically improved prediction by enabling scalable representation learning (LeCun, Bengio, and Hinton 2015). Economically, this represents a sharp decline in the cost of prediction across a wide range of tasks, potentially raising productivity in sectors from manufacturing to services. However, as Kleinberg et al. (2015) emphasize, prediction alone does not generate welfare; outcomes depend on how predictions are embedded in decision-making systems.
Agrawal, Gans, and Goldfarb (2022) argue that when prediction becomes cheap, the binding constraint shifts to complements: data, organizational redesign, and authority over decisions. These complements are unevenly distributed under capitalism. Firms and countries that control data, infrastructure, and decision rights are positioned to capture the gains from AI, while others may remain users rather than innovators. This distinction is particularly apparent for LMICs, which often access AI through global platforms rather than developing domestic capabilities.
Historical experience suggests that growth depends on more than access to technology. Akcigit, Grigsby, and Nicholas (2017) show that U.S. growth during the late nineteenth and early twentieth centuries relied on broad participation in innovation, geographic mobility, and relatively low barriers to entry. Capitalism functioned as a mechanism for reallocating talent toward inventive activity, producing diffuse innovation and upward social mobility.
Many LMICs, by contrast, entered global capitalism through colonial systems designed for extraction rather than innovation. Colonization constrained economic freedom, limited access to education and capital, and suppressed social mobility, producing institutional environments poorly suited for domestic inventive activity. Even after decolonization, these legacies persisted, shaping dependence on imported technologies and weak innovation ecosystems. The World Development Report 2024’s emphasis on stalled transitions from technological infusion to innovation can be understood through this historical lens.
Generative AI intensifies this tension. While it enables rapid adoption of frontier capabilities, platform-mediated access may replicate colonial patterns of technology use without ownership, learning, or control, raising the risk of innovation without autonomy.

Research Design
The project uses a cross-country panel design to test whether AI adoption in low- and middle-income countries is associated with expanding domestic innovative capacity or continued technological dependence. AI diffusion is measured using patent data, capturing both the share and technological breadth of AI-related inventions.
Domestic innovation outcomes are measured using indicators such as new firm formation, inventor entry, and growth in non-AI patenting, which proxy for whether AI adoption coincides with expanding local innovation rather than substitution. To account for capitalist structure and historical institutions, I compare formerly colonized and non-colonized countries and interact AI diffusion with measures of economic freedom, such as competition intensity or barriers to entry.
The core test estimates whether increases in AI diffusion predict subsequent changes in innovative activity across these contexts. The main visualization plots AI diffusion against changes in domestic innovation, with separate trend lines for colonized and non-colonized countries, indicating whether AI supports transition from infusion to innovation or reproduces technological dependence.

Name: Felice Chan

[johnab20]

John Butka

Research Question: To what extent does the scale of human capital investment in data science (analyst headcount) drive organizational growth efficiency, and at what point does "Innovation Saturation" lead to diminishing returns in Major League Baseball?

Research Design: In Power and Prediction, Agrawal et al. argue that the most profound economic shifts occur when AI enables a redesign of an entire "system." In baseball, "Point-Adopters" use data to solve isolated problems, while "System-Innovators" use AI as a core engine to synchronize scouting, development, and tactics. My research design tests this by performing a regression analysis of all 30 MLB teams, using Analyst Headcount as a proxy for Innovation Input and Wins per $10M of Payroll as the metric for Growth Efficiency. This methodology seeks to validate whether a "Middle-Income Trap" exists in sports, where teams spend heavily on data staff and player salaries but fail to achieve the elite efficiency of leaner, system-integrated organizations.

Using 2024 empirical data, the design utilizes a linear regression model to map the correlation between these variables. The presence of a negative slope (although slight) in the regression line suggests a paradox: as analyst counts increase beyond a certain threshold, organizational efficiency often declines. This allows for a critical qualitative interpretation of "Innovation Saturation." For instance, the Tampa Bay Rays (42 analysts, high efficiency) serve as a model for "System Innovation" where scale enhances the redesign, whereas teams with high analyst counts but lower efficiency scores may be bogged down by "point-solution overload."

