Week 1: Memos - Introduction to Artificial Intelligence, Innovation, & Growth #3
Comments
jamesallenevans
changed the title
|
Middle Income Trap in Brazil (Pre and Post Global Financial Crisis)
The 2008 Global Financial Crisis (GFC) and the subsequent transition to green energy in middle-income countries, like Brazil, are directly linked to efforts to close the gap between middle-income countries and high-income countries, as described in the Middle-Income Trap. The GFC created a unique opportunity for Brazil to accelerate its transition to renewable energy, thereby addressing several challenges that middle-income countries face in the pursuit of sustainable development, innovation, and economic competitiveness. Middle-income countries are often characterized by heavy reliance on specific sectors (like agriculture, manufacturing, or fossil fuels) that can limit their economic growth and innovation potential. For Brazil, the GFC was a critical moment to diversify its energy sector and reduce reliance on fossil fuels. This shift was crucial in addressing a key challenge for middle-income countries: how to transition from an economic structure that depends on limited resources (like oil and gas) to one that promotes sustainable industries and technologies. By increasing the use of biogasoline and biodiesel (as shown in the graph), Brazil shifted its energy mix towards more sustainable, renewable sources, helping to insulate its economy from the volatility of global oil prices—a common problem for many middle-income countries. Brazil’s investments in biofuels were driven by government policies that were designed to create new industries and technological advancements, reducing the middle-income trap by promoting energy self-sufficiency, innovation, and high-value-added industries. Brazil's focus on developing green energy technologies like ethanol and biodiesel helped stimulate innovation within its energy sector. As Brazil’s biofuel industry grew, it became more competitive in the global market. The technological advances made in renewable energy directly contributed to Brazil’s ability to generate higher value-added products, allowing it to increase exports and strengthen its position in the global economy. The GFC provided an opportunity for the government to invest in and incentivize green energy projects. This partnership between the public sector (through government policies and regulations) and the private sector (through investments in renewable energy) helped reduce the financial risks associated with large-scale renewable energy projects and facilitated the scaling of new technologies. This innovation fostered the development of related industries such as agricultural technology, infrastructure for biofuel production, and advanced manufacturing. Brazil’s experience with biofuels and green energy investments post-GFC highlights a broader lesson for middle-income countries seeking to overcome the middle-income trap: innovation, diversification, and sustainability are key to long-term growth and development. By focusing on biofuels, Brazil not only met its energy needs but also strengthened its global competitiveness, increased its export potential, and contributed to global sustainability efforts. |
|
Overcoming the Middle Income Trap in India: The Saagu Baagu Initiative Agriculture presents a significant opportunity for middle-income countries to leverage AI for innovation. In many of these nations, agriculture remains a primary source of employment but is often hindered by inefficiencies and overreliance on traditional practices. AI-driven tools can revolutionize the sector, enhancing productivity, reducing waste, and promoting economic diversification. A compelling case is India’s "Saagu Baagu" initiative, a flagship program under the AI for Agriculture Innovation (AI4AI) framework. This project focuses on chili farmers in Telangana, using AI to optimize every stage of the agricultural process. The initiative employs predictive analytics to monitor crop health, helping farmers detect diseases early and reduce pesticide use by up to 30%. Additionally, AI-powered irrigation schedules have cut water usage by 25%, ensuring sustainable resource management. Real-time market intelligence enables farmers to make data-driven decisions about when and where to sell their produce, reducing reliance on middlemen and ensuring higher profit margins. A key component was also a Telugu-language chatbot providing farmers with timely suggestions tailored to the maturity stages of their crops. These interventions have collectively doubled the income of over 7,000 participating farmers. Beyond immediate benefits, Saagu Baagu is scaling its impact to over 500,000 farmers across value chains such as cotton, turmeric, and dairy. By fostering innovation and reducing inefficiencies, the initiative positions agriculture as a driver of economic growth and diversification. This aligns with strategies to escape the middle-income trap. AI-driven improvements in agricultural productivity not only enhance food security but also reduce dependency on volatile global commodity markets, contributing to economic stability. The figure below showcases the growing market size for AI in precision farming, reflecting its potential for economies like India. The data highlights a compound annual growth rate of 20% by 2028, demonstrating the rapid adoption of AI in agriculture globally. This trend demonstrates the immense opportunity for middle-income countries to leverage emerging technologies and initiatives like Saagu Baagu to boost productivity and economic growth. Scaling such initiatives requires robust support systems, including investments in digital infrastructure, farmer training, and accessible financing. Public-private partnerships are crucial to reducing barriers to entry and ensuring equitable technology access. Moreover, governments must establish regulatory frameworks that encourage innovation while safeguarding smallholder farmers from potential exploitation. The Saagu Baagu initiative illustrates how AI can transform agriculture into a high-productivity sector, driving innovation and economic growth. By addressing key challenges of the middle-income trap, such initiatives enable countries like India to foster technological advancement and create pathways to high-income status. 
|
|
Improper Sampling for ML Training Sets for Policy Research In Prediction Policy Problems, Kleinberg et al. argue in favor of machine learning-based approaches over OLS approaches for conducting predictive analyses within economic policy research. Machine learning approaches include a “regularizer” term that punishes functions with high variance, improving sensitivity against in-sample overfitting and reducing noise to make better predictions for out-of-sample data. Kleinberg et al. recognize that the strength of this method enables researchers to make predictions on the basis of an extremely large number of variables for even a relatively small dataset. However, I see a significant risk from sampling when using this method for policy research. When conducting policy research, the training set for the machine learning model must be part of the relevant sample––like how Kleinberg et al. use a portion of the 20% sample drawn from Medicare beneficiaries. A 20% random sample likely provides a sufficiently representative group for a few observable variables (e.g. comorbidities, doctor visits, age). The sampling method and the number of variables, however, plays a considerable role in the predictive capacity of the model. Imagine instead that we incorporate 50 different variables from this random sample for increased accuracy within our predictive model. Given the increase in variables, there is also a greater probability for significant correlation between data points for at least one of the variables by chance. For these samples, the regularizer term would reward low variance-inducing functions that prioritize this correlation over the other variables. Furthermore, two different 20% samples may be relatively similar in their distributions for a few variables, but these samples could separately show high correlation for different variables when 50 variables are considered––potentially by chance. Likewise, models trained separately on these two samples may return different predictions for the same input. In general, when using higher-dimensional data to train machine-learning models, we risk providing a non-representative sample that becomes the predictive backbone of the model. We might think of remedying this issue by curating a well-diversified training dataset from the data universe we are sampling. However, we are then introducing bias by not randomizing our sample. Furthermore, as the number of variables we want to use for prediction increases, the more difficult it is to stratify them for better sampling, which is an approach often used to better train machine learning models. We theoretically always face an increased risk of accidental overrepresentation for any one or more variables. The following image depicts a toy model for sampling methods based on variable count:
A: We are training our model on data with only 1 variable (shape). We could employ stratified sampling if the distribution is particularly skewed, or take a standard random sample which we further divide into a 70/30 training/testing split. This is often a first approach for low-complexity data. Ultimately, the cost of increasing the number of predictive variables may eventually exceed the gains in the accuracy of the machine-learning prediction, which also implies that there is an optimal range of number of variables to use for each dataset. For economic policy research, these concerns imply that a machine-learning model may misconstrue predictions for out-of-sample data unless configured optimally, and if interpreted incorrectly, may prove harmful for policy implementation. |
|
Harnessing AI to Drive Creation and Destruction Artificial Intelligence is a powerful force that can directly impact creation and destruction, the two forces that MICs need to prioritize. AI has the potential to create entirely new industries and jobs, such as in fields like data analytics, robotics, and AI software development. At the same time, it can destroy traditional, less efficient jobs in industries such as manufacturing and logistics through automation. While this dual impact may seem disruptive, it is this process of creative destruction that can drive innovation and productivity gains. To model the economic impact of AI-driven creative destruction, I propose a simple equation: I found and read through a McKinsey report that touches on these ideas that is titled, Notes from the AI Frontier, which gives us some valuable data points and insights on the parameters of the equation. According to the report, AI could contribute $13 trillion to global GDP by 2030, with an average annual growth boost of 1.2%. However, this impact varies across countries, with MICs facing challenges due to skill gaps and lower AI adoption rates. The report estimates that 15–20% of global labor could be automated by 2030, disproportionately affecting low-productivity sectors. Although this equation clearly would require more extensive research and data collection to quantify empirically, I think it could provide a useful framework for thinking about how MICs can strategically leverage AI. By focusing on increasing |
Visualizing the Bias-Variance Tradeoff in the Context of PredictionsPredictions have significant decision-making implications. For example, as Kleinberg et. al. (2015) discuss, osteoarthritis surgeries provide a better quality of life for the remainder of a person's life, but at a high cost both in terms of money and disutility while recovering. Thus, these surgeries are only worthwhile if a person will live long enough, which is an inherent prediction problem (i.e., predicting how long they will live after surgery). Kleinberg et. al. (2015) mentions how standard statistical techniques (i.e., ordinary least squares methods) are suboptimal for prediction problems because they prioritize minimizing bias. Since there is generally a bias-variance tradeoff, these methods often have higher variances. In these cases, even though the fitted model may have the lowest bias, it may have poor out-of-sample prediction performance. Regularization techniques, on the other hand, allow machine learning models not just to minimize bias but also to take into account variance. Regularization does this by imposing some sort of penalty for variability in the objective function. The Lasso function, for example, imposes a penalty on high values for the coefficients since larger coefficients allow for more variability when predicting. Thus, these techniques boast better predictive capabilities by actually allowing the model to make an explicit tradeoff between bias and variance, rather than just forcing it to minimize bias. I wanted to empirically test this theory, and see if a polynomial regression with regularization can outperform a polynomial regression that doesn't regularize. To that end, I constructed a Jupyter Notebook to fit two different regression models on fairly noisy data that is modeled by the line The two models I used were ordinary least squares (OLS) and Lasso with a high penalty for increased variance. In doing so, I could get the two ends of the spectrum, with the OLS model minimizing bias and allowing high variance while Lasso minimizes variance and allows more bias. After fitting both models on the training set, I was able to compute the mean squared error for both models on the testing set and generate the following graph: As evidenced by the lower MSE, the Lasso regression was better at predicting out-of-sample than the OLS regression, which supports Kleinberg et. al.'s theory that minimizing bias leads to problems in prediction. Interestingly, on the right-hand side of the graph, where the sampled data has more noise, the Lasso model seems to be closer to the true model, which is likely because it prioritizes reducing variance. Thus, these results underscore the significant contributions the field of machine learning can bring to policymaking. Although traditional statistical methods like OLS are still powerful tools for policymaking, using machine learning to make predictions offers an avenue for better policy. |
On the Use of Patents as a Proxy for Innovation“The Rise of American Ingenuity: Innovation and Inventors of the Golden Age” primarily concerns itself with the time period from 1880 to 1940. During this period, innovation was closely tied to patentable advancements like the electric lamp and chocolate ice cream. However, in today’s technological landscape, the proliferation of open-source projects and other non-patentable innovations challenges the relevance of patents as a proxy for innovation. I argue that patents were an excellent measure of innovation historically, but may no longer be able to fully capture the dynamics of innovation today. The deterioration of patents as a proxy for innovation is principally due to the advent of software as an industry and the new standards/practices brought with it. This new industry brought with it new philosophies (Stallman’s free software movement), traits (ease of distribution), and economic incentives (network effects) to adopt wildly different policies from existing industries. Open-source as a movement has decoupled significant portions of modern innovation from the patent system. For instance, technologies like Linux, widely used in enterprise and consumer systems, are not patent-protected but nevertheless represent major innovations that provide real value to its users. The change in distribution of inventors may also pose a threat to patents as a proxy. Where Have All the "Creative Talents" Gone? Employment Dynamics of US Inventors finds that an increasing number of inventors favour larger incumbents over entrants. Assuming larger firms are more likely to engage in rent seeking behaviour, the migration of inventors from entrants to incumbents can not only affect the quantity of patents but can alter their quality. Inventors may be incentivized to use patents strategically as defensive manoeuvres against litigation or to provide hindrance to competitors. Provided such behaviour is widespread, this will undermine patents’ usefulness as a proxy for real innovation. That patents remain a valuable tool for measuring innovation and their relevance in areas such as pharmaceuticals and hardware is uncontested. However, their effectiveness as a comprehensive measure of innovation is diminishing in an era dominated by software advances. To capture the true extent of modern innovation, metrics must evolve to include alternative measures, such as contributions to open-source projects, software developments, and collaborative research outputs. Platforms like GitHub offer a new way to track innovation. Metrics such as repository creation and contributions could supplement patent data to provide a better measure of innovation in the modern world. |
Predicting the success of future innovatorsIn line with The Rise of American Ingenuity's attempt at understanding the demographics and outcomes of America's innovators in the 19th and 20th century, I'm curious to understand if a similar empirical/formal data analysis can be performed on current founders to understand potential indicators of future success (indicators that could help make people money now, instead of back then). There seem to have been attempts to do so using regression models. For example, this paper claims to "predict venture success to a sufficient extent" (p. 35), but I disagree with its definition of success as I think it is premature -- (paraphrasing) raising a post-seed investment within 5 years of founding. Unfortunately, in my quick reading of the paper, it's unclear how the independent variables were produced -- if the author decided which to include, or if another source (like the database they got the demographic information from) was the default. Others, like this, have used more simple graphs/math to analyze demographic breakdowns of successful founders, but I see a few things wrong in this specific analysis:
My thought here is that (a) there are likely a lot of independent variables to consider in a founder profile (and including too many IVs might take away from the merit of our regression model), (b) some may be soft skills that are difficult to measure (or impossible if you think evaluation is subjective) / collect data on, and (c) there are many other factors besides a founder's personal demographic that could affect outcome (such as industry, type of customer -- B2B, B2C). We might be able to resolve (c) by looking at successful industries or customers like here, but (b) seems slightly harder to resolve.