By plotting the 95% confidence interval (the shaded red region), the design identifies outliers that successfully break the league-wide trend. Following the framework of Kleinberg et al. on "Prediction Policy," I will assess if these outlier teams are better at decoupling raw data predictions from roster-management decisions. This study ultimately validates whether the value of AI is found in the volume of analysts or the integration of those analysts into a cohesive competitive strategy that avoids the stagnation typical of high-capital legacy models.

Note: I understand that the efficiency metric is flawed, but given the limited time it will suffice, it heavily favors the teams that don’t spend (which I strongly disagree with as a fan). The point still stands to where the underfunded teams have to rely on the data systematically as to where the well funded teams can just out spend and can afford to pay players more for less statistical output.

[Rjlee22]

Question: Did Chile’s dependence on copper exports and limited downstream industrial upgrading contribute to its persistence in the middle-income trap following market liberalization?

Background and Motivation:
This project investigates the intriguing case of Chile, which stands out among South American nations for its robust macroeconomic policies and early embrace of market liberalization. Despite these foundational strengths including institutional stability and economic openness, Chile has experienced prolonged stagnation in productivity growth when compared to other advanced economies. The nation's continued, heavy reliance on copper exports raises a critical question: did this export concentration and a lack of economic upgrading create a barrier, preventing its complete transition to high-income status?

My proposed research design draws directly from the argument that commodity dependence can impose a "structural ceiling" on long-term growth. This ceiling arises because it limits key growth mechanisms, such as learning-by-doing, the diffusion of technological spillovers, and the necessary diversification into higher value-added sectors.
Rather than covering topics like institutional quality or educational attainment, this design centers on industrial structure as the primary explanatory variable. The analysis will compare Chile’s post-liberalization economic path with what could have happened if its substantial copper revenues had been more effectively channeled into downstream manufacturing and knowledge-intensive services.

Research Design:
Empirically, this project will utilize a Synthetic Control Method. We will construct a "synthetic Chile" using a pool of comparable middle-income countries that shared similar characteristics in the early 1980s (initial income, openness, and institutional quality), but crucially, possessed more diversified export baskets (e.g., Malaysia, Thailand). The "treatment" in this analysis is defined as Chile's sustained reliance on copper extraction rather than significant diversification into downstream processing or other manufacturing. Key outcomes we will track include labor productivity growth, changes in export complexity indices, and GDP per capita convergence relative to the United States.
To motivate the design, Figure 1 presents a conceptual diagram. The figure illustrates how copper export dominance crowds out downstream upgrading by dampening incentives for firm-level innovation and skill accumulation

This project shows that even countries with strong institutions may stall when their growth model remains structurally narrow. Chile’s experience highlights the importance of export composition and industrial upgrading, not just market liberalization, in sustaining long-run development.

[CristianG-png]

How, in particular, can AI-driven productivity offset the economic drag of demographic aging in middle- to low-income countries that are aging out of work instead of economically flourishing?

While a concern in the US is that AI is taking over too many jobs, “World Development Report 2024: The Middle-Income Trap” shows that such a fear can actually be a desire for lower-income countries, where the working population is going to gradually decrease. As the primary target demographic for this initiative is lower-income countries, which do not currently face low birth rates, they still need to fulfill the roles of the aged that the ever-shrinking aging population cannot. Figure B3.1.1 from the reading exemplifies this concern with the indication that middle-to-lower income countries are “aging” quickly. This could reach the point where the retired work population (noted to be 12% of the population) far exceeds the number of aging future workers.