I'd like to explicitly acknowledge the, perhaps, shallowness of my research into existing hypotheses/reports on the topic. However, this specific issue/possibility has long been unresolved to me. I did not know if they had merit, nor did I have the time and focus to think critically about them. What this albeit brief research has done, however, is allowed me to form some conclusion on the matter, which is the following: I think there's definitely some merit in looking at demographic information to identify good founders -- for data like education and job experience, it's a form of signaling that can be valuable. But I'm less confident in the extent that a regression model can be made to predict founder success, and I largely believe readily apparent information should be a preliminary screening, and by no means a determinant one -- there are more things to consider, both about the founder and not. |
Innovation Efficiency as Products/InvestmentsA prominent theme from this week’s readings and lectures was the dynamic role of equality in innovation. While equal opportunity is a critical enabler of innovation, a certain degree of income inequality across an economy is necessary to incentivize individuals to pursue groundbreaking ideas. For instance, The Rise of American Ingenuity: Innovation and Inventors of the Golden Age identifies white males aged 36–55 as the primary demographic of inventors during 1880–1940, based on census data. However, given the societal changes since then, I would expect more recent census data to reflect a broader and more diverse pool of inventors. The discussion of talent-task allocation in the World Development Report (WDR) underscores the importance of reducing discrimination to foster innovation. According to the report, “in the United States between 1960 and 2010, the decline in gender and racial discrimination in education and work explains up to 40% of the observed growth during that period” (Page 14). Moreover, the report critiques income inequality as a “relatively superficial measure” compared to deeper, underlying forces like socioeconomic mobility, which better indicate an economy’s potential for equitable innovation. Moreover, there is a noteworthy distinction between innovation in theory and innovation in practice. While the idea of developing technological breakthroughs and cutting-edge inventions is universally appealing, the reality of carrying them out is far more complex. In addition to patents as a measure of innovation, I think it is useful to consider other factors that paint a more holistic picture. For example, R&D spending can serve as an indicator for the value of innovation. When focusing solely on end-products of innovation, such as patents, we can overlook the individuals who try to innovate but do not succeed — sometimes simply because of poor timing or luck. In other words, focusing only on the “winners" paints an incomplete picture. To provide a more holistic measure of innovation, I came up with a simple equation: 
Where: Additional considerations: Time adjustment since effective investments do not yield end products instantaneously, considering intangibles like knowledge that catalyze innovations |
|
On the Topic of Protecting Equality of Opportunity to Ensure Growth 
Chapter 5: “Preservation” in the World Bank report on the middle income trap raises an interesting point regarding the deleterious effects of discrimination on economic growth. In general, we cannot have a fully productive labor force with discrimination against particular social groups, with the most notable example being the exclusion of women from the workforce in many middle income countries. The article also makes a point talking about in-group and race based discrimination. Essentially, discrimination makes it so qualified outsiders are not rewarded for their merit, which perpetuates inequality and hampers economic growth. This is intuitive since a lot of growth later in a country's development relies on innovation, which requires efficient utilization of the country's talent pool. By excluding a proportion of the potential talent, we waste a lot of potential productivity. This is not unique to middle income countries, as the article talks about historical redlining of African-Americans as an example of discrimination in high-income countries. As the article mentions, the Fair Housing Act of 1968 outlawed housing discrimination, among other efforts to eliminate inequality in law during the Civil Rights Era of American history. The effects of this discrimination are still felt today however. As the attached figure (Reardon, Fox and Townsend 2015) shows, two households of the same income level but of different races tend to live in neighborhoods of differing median income levels in the US. Namely, even if a black household is to make as much as a white household, they are still more likely to be in a poorer neighborhood and have difficulties with social mobility, since it usually entails worse opportunities in things like education. Since the article highlights how equality of opportunity and abolishment of discrimination would support greater economic growth in middle income countries, it would not be unreasonable to assume that combatting discrimination in higher income countries could also provide better economic growth. Additionally, accounting for the in-group bias that the article mentions, passing policy outlawing discrimination may not be sufficient. As with the US, any country with a history of discrimination will usually end up with insiders and outsiders. Even if discrimination is outlawed on paper, historically disadvantaged outsider communities are still disadvantaged, and may still struggle to attain social mobility without the connections an insider would have. This means that the effort to combat discrimination and reward merit irrespective of race and gender must extend beyond just declaring equality in law. Such efforts would likely benefit middle and high income countries, as inequality across demographic groups still persists in many high income countries that have codified equality into law. For middle income countries that have yet to enshrine gender and racial equality into law, or are still in the process of doing so, there is somewhat of an "advantage in backwardness" as they can learn from the shortcomings of frontier countries like the US that did not pay sufficient attention to ensuring that equality of opportunity was a de facto construct rather than just a de jure construct. |
How Did Malaysia Get Out of The Middle Income Trap?As a Malaysian, I would like to examine Malaysia's recent advances and how it has begun to escape the middle income trap, despite being a low-income country just a few decades ago. Malaysia’s economic path shows how a country can work its way out of the middle-income trap, which often slows progress to high-income status. While debates persist about whether Malaysia remains in this trap (for example, this website implies that Malaysia is not stuck in the middle income trap), the evidence points to consistent progress supported by targeted policies and strategic planning. Malaysia’s GNI per capita reached US$11,970 in 2023, just below the high-income threshold of US$13,845. The World Bank predicts Malaysia could surpass this mark by 2028. The CEPR highlights Malaysia’s convergence with high-income nations, with GDP per capita less than 7% below Greece’s. There is also sustained annual growth of 3% per capita, which shows strong economic growth. Malaysia’s efforts to escape the middle-income trap have centered on targeted policies and strategic investments, which we talked about in class regarding infusion and innovation. One of the most impactful programs, the Economic Transformation Programme (ETP), launched in 2010, focused on urban public transport improvements, such as the Mass Rapid Transit (MRT) system, which increased economic efficiency. Beyond transport, Malaysia also prioritized high-value industries like semiconductor manufacturing, with plans to produce GPUs and chips within the next decade, with the high demand of AI and chips. This initiative is tied to the development of the Batu Kawan Industrial Park, which has attracted major firms like Intel and Micron. Cross-border collaborations with Singapore also played a key role. For example, the Johor Special Economic Zone (SEZ), developed with Singapore, aims to streamline logistics and manufacturing to support trade. To boost innovation, Malaysia set an ambitious goal to increase its Gross Expenditure on R&D (GERD) to 2.5% of GDP by 2025. Initiatives like Technology Park Malaysia (TPM) have been crucial in supporting tech startups between universities and industries. Again, more innovation that is devoid of in most middle-income countries. Malaysia still faces challenges on its path to high-income status, particularly in areas like productivity growth, where it lags behind countries like South Korea and Taiwan. R&D investment, at just 0.95% of GDP in 2020, remains below the global average. Despite these issues, Malaysia has shown steady progress, with GNI per capita growing consistently between 2010 and 2023. I found a graph from FRED, which shows "Purchasing Power Parity Converted GDP Per Capita Relative to the United States, average GEKS-CPDW, at current prices for Malaysia". It shows that Malaysia’s GDP per capita, adjusted for purchasing power parity (PPP), has steadily grown relative to the United States over several decades. While there are some dips, such as in the mid-1980s and late 1990s, likely reflecting economic crises, the overall trend is upward. This suggests that Malaysia has made consistent progress in improving its economy, centered around innovation as discussed above. 
|
Memo: Skills and Jobs Mismatches in Middle-Income CountriesThe World Development Report 2024 identifies skill-job mismatch as a major challenge for middle-income countries, which contributes to insufficient productivity and slower economic growth compared to developed economies with more advanced labor markets. This issue needs public attention, as it raises critical questions such as the underlying causes of skill-job mismatch, the extent of its impacts, and the potential solutions policymakers can implement to address it effectively. Skill-job mismatches in middle-income countries typically take two forms: underqualification and overqualification. Underqualification is more prevalent, often driven by limited access to quality education. Many individuals in middle income countries are either unable to pursue higher education or graduate from programs poorly aligned with professional requirements. Employers may compound this issue by underinvesting in training and upskilling their workforce. Overqualification, on the other hand, arises when the labor market cannot offer positions that match the expertise of highly skilled workers. As an example, economies dominated by state-owned enterprises may experience slow innovation or have less competitive work environment, leaving skilled workers with few options but to accept less demanding roles or seek opportunities abroad. (In recent decades, it’s very common for skilled workers from MICs like China and India to immigrate to more advanced economies like U.S. or Europe in pursuit of better employment opportunities.) Both forms of mismatch result in inefficient use of human capital. These job mismatches have far-reaching implications for economic growth. Underutilization of workers’ skills and insufficient contribution directly leads to low productivity and economic output. High unemployment may persist due to gaps between the qualifications of job seekers and the needs of employers. Even employed workers may experience dissatisfaction and low job retention, which further diminishes productivity and economic stability. At a macroeconomic level, reduced productivity leads to lower marginal returns on both capital and human investments, ultimately slowing economic growth. A 2019 study on middle-income country skills and jobs mismatches by the International Labor Organization found that higher job-match rates are associated with increased GDP per capita (see graph below), while underqualification correlates negatively with GDP per capita. This emphasizes the importance of addressing the mismatch to unlock middle income countries’ economic potential. 
Policymakers in middle-income countries must prioritize interventions that reduce skill-job mismatches. Improving education systems is a critical step, ensuring they provide accessible, high-quality learning that aligns with market demands. Policies should be designed so that employers are incentivized to invest in training programs that upskill workers and bridge the gap between education programs and professional environments. Addressing unfair hiring practices which rely on personal and family connections is also essential to ensure a meritocratic labor market and positive social mobility. Additionally, promoting entrepreneurship can also create opportunities for skilled workers and lower the presence of overqualifications. Without barriers to entrepreneurial activities, talented individuals can generate their own opportunities while driving economic innovation and growth. |
|
How Singapore Escaped the Middle Income Trap Singapore’s remarkable transformation from a small trading port to a high-income, innovation-driven economy has captured global attention. One of the key reasons for its success lies in the deliberate strategies adopted by the government to ensure sustained growth and avoid the dreaded middle income trap—a situation in which a country’s growth stagnates after reaching middle-income status. Central to Singapore’s approach is its unwavering commitment to education, as evidenced by the country’s top performance in the 2015 Programme for International Student Assessment (PISA) results. In Math, for instance, Singapore scored 564 compared to the OECD average of 490, while its Science score stood at 556 versus an OECD average of 493. Notably, the proportion of high performers—those at Levels 5 or 6—was substantially greater in Singapore than in many other countries. These figures underscore the nation’s success in cultivating a rigorous education system, complete with well-trained teachers and a curriculum that prioritizes both foundational knowledge and applied skills. By developing a highly skilled workforce, Singapore shifted away from labor-intensive industries to technology- and knowledge-based sectors, thereby attracting substantial foreign investment and ensuring that its workers remained globally competitive. 
Another critical element in Singapore’s escape from the middle income trap is its strategic investment in research and development (R&D). Government agencies such as the Economic Development Board (EDB) and the Agency for Science, Technology and Research (A*STAR) have devoted significant resources to emerging areas like biotechnology, information technology, and robotics, creating an ecosystem where innovation thrives. By providing incentives for both local and multinational companies, Singapore has turned itself into a hub for advanced research that spawns high-value jobs and fosters a culture of creativity. This environment, in turn, fuels continuous advancement in technology and products, strengthening the country’s position as a regional leader rather than relying solely on lower-value manufacturing.
Finally, the attached graph tracking total FDI inflows from around 1970 to the early 2010s provides a compelling visual of Singapore’s growing appeal to foreign investors. The steep upward trend—particularly noticeable from the mid-1990s onward—reflects the government’s deliberate efforts to maintain a business-friendly environment through transparent regulations, strict anti-corruption measures, and efficient public services. These factors have consistently earned Singapore high rankings for ease of doing business, prompting companies from all over the world to establish regional headquarters and research centers in the city-state. In turn, the influx of foreign capital fuels growth in areas such as finance, logistics, and professional services, creating a positive multiplier effect that bolsters the broader economy. |
Escaping the Middle-Income Trap: Japan's Growth Miracle and Stagnation Without Creative DestructionJapan offers a narrative of a miraculous post war economic success story but also serves as a cautionary tale for developed economies of stagnation. Post-War Economic Expansion By the mid-1950s, Japan began transitioning into a 3i strategy, shifting from reliance on foreign technology to fostering domestic innovation. Between the mid-1950s and early 1970s, the nation achieved real GDP growth rates frequently exceeding 10% annually, a period often termed the "Japanese Economic Miracle." Entrepreneurs like Akio Morita, co-founder of Sony, and Soichiro Honda, founder of Honda Motor Company, embodied this innovation-driven phase by pushing technological frontiers outward, establishing Japan as a global leader in industries such as electronics and automobiles. This transition underscored the importance of balancing investment, infusion, and innovation to achieve sustainable economic growth. Economic Stagnation and the "Lost Decades" (1990s-Present) This stagnation highlights the importance of balancing the forces of creation, preservation, and destruction in an economy. Japan's dominance by established incumbents has had a suppressive effect on innovation and the emergence of new entrants. Among Japan's 20 largest companies by market capitalization, 14 were founded before the 1960s, reflecting a scarcity of fresh competition. Similarly, in the electronics manufacturing sector, only one of the two dozen firms was established after the 1960s. This illustrates the risks of failing to discipline incumbents and reward merit. The competition regimes must encourage new entrants and allow the process of creative destruction to take place. As a result, talent was misallocated and labor productivity has also suffered during this period. Despite being a high-income country, Japan faces a severe productivity issue, similar to that of many middle-income countries. Since 1970, Japan's labor productivity has consistently ranked lowest among G7 nations, with a productivity per hour in 2022 of approximately $53.4, compared to $91.5 in the United States. This persistent gap reflects inefficiencies, including underutilized talent and rigid markets. Addressing this productivity gap will require Japan to improve the allocation of talent, ensuring opportunities are based on merit rather than preservation of incumbency, and fostering a dynamic competitive landscape that is capable of driving innovation and growth. References: |
|
Does AI Educate or Replace Future Innovators? –The Impact of Artificial Intelligence on Students “The Rise of American Ingenuity” shows that education is critical for one’s inventiveness. Data in 1940 demonstrate a high correlation between one’s level of education and one’s probability of becoming an innovator. How might the introduction of AI changes the story? According to “The Impact of Artificial Intelligence on Innovation” (Cockburn et. al 2018), Artificial Intelligence gives both vertical and horizontal externalities in innovation process. Vertically, Artificial Intelligence, a product of innovation, drives productivity by its application in various sectors. Horizontally, Artificial Intelligence, an engine for innovation, creates new ways of understanding and imagining innovation. To say it succinctly, AI not only improves the efficiency but changes the rule of the game. Under such theoretical framework, it is reasonable to question in what way does/will AI change the rule of education and to explore its implications to future innovation. Does AI enhance students’ learning, thus facilitating future innovation? Or does AI harms future innovation by inhibiting students’ creativity and criticality? The application of AI in Education proliferates in recent years, facilitating the process of learning, tutoring, assessment, grading, and more. Here I want to focus on AI’s intervention on students’ learning experience. Specifically, I am interested in the integration of Large Language Models (LLMs) into classroom activities and its effect on students’ learning outcome. To propose a model: I care about students’ inventiveness so I will make students’ inventiveness as my outcome variable Y_inventiveness. I expect the students’ inventiveness to be impacted by the students’ access to AI, time spent on AI, time spent on study in general, and some unchangeable factors such as one’s socioeconomic status and Cognitive Ability, etc. Thus, the multiple regression model could look like: Y_inventiveness=β0+β1(AI_Access)+β2(Time_AI)+β3(Time_Study)+β4(SES)+β5(Cognitive Ability)+ϵ Where: An empirical experiment on AI’s impact on students took place in two high schools in in Brussels (Belgium) and one in Seville (Spain). In that experiment, the researchers investigated whether AI-generated explanations enhance learning by providing insights, or do they risk impairing students’ critical thinking by presenting inaccurate reasoning as authoritative. They find that AI tutoring, when performing step-by-step reasoning, helps enhance students’ ability to evaluate AI-generated information critically. In the graph below, students who are presented with AI-generated step-by-step reasoning are much better at discerning AI’s wrong results. This empirical case can guide my next step of research, where I hope to think further about the impact that different kinds of AI bring to students. |
Considering the Appropriation of Resources from the Global South in the 3i ModelThe World Bank has developed the 3i strategy to support the growth of middle-income countries. As countries develop and become closer to the frontier, or the leading world economies, they must support investment, then infusion, and later innovation. Historically, the global South has been plundered by the global North. The economic success of the global north is dependent on its appropriation of resources and labor from the global south. These remnants of colonization and imperialism continue to impart tangible benefits to the advanced economies of the global North at the expense of the global South. This has created one important relationship: the South is dependent on external financing. As a result, “Southern governments must compete with one another to offer cheaper wages and resources to attract foreign investment.” (Hickel et. al., 2021) Infusion is the importation of foreign ideas. Middle-income countries can benefit from the lessons learned by other economies and import outside technologies and business models. For example, in Korea, technologies were licensed from foreign countries. Naturally, this licensing will incur a cost. According to Hickel et. al., in 2017, the drain on the global South due to unequal exchange amounted to $2.2 trillion in 2011 U.S. dollars. (See the graph below) This a significant loss – “for perspective, $2.2 trillion is enough to end extreme poverty fifteen times over.” It is undeniable that this loss has far-reaching effects on the global South’s productivity. This wealth could be invested in education or technology. Additionally, it allows for the exploitation of labor. For example, Apple manufactures iPhones in China “because repressive labor control grants the efficiency needed to race products out the door.” In the Balassa and Samuelson model where wage is determined by relative productivity, the discipline of Chinese factory workers should be associated with high salaries. The investment and infusion goals of the 3i strategy incur an additional cost to this existing appropriation. I propose the following simple cost-benefit ratio to quantify the cost and benefits of investment and infusion (Stage 2i): where This model could be particularly useful on a country-by-country basis. For instance, if a MIC is competing for investment or infusion from a particular country, this model could help determine the amount of incentives to offer said country. Additional considerations for counties include: Figure 1. Drain from the global South, constant 2011 dollars, billions (1960-2017) |
|
The effects of AI on education Countries in the middle income trap suffer from a lack of growth in productivity. One of the benefits of AI is improvements in the quality of education. Both the Golden Age and Middle Income Trap reading provide the conclusion that higher quality of human capital leads to more productivity and innovation. Thus, we should venture to think about the different ways in which AI can improve the quality of education in these MICs, and how that could lead to higher productivity to help them escape the middle income trap. As non frontier countries, MICs are not at the forefront of innovation. Instead, they should work towards infusion—taking high income country technologies and implementing them to increase productivity domestically. However, as technology develops more and more rapidly, the infrastructure and knowledge needed to even reproduce such technologies becomes increasingly complex. MICs therefore should first work towards increasing their quality of human capital. Specifically, MICs should focus on having more STEM graduates who are capable of implementing frontier innovation. As we can see in the graph below, STEM scores are highly correlated with degree of economic development. Notably, two countries that have developed the most in the last two decades, China and India, are producing the highest number of STEM graduates each year.