A proposed solution to this would be the addition of AI to these lower-income countries. AI acts as a substitute for the shrinking labor supply by lifting total factor productivity (TFP), allowing economies to grow even as the workforce weakens. Additionally, the World Development Report 2024 finds that migration is one of the most powerful tools aging economies have to expand labor supply and sustain growth. As the child of an immigrant family, I thought incorporating migration would be an interesting avenue to take.
My project would focus on a three-path control-style simulation, inspired by the 3i model, which compares how GDP per worker evolves under different combinations of demographic pressure, AI adoption, and migration inflows. One would be the baseline aging path, represented by Figure B3.1.1. Another would be a prediction-driven AI-only path, which would show increased TFP growth by a significant amount. The final one would include migration into the AI path, with the same productive increases from AI, while also considering how migration inflows work to also offset the shrinking labor force. The conclusion would be that while AI raises productivity, migration can stabilize labor supply, and together they generate a growth path that neither mechanism can achieve alone.

To predict the effect of AI, I’d cite long‑run macroeconomic projections. One suggests that generative AI could raise productivity by 1.5–3.7% over the next fifty years, depending on adoption speed (Penn Wharton Budget Model, 2025). As for evidence of how migration affects economic growth, the United States is an incredible source, as it has gained a reputation for being the place where people from these low-income countries (such as my parents) see fruitful economic opportunities. Extensive evidence shows that migration expands labor supply, raises innovation, and increases GDP. In the U.S., immigrants contribute roughly $2 trillion to GDP annually, and similar mechanisms apply to aging middle‑income countries.

Cristian Garcia

[csardinas]

Research Question:

Does AI adoption create an insurmountable development barrier where only upper-middle-income countries with pre-existing digital infrastructure can leverage AI for economic growth, while lower-middle-income countries remain permanently trapped regardless of current policy interventions?

Theoretical Foundation:

Akcigit et al. (2017) documented that innovation clustered in regions with existing infrastructure and financial development. Their Fact 7 and Figure 5B revealed strong correlations between patents per capita and deposits per capita, demonstrating that innovation required prerequisite institutional capacity. Inventors migrated toward regions already equipped with financial and industrial infrastructure, creating self-reinforcing geographic concentration.

I hypothesize AI adoption amplifies this pattern at a global scale. Unlike historical innovation requiring primarily financial capital, AI demands a convergent set of prerequisites: digital infrastructure, technical human capital, and critically, recent innovation stocks to build upon. Only upper-MICs possess this combination, creating a threshold effect that locks out lower-MICs.

Research Design:

I construct a Digital Readiness Index using World Bank and ITU data measuring: internet penetration rates, electricity access coverage, tertiary STEM enrollment ratios, and existing patent stocks per capita. This index spans 2010-2024 for all middle-income countries, classified by World Bank income thresholds into lower-MICs (GDP per capita $1,086-$4,255) and upper-MICs ($4,256-$13,205).

AI adoption is measured through three proxy indicators: AI-related patent applications filed with WIPO, venture capital investment flows into AI startups from Crunchbase, and enterprise cloud computing expenditure as a proxy for AI service adoption. These metrics capture both innovation production and commercial deployment. The empirical test examines whether 2015 digital readiness levels predict 2020-2024 AI adoption growth, and crucially, whether this relationship differs systematically between lower-MICs and upper-MICs.

Regression specification::

AI_Adoption_Growth = β₀ + β₁(Digital_Readiness_2015) + β₂(Upper_MIC) +
β₃(Digital_Readiness_2015 × Upper_MIC) + ε

The critical coefficient is β₃. If positive and significant, it demonstrates upper-MICs extract disproportionate returns from digital readiness investments. An upper-MIC at the 70th percentile of digital readiness achieves fundamentally different AI outcomes than a lower-MIC at identical readiness levels, proving the barrier operates through country income classification itself, not merely infrastructure.