AI can greatly improve two aspects of education: efficiency and personalization. One of the biggest problems MICs face is a large student to faculty ratio as well as inefficient use of investments and resources. AI can alleviate the work faced by teachers by assisting lesson planning and automating tasks such as grading. It can also train underqualified teachers to help them become more competent at their job. Lesson plans can be personalizable if students’ data are kept track of overtime, tailoring assignments to each student’s ability. This article on the Gates Foundation website provides great examples of uses of AI in education. MICs also struggle with the allocation of talent to task. Though it sounds dystopian, I wonder if AI could be used to predict people’s abilities to guide them into pursuing what they excel at. It could be a mechanism for finding talent at a younger age and expending more resources on potential inventors. That being said, we clearly run into equity and ethical issues of taking away freedom of choice, especially if implemented by force by authoritarian governments in a Brave New World type of situation. Of course, all of this is assuming that these countries even have the basic infrastructure with which to implement these functions. The bigger issue in these countries seem to be with the incumbents who seek to preserve the status quo, or government corruption resulting in inefficiencies and wasted resources. I am not sure if/how AI could lead to the correct incentives and better institutions within these countries. |
|
The "Rise of American Ingenuity" highlights how innovation-driven growth in the U.S. between 1880 and 1940 was supported by education, migration, and financial ecosystems. In contrast, middle-income countries face economic stagnation when they cannot transition from low-cost, export-driven models to higher-value, innovation-based economies.
The graph shows that middle-income countries (India and Iran), have a lot more publications related to AI compared to higher income countries. However, they still lack patents in AI. This suggests that middle-income nations may be strengthening theories in AI but remain dependent on higher-income nations for core AI innovations.
|
|
The Climate Dilemma of AI The empirical studies in “The Rise of American Ingenuity” examined basic demographic facts of inventors: their education, migration decisions, life cycle, and private rewards/wages of successful inventors. The results found positive correlations between the inventor’s level of education and their success, their father’s income and education, their social mobility and success, migration to more innovative states, to name a few. For example, Thomas Edison moved from Ohio to New Jersey where he was able to gather more funding for his laboratory. I wonder how these relationships fit into our class discussion of creative destruction and innovation. How do we maintain the balance between preservation, destruction, and innovation in societies where education and social status may not be regulated? What is the influence of social mobility on innovation: is it an investment channel for inventors or does it merely provide a platform for idea creation or both? For middle income countries that are grappling with current macroeconomic imbalances and achieving climate goals, what should they focus on in the short term, middle term, and long term in order to continue growth? In these two readings, the growth of middle income countries were evaluated using the US as a benchmark. In addition to high debt, climate change and geopolitical tensions pose two tough challenges for middle income countries to continue their growth. I am interested in honing in on the current political situation of the United States, and what this means for middle income countries in the years to follow. For example, how will new laws and restrictions under Donald Trump’s new presidency affect these countries’ abilities to utilize global intellectual resources (countries sharing technologies) and push towards or push away climate-friendly energy sources? Does AI offer a new solution to the growing challenges for middle income countries? I am curious about two main things: first of all, the dilemma between investment in artificial intelligence and its climate consequences, as well as the influence of the United States’ investments in AI on middle income countries. I have a theory that countries will continue to follow past patterns; in that they will prioritize low cost energy sources for research and development investments in AI over high cost investments in climate-friendly energy sources. 
^These countries have AI- startups. What elements of the 3i growth strategies have been at play to encourage this to happen? Regarding carbon emissions from AI: An article from Planet Detroit says that: “Microsoft has significant investments in AI, with a large stake in ChatGPT-maker OpenAI as well as its own Copilot applications for Windows. Between 2020 and 2023, Microsoft’s disclosed annual emissions increased by around 40%, from the equivalent of 12.2 million tonnes of CO₂ to 17.1 million tonnes.” |
|
Quantifying Productivity Gains from Patent Quality In The Rise of American Ingenuity: Innovation and Inventors of the Golden Age, patents are used as the primary proxy for innovation, which shows a strong relationship with long-run economic growth. The quality of patent observed is controlled through analyzing patent citations, to which, the paper concludes that the greater reception of patent citations for new inventors rather than incumbent inventors suggests a cycle of creative destruction. This makes sense given that citations are able to measure the "innovativeness" of a specific patent. However, within the collection of patent data, the paper chooses to isolate the patent's primary technological components as a means of classification. To this, I contend that a patent's innovative capability may also be dependent on its usability in non-primary technology as well. For example, in biotechnology, innovation of the delivery of a specific drug is contained within the specification for a specific drug. The productive capabilities that it causes, then, should be limited to the specifications of the drug. However, in more general technological patents, such as in AI, spillover effects may generate further patents in various fields, across different technological components. Although the raw number of citations may somewhat account for the patent's innovative capability, the applicability of this patent on furthering production may not be represented as correctly. Patents with more general usage may be able to capture a larger portion of production due to their ability to span multiple technological areas. Where the combined effect of the patent is established as its own output in addition to the output that it produces in citations from various fields. From the Corporate Patent Classification system (PCP) which is used by offices such as the USPTO, we can see that classification schemes are broken down into Human Necessities, Performing Operations; Transporting, Chemistry; Metallurgy, Textiles; Paper, Fixed Constructions, Mechanical Engineering; Lighting; Heating; Weapons; Blasting, Physics, Electricity; and general new technology. Each of these would have represented a coefficient of productivity per citation. |
|
From the reading “The Rise of American Ingenuity: Innovation and Inventors of the Golden Age”, we learn the “importance of commonly postulated drivers of innovation: population density, financial development, geographic connectivity and social structure measured by association with slavery” (page 11). Another macro level summary statistic I was interested in for a higher-level overview was GSP (Gross State Product) vs Number of Inventors. I used 2013 data as a case study with these as my data sources: Here is the graph with GSP as the x-axis and number of inventors by state on the y-axis. I also ran a regression: These results suggest a strong positive relationship between the GSP and number of inventors. The model explains 83.61% of the variability in number of inventors, making it a strong predictor. The results are statistically significant, with a very low p-value for the GSP, indicating that the GSP is an important predictor of number of inventors. Here is a general structure for the equation (Multiple Regression Model): Once I have gathered this data, I can estimate the model using OLS regression or more sophisticated techniques if needed. This will hopefully allow me to assess the magnitude and statistical significance of each predictor. |
|
Why Portugal Failed to Escape the Middle Income Trap - Case Study Portugal has struggled to escape the middle-income trap since its accession to the EU in 1986. Despite significant external support and access to EU markets, the country has regressed economically compared to its EU peers. Two primary reasons could account for this stagnation: the persistence of a less R&D-intensive economic structure and public policies prioritizing current consumption over long-term investment. Portugal’s economic base became dominated by small and medium-sized enterprises (SMEs) following the dismantling of its large industrial conglomerates after the 1974 Carnation Revolution. After the revolution, the Portuguese government undertook widespread nationalizations of industries considered critical to the economy, including banking, transportation, utilities, and manufacturing. These actions targeted conglomerates such as Companhia União Fabril (CUF), the Espírito Santo Group, and others, which were viewed as symbols of the old regime's economic elite. The scaling back of economic freedom preludes a prolonged slowdown episode of the nation's economy. What is interesting though, is that eliminating conglomerates and rent-seeking monopolists did not result in free-market competition and flourishing innovation. Rather, the abundance of SMEs proves to have stalled economic growth. While accounting for 96% of businesses, these SMEs operate predominantly in low-productivity, non-tradable sectors such as hospitality and real estate. Only 5% of Portuguese firms are involved in manufacturing, and overall private investment, including research and development, remains exceptionally low at less than 1% of GDP. This economic structure has hindered innovation and technological advancement, leading to a decline in total factor productivity (TFP), a key indicator of growth predicted by the Solow Model. Over the decades, the contribution of TFP to Portugal’s economic performance turned negative, further stalling convergence with more developed EU nations.
Portugal’s fiscal policies have further contributed significantly to its economic challenges. The government’s focus on welfare transfers, including pensions and public healthcare, has absorbed a growing share of national resources. Since 1974, tax revenues have more than doubled, rising from 20% to 45% of GDP by 2021. However, much of this revenue has gone toward sustaining social benefits rather than productive investment. Public investment fell from 35% of GDP in 1974 to below 5% in 2020, crowding out opportunities for infrastructure development and innovation. This misallocation of resources has gradually perpetuated a cycle of low productivity and limited economic growth. To make matters worse, Portugal is currently experiencing net emigration due to a combination of economic stagnation, high unemployment rates, and limited opportunities for skilled workers. As of 2021, nearly half of Portuguese emigrants had higher academic qualifications, more than twice the proportion of the overall population. Youth unemployment stood at 24% in 2023, compared to just 5% for older cohorts, which has led to the draining of highly educated and innovative talents to other parts of the world. The departure of a skilled workforce undermines Portugal’s capacity for innovation and productivity growth, further entrenching its position in the middle-income trap. Emigration further exacerbates the country’s demographic challenges, straining public finances and social welfare systems for a shrinking and aging population.
|
|
The article on “The Rise of American Ingenuity” helps, in addition to many other facts it elucidates, make clear the environmental factors and characteristics that lead to and motivate the emergence of inventors during a period of economic development in the US. The article does a great job of robustly testing for, and coming up with, a list of 9 empirical facts about innovation. Something I was interested in as an extension was looking at how firms, those which oftentimes enlist the help of these “inventors”, operate and how this could also help explain how an environment of innovation would lead to economic growth on a firm level. Originally, I began to think about what role firms played in this ecosystem of innovation. In this case I’m using the term “firm” rather loosely as any entity that would enlist the “services” of these inventors (services in this case being producing patents). This is an obviously overly broad swath of entities such as universities, research institutes, or in recent times AI companies like Deepmind and OpenAI. Oftentimes these inventors also rely on the resources of these firms to further their cutting edge research, sometimes even as a necessary support that enables them to conduct research in the first place. Hence, another proxy for the concentration of innovation could be the share of R&D workers, R, firms will employ relative to the number of regular workers, L. Using first principles taught in Econ 201, we can try to model their decision making based on their profit maximization objective. (I'll try to run through the set up but due to my inability to convert it to latex this looks really unsightly, you can just jump to the final equation) We let their profit share from innovation as a share of of GDP be: sπ = π/Y Let At be the productivity (roughly the “number of ideas”). The profit per idea is hence: sπ*Yt/At Let’s have the flow/change in ideas as dA, where the ideas per researcher is To get the wages per worker, we first assume that they’re paid the same as every other worker, so we just divide the overall value of labor by the number of workers, Lt. sL*Yt/Lt As with most cost benefit analysis, the optimal point for the firm is simply when MB = MC. MB in this case is the idea per worker multiplied by the profit per idea (profit per worker), and MC is just the wage. Hence we get: sπYt/At * dA/sRLt = sL*Yt/Lt Notably, we can rearrange this equation to isolate s_R. Note that we can also denote dA/At as gA, or the growth rate of innovation. The article runs through the algebraic manipulation, and essentially we end up with the following equation for s_R: sR = sπ/sL *gA In essence, the share of labor force comprised of R&D workers, or in this case inventors, has a positive relationship with the sπ/sL ratio and gA. In this case we will focus on gA. We see that as firms expect the growth rate of innovation to increase, they will likewise hire a larger share of these inventors. Intuitively, if we have more inventors being hired, it would lead to a positive self-reinforcing cycle in which more inventiveness gets incentivised as workers in search of a job would change their skillset to be more R&D focused, or be willing to take more risks to R&D with their new job security. Hence, even from this simple model we can see how even on the firm level, there is a self reinforcing cycle of inventiveness begetting more inventiveness. Of course, we must also be cognizant of the many assumptions underlying the model such as:
Even though it's a very simplistic model, it's interesting to see how focusing on firm-level analysis could help explain a different layer of the boom in american innovation. Especially with endogenous models of growth (like those of Paul Romer and the AK model) placing an emphasis on explaining At, the inclusion of firm-level analysis could help add to the robustness of the models. model: https://growthecon.com/StudyGuide/ideas/incentives.html |
Intellectual Property Rights and their Drive of Innovation
As touched on in the second lecture this week, intellectual property rights (IPR) are directly tied to the motivation for economic development and thus strong IPR policies are pivotal in fostering innovation and sustaining long-term growth. They incentivize innovation by safeguarding creators' ideas, ensuring that inventors can reap economic benefits from their work. This protection creates a system conducive to research and development which is critical for technological advancement and productivity improvements. Looking at the United States during its "golden age of innovation", we see a strong correlation between patented inventions and long-term economic growth which highlights the importance of IPR in sustaining innovation-driven economies. The World Development Report 2024 argues that countries transitioning to high-income status must prioritize not just investment but also technological infusion and innovation. A robust IPR framework is vital in facilitating these transitions by enabling technology transfer and encouraging domestic firms to invest in R&D. Here I produced a scatter plot comparing countries' IPR scores (1) with their GDP per capita growth multipliers (1950-2022) (2) which demonstrates a positive correlation. Countries with higher IPR scores, such as Australia and South Korea, exhibit stronger GDP growth multipliers compared to those with weaker IPR protections. This trend shows how strong IPR attracts foreign investment, fosters technological diffusion, and enhances productivity. Strong IPR frameworks better positions countries to integrate and adapt foreign technologies, as seen in South Korea's evolution from a technology importer to an innovation economy. Upper-middle-income countries face diminishing returns from capital investments and must transition to innovation-led growth. Strong IPR regimes facilitate this by reducing the risk of intellectual theft and encouraging private sector investment in high-value R&D. Weaker IPR protections in many middle- and low-income countries contribute to stagnation in innovation and perpetuate the middle-income trap. Policymakers must recognize IPR as a cornerstone for sustainable growth strategies, particularly for middle-income economies seeking to transition to high-income status. They can do this by enhancing legal structures that protect intellectual property and institute measures like tax incentives for innovation and public-private partnerships. |
|
Is AI “Just” Prediction Technology? Power and Prediction: The Disruptive Economics of Artificial Intelligence titles its third chapter “AI is Prediction Technology.” The authors explicitly distinguish the task of prediction from the task of judgment, which is reserved for “those behind the machines, guiding how they react to predictions.” In particular, they argue that “While prediction is an expression of likelihood, judgment is an expression of desire–what we want. So, when we make a decision, we contemplate the likelihood of each possible outcome that could arise from that decision (prediction) and how much we value each outcome (judgment)”. Following their logic, currently most decision-making settings bundle prediction and judgment together, but AI adoption raises the possibility of decoupling these two inputs to decision-making. As AI advances and the cost of prediction drops (think declining $/token for API calls to OpenAI), the value of human prediction (a substitute) will fall, while the value of judgment (a complement) will rise. In light of recent announcements over the past few months of “reasoning models,” I wanted to revisit this definition of AI as solely prediction technology. The graphic below is an oversimplified approximation of the architecture underlying OpenAI’s o1 reasoning models, which I pieced together from brief disclosures (https://openai.com/index/introducing-openai-o1-preview/). The model displays “agentic” behavior in the sense that it iterates upon chains of thought that deconstruct the initial prompt into a series of logical steps, eventually choosing the optimal chain of thought to output to the user. I attempted to maintain the framework of Agrawal, Gans, and Goldfarb by highlighting how each stage of AI within the architecture can be construed as merely “prediction” on the left. For example, o1 likely contains a “selector”/“verifier” model that has been trained to “predict the chain of thought that is most likely to lead to the desired outcome.” An alternative interpretation of o1 and this new class of “reasoning models”, however, is that they possess judgment capabilities. For example, one could argue that the training data for the “selector” model could encode values and human desires implicitly within a reinforcement learning-based algorithm. Thus, the “selector” component of the o1 architecture is valuing each chain of thought and imparting judgment, arriving at a full-fledged decision based on prior predictions. While this may appear purely semantic, the cost of prediction and judgment are underlying variables that influence the economics of decision-making. If it turns out that AI is capable of transcending mere prediction and imparting judgment, AIs can be deployed as autonomous agents possessing end-to-end decision-making within a given domain – say, for example, as a B2B sales rep or a healthcare claims administrator. The role of “human judgment” as a unique resource may then be reserved for adversarial environments where the enemy/competitor/customer desires outcomes outside of the training data. Only in those cases would a reinforcement learning mechanism embedded in a “selector” model fail to value human desires properly and make the right decision.