Visualizations:

Expected Findings and Implications:

Preliminary analysis reveals β₃ is positive, large, and highly significant on F-tests. The interaction term dominates the model, indicating upper-MIC status amplifies digital readiness effects by factors of 3-5x. Countries marginally below the readiness threshold in 2015 cannot catch up through contemporary investments because AI innovation demands building upon recent innovation ecosystems they lack.

This creates a development trap analogous to Akcigit et. al's geographic concentration but more severe. In their historical analysis, regions without financial development failed to attract inventors despite potential talent availability. Similarly, lower-MICs now face permanent exclusion from AI-driven growth trajectories. They missed the critical 2000-2015 digital infrastructure investment window. Current attempts to build AI capacity fail because AI requires compounding upon existing innovation stocks, computational infrastructure, and technical ecosystems that take decades to develop.

[jgoberman]

Julia Goberman

Research Question: Will AI solve the historical barrier to innovation (education/access) or create new barriers that are more exclusionary?

Akcigit et al. (2017) paper The Rise of American Ingenuity: Innovation and Inventors of the Golden Age document that during America’s innovation golden age (1880-1940), father’s income strongly predicted who became an inventor but not their productivity. Children of top 5% earners were 2.5 times more likely to patent, yet once individuals became inventors, parental wealth didn’t correlate with patent quality or quantity (Tables 2-3). With education access, college graduates were 4 times more likely to invent, and controlling for education eliminated the father income effect in Table 2. This distinction between who can attempt innovation and who succeeds may provide a lens for evaluating AI’s impact on future innovation.

AI appears to solve the access problem Akcigit identifies. If education was the barrier and AI provides educational capabilities, we might expect a more democratized innovation. However, this view assumes resource requirements for productivity remain constant. Unlike the 1880-1940 era used in the paper, where education was necessary but additional resources did not improve inventor productivity (Table 3), AI may create new barriers as to who succeeds. Computational infrastructure costs, access to proprietary datasets, and network effects in AI capability development. Training frontier models costs $100M+, fine-tuning requires specialized expertise, and productive AI use depends on knowledge about prompt engineering and model selection.

This results in a shift in that more people can attempt innovation while fewer can achieve transformative breakthroughs. AI may couple these intensive and extensive margins, requiring both access and resources for success.

This framework suggests current policy focus on AI literacy and tool access address the wrong margin. If the historical lesson is that education access mattered but post-education resources didn't, and AI inverts this pattern, we need policies targeting the intensive margin: public compute infrastructure, data commons, and network access programs. We need empirical work replicating Akcigit's Census to patent matching methodology for the AI era, separately measuring how AI affects who attempts innovation versus who succeeds. I would like to continue to explore how the barriers to innovation have continued to shift through the rise of AI.

Research Design:
Extending Akcigit's model (equations 5, 7), I decompose innovation costs into entry (cᵢᴱ) and productivity (cᵢᴾ) components. While Akcigit shows cᵢᴱ varied by background but cᵢᴾ = 0, AI inverts this: cᵢᴱ becomes negligible but cᵢᴾ = g(Resources) emerges. This generates testable predictions: in regressions parallel to their Tables 2-3, $\beta$(Father_Income) should approach zero in extensive margin regressions but become positive in intensive margin specifications.

Akcigit et al. (2017) baseline: Individuals become inventors when $\eta V_t - c_i Q_t \geq w_t$ where $V_t = \pi_t/(\rho + \tau)$ (equation 7). Their finding: $c_i$ varies with father's income but $\pi_t$ does not.

AI era extension: Decompose $c_i$ into entry cost $c_i^E$ and productivity cost $c_i^P$.
Stage 1 Extensive: Attempt if $\eta \cdot \mathbb{E}[\pi_i]/(\rho + \tau) - c_i^E Q_t \geq w_t$
Stage 2 Intensive: Succeed if $\theta_i - c_i^P \geq 0$

I look to formalize this as a two-stage model where AI reduces entry costs $ciEc_i^E ciE​$ but introduces productivity costs $ciPc_i^P ciP$​ that vary by resource access.