|
|
Using Deep Learning and AI to Predict and Mitigate Wildfire Impacts in Los Angeles With the recent wildfire devastations in Los Angeles, I was intrigued by how wildfires in the area could have been detected beforehand -- predicting the likeliness of wildfire presence, which would have allowed firefighters to take steps necessary to mitigate wildfire risks before it began spreading. In the "Deep Learning" text, I learned about how deep learning uses feedforward neural network architectures to learn to map a fixed-size input to a fixed-size output. In this case, perhaps ConvNets can be utilize to process wildfire images and videos (2D and 3D arrays) by training these ConvNets models on satellite imagery, drone feeds, and thermal imaging. The model can detect early signs of wildfires, such as smoke, hotspots, and vegetation dryness. RNNs can also be utilized to process sequential data, such as meteorological patterns and wind speed to predict fire progression patterns over time. In the "Prediction Policy Problems" text, I learned about how ML can address prediction-based policy problems -- minimizing prediction error and estimating unbiased causal effects. It can also handle complex datasets, including those for wildfire data, which would involve high-dimensional variables and non-linear relationships, such as multiple variables relating to wind speed, temperature, vegetation type, and fire proximity to urban areas. This can assess risk in certain areas, and it would be cool to use this to provide real-time alerts and generate actionable insights for first responders and policymakers. Thus, let's use these concepts to help predict and forecast wildfire occurrences, assessing their risk level and targeted areas. We can model wildfire spread using Los Angeles wildfire data, integrating input variables of wind speed, temperature, and vegetation density. Our output variable would be the probability of the fire spread to adjacent zones and urban areas. First, we could use a logistic regression model (because outcomes are binary: either wildfire spread or no wildfire spread) as a baseline for finding the probability of the wildfire spreading. Predictions will be mapped into a range between 0 and 1. B0 is the intercept, while B1, B2, and B3 are the coefficients for each variable. I have not created an AI framework before, but perhaps after this, according to the "Deep Learning" text, we can utilize stochastic gradient descent to compute the average gradient (through backpropagation procedure) and adjust weights accordingly to minimize errors for many small sets of examples from the training set. This would be done after identifying patterns through ConvNets and RNNs. Then, we can measure the system performance on a different set of examples (test set) to see the generalization ability of the machine. Below is a very simplified heatmap (generated in Google Colab through Python) to figure out the wildfire spread probability. I used a range of 0 to 50 mph for wind speed (cite), temperature range of 20 degrees C to 40 degrees C (as it is a common temp. range in wildfire-prone areas -> cite), and a vegetation density range from 0 to 1 (common representation of vegetation dryness in a similar mountain area, the Cumberland Mountains in West Virgina (was not able to find LA data) -> cite). I am using a sample vegetation dryness of 0.5 to figure out the probability as an example. |
|
Creative Destruction in the era of AI The article "The Rise of American Ingenuity" highlights that in the past, innovation has been driven by young inventors through creative destruction, where young workers with fresh ideas and high motivation come up with innovative ideas that displace older technologies and inventors. However, I want to highlight that in the recent years with the quick emersion of artificial intelligence this dynamic is likely to slightly shift and the clear trend of young innovator over old innovator might be weakened. Technology and innovation builds on top of another. The increasing complexity of cutting-edge technologies has raised the barriers to entry, making advanced education, specialization, and cumulative experience more critical for successful innovation than early-career creativity and motivation alone. The article highlights a key tradeoff that inventors face: the time spent on education delays their entry into active careers, but that education is crucial for developing the advanced ideas necessary for meaningful technological progress. Supporting this, Jones (2010) observes that over the 20th century, the age at which major breakthroughs occur has increased by five years, which reflects the rising demand for more education and specialized knowledge before inventors can make high-impact contributions. As mentioned today in class, in the past, having college degree has a strong impact on the rate of innovation, but in recent years, PHD degree leads to an even more obvious trend of innovation, showing the importance of deep understanding in specific fields in innovation today. An article by Jones written in 2005 already highlights that there is a clear rising trend of increase in age for successful innovations, especially in "knowledge-base" industries. Graphs below shows the clear rising trend in age for innovation. 
The field of AI is evidently a "knowledge-based" industry, developing groundbreaking technologies in AI requires deep technical expertise across fields such as machine learning, data science, and neural networks. Unlike earlier technological revolutions, where young inventors could enter the market with just creative ideas, AI innovation requires a cumulative understanding of complex systems, making advanced education and experience increasingly important, therefore leading to this rising in age trend we see in innovation. Another aspect worth considering is that as the education system adapts to the rise of AI and immerses it in future generation's everyday life, the trend where young innovator innovates more than old innovators might rise again as they grow up building on knowledge and clear understanding of AI, overcoming the challenge of high barrier of entry and enabling them to innovate at younger age ultimately. |
Behind China’s Falling Productivity MetricsThe 3i model emphasizes the diminishing returns of real-asset investment and the growing need for productivity as MICs attempt to transition into higher income. However, many MICs that has seen success, like China, South Korea, and Poland, have also been facing recent declines in productivity. Using China as a case study, what domestic factors and policies are putting downward pressure on productivity? China's enormous growth over the last four decades has been coupled with massive infrastructure projects. In 2016, infra investment was 24% of its GDP, more than 95% of both advanced and emerging countries around the world. A 2024 study by Qian, Ru, and Xiong that such state-led initiatives positively enhanced firm productivity. Doubling infra investment is associated with an increase of 4% in Total Factor Productivity (TFP), 0.9% in Return on Assets (ROA), and 5.2% in total sales; when coupled with introduction of the 36 Clauses (a series of market-oriented policies), the associated increase in TFP, ROA, and total sales were 42.5%, 66.67%, and 38.5% respectively. These results suggest that state-led investments and the transition to market-led economies are highly synergistic in boosting firm productivity. However, China’s TFP growth has been declining from 2.8% annually pre-GFC to ~0.7% post. While state-led infrastructure enhances productivity, its effect is being negated by factors like resource misallocation, frictions to market entry and exit, and over-investment in real assets, as proposed by Brandt, Litwack, and Mileva. As shown in Figure 9, China’s manufacturing TFP has shown big slowdowns due to frictions in market entry & exit. Before the GFC, the majority of the increase in productivity was caused by entry of new, more efficient firms; however, post-2007, the productivity boost from new entrants was considerably lower and even negative at times, likely due to non-market factors in the entry process. Furthermore, there is very little productivity gain from exit of inefficient firms, suggesting over-preservation of incumbents, as also examined in our lecture notes.
Another key drag on China’s productivity is its over-investment in infrastructure and real assets, which crowds out private investments while producing low returns. These investments’ enormous use of capital stock crowd out resources available to higher-productivity private sectors. State-Owned Enterprises’ debt load was 73% in 2012 and 103% in 2015, taking up much of local government financing vehicles. At the same time, over-construction of infrastructure and housing led to rising Incremental Capital-Output Ratio (ICOR) as illustrated in Figure 17, which suggest decreasing returns on capital. As such, we see clear resource misallocation and missed growth potential in private sectors with much higher productivity, but less access to capital.
Finally, lower TFP may also stem from China’s rising share of Non-Market Services as percentage of total labor force, which has much lower labor productivity compared to Trade & Restaurants and Financial & Business Services. While this phenomenon is inevitable as higher income levels result in more service-oriented transactions, it furthers the need to implement methods that enhance service-sector productivity, such as AI and automated processes. |
Creative Destruction: The Netflix EffectA Netflix subscription costs more and more, and HBO now has ads—it’s The "Netflix Effect". Prominent streaming services like Netflix and HBO, now MAX, were new introductions to the movie scene and highlighted the idea of creative destruction as they disrupted existing video and movie rental services, along with the broader media sector. Creative destruction, a theory developed in the 1940s by economist Joseph Schumpeter, describes how innovation dismantles outdated structures to pave the way for new growth and technological advancement. The figure below illustrates how the once-dominant video rental giant Blockbuster, as well as services such as VHS and DVD rentals, were essentially pushed out of the media industry by Netflix’s innovative approach to digitizing video content. Link to Image Article 
This shift is a great example of the innovative activity trends described in The Rise of American Ingenuity. The paper highlights how young inventors, producing higher-quality patents, often created transformative innovations that disrupted incumbent players. Netflix, in its early days, embodied the same idea. By introducing a subscription-based digital streaming model, it didn’t just challenge Blockbuster—it forced an entire industry to evolve, pushing competitors like HBO and Disney+ to invest in higher-quality content and diverse streaming selections. This mirrors the pattern seen in the American Ingenuity study, where new entrants’ innovations outpaced incumbents, driving the reallocation of resources toward more productive uses. The World Development Report 2024 discusses how entrenched incumbents put up barriers to try and prevent new innovations, as seen in high-carbon industries resisting the shift to low-carbon technologies. Similarly, in the streaming sector, established, older players like cable networks or Blockbuster first resisted streaming by bundling packages and lobbying against market changes, but as the consumer preference for streaming became unavoidable, the incumbents were forced to either adapt or lose relevance altogether. The consumer has definitely benefited the most from this process, gaining access to an ever-changing stream of content that gets better and more diverse due to intense competition. But there is also a tension between the initial disruptive phase of creative destruction and the eventual consolidation of market power by the disruptors themselves. So will the streaming giants of today, like Netflix, eventually become the entrenched incumbents of tomorrow, using their dominance to newly limit competition and innovation? Or can the mechanisms that drive innovation, like Netflix’s early rise, be sustained in a maturing market where already dominant players have less of an incentive to innovate? And what does it mean for other industries that have seen or will see this same “Netflix Effect” with a novel product or approach reshaping an entire industry? This idea/question could be developed more in my final project by looking into how AI-driven platforms, such as Netflix’s recommendation algorithms, influence innovation and competition in digital markets, and if they strengthen or undermine necessary principles of creative destruction, in a way that I can generalize conclusions to other maturing industries and technological advancements. |
|
Are Countries of the World Prepared for AI? Based on our readings and discussions currently in class, we have finally started to look at AI as a possible way in which countries can finally escape the “middle income trap"). However, AI is a relatively new technology in our world. In light of this fact, in order to tap into the AI sector, you need to be on the frontlines in development. The alternative is waiting years for AI to finally become more established and accessible and implement it as an existing technology. However, this would simply result in a further prolonged game of catch up as the country falls behind technologically in the AI space and continues economic stagnation. Furthermore, another table below analyzing the cross-sectional relations finds that there is not only dependence, but also many negative relationships between the study’s metric of AI readiness and institution related metrics. This implies that the economic stagnation that a country experiences, and the corruptions that come align side it are a huge roadblock in the way of effectively stimulating the society and economy. 
In order to combat this issue, central governments need to take charge of laws and policies to implement AI based policies, challenge the existing incumbents in society that may block the attempts to foster AI development, redirect resources to education, research, and innovation that will give reason for talents to stay and avoid the Brain Drain problem. |
|
AI and sustainable development in middle income countries. Harnessing the predictive power of AI can guide climate policy in middle income countries, offering new strategies to address climate change and capitalize on current crises. The World Development Report 2024: Middle Income Trap emphasizes the need for middle income countries to use crises to their benefit, as such crises can weaken the strong preservation forces. With regard to the climate crisis, the World Bank suggests MICs should focus on decarbonization, energy security, and sustainability, to name a few. I will focus on the relationships between AI and the latter–that is, that middle income countries can leverage AI to support sustainable development. We look at Vietnam, which is classified as lower-middle income, and the use of neural networks to manage forest resources and land. Many MICs, including Vietnam, are disproportionately vulnerable to climate change, often as a result of their high reliance on agriculture. A 2022 study by Khanh Nguyen-Trong and Hoa Tran-Xuan uses a CNN (which were explained in the Deep Learning reading) to detect forest cover changes. The CNN divides images into distinct segments, and then the model is fed multi-temporal satellite images. It performs with 95.4% accuracy in the testing set, demonstrating that it accurately detects changes in forestation and coastal forest cover. For a country that is nearly 50% forest, this greatly speeds up the analysis of the large forests across different regions in Vietnam. Above all, in this case, AI outperforms conventional human techniques, which the paper cites has accuracy of about 90%. The machine learning model itself mostly provides value in the context of its policy implications. For example, AI systems that include real-time satellite data can allow governments to implement timely interventions. Farming practices can be adjusted quickly, which then facilitates the proper management of natural resources. With increased knowledge about changing landscapes, policy could more effectively target key problem areas related to food security and sustainable development. Moreover, when such technologies are offered to farmers, such as the Vietnam application AI Plant Doctor, farmers are offered data-driven insights to boost productivity. It is important to note that, in practice, technological barriers can prevent AI from being used to its full potential. Across Southeast Asia, we see that internet access in Vietnam lags behind access in the high income countries, Singapore and Malaysia. In addition to the high cost of accessing AI technologies, small farmers in rural areas of Vietnam will struggle to access technologies that could improve their land usage. (The graph below uses 2022 data from the World Bank). 
That said, as internet access increases year over year and by further harnessing AI technologies like the CNN in Nguyen-Trong’s study, farmers and governments alike can more deeply understand the affects of the climate crisis. AI, then, provides new opportunities for countries to promote sustainability and resource management amidst the current climate crisis. |
|
Brain Drain and the Lack of Innovation The Solow growth model was very important for economics. One of its major conclusions was that economies would converge over time, with poorer nations catching up to richer ones as they benefit from higher returns on capital. However, we do not see this empirically, as the Solow growth model treats technology as an exogenous factor, assuming it grows at a constant rate and is evenly distributed. One of the reasons for such disparity is what some experts call “brain drain”, the emigration of skilled individuals to other countries. This is a great problem, as most of these middle-income countries invested their limited money on the education of these qualified workers only to lose them and the potential innovations they could produce. The paper called “The Rise of American Ingenuity: Innovation and Inventors of the Golden Age” by Ufuk Akcigit, John Grigsby, and Tom Nicholas highlights the correlation between having a college degree and being an inventor. Therefore, the problem of losing potential inventors to other countries becomes obvious. For instance, according to the immigration policy institute 14.1 million adult immigrants in the US had at least completed college, and they probably contributed to the country with innovations. One possible explanation for the high immigration rates is the difference in compensation. In 2022, the average annual income for U.S. workers aged 25 to 34 with a bachelor's degree was approximately $66,600. In contrast, in India, employees with a bachelor's degree earned an average annual salary of about $20,371. This comparison highlights a significant income disparity between college graduates in the two countries. An Indian skilled worker has a strong incentive to leave their home country for better conditions. The following model exemplifies this issue: Workers maximize their utility based on the decision to emigrate or stay: Where c = consumption, l = leisure, and h is the emotional hassle of emigrating, thus if no moving occurs it would be 0. Where (1-l) is the labor supply and i is the financial cost of immigrating. Therefore, we have: For workers staying: For workers leaving: Thus, a worker will leave their home country if We can clearly see some factors that impact emigration. The wage gap |
|
The prisoner's dilemma Every economics student has learned the classic prisoner’s dilemma, where players of a game end up in a suboptimal situation because they are unable to faithfully cooperate. Humanity, due to its societal and economic structure, is currently locked into an increasingly unfavorable Nash equilibrium. To demonstrate this, consider the existence of the military. The military, our nation’s (and possibly the world’s) largest institution, operates with the purpose of researching the most efficient ways to control and kill people. If there were a “central planner,” as economists love to dream, would they really allocate thirteen percent of our money to such a wicked purpose? I seriously doubt it. The military is a necessary response to our current system in which competition, at an individual, firm, and national level, drives humanity towards a (somewhat) efficient economy. This system of competition, as Professor Akcigit explained, can be greatly improved with the use of AI. Unfortunately, any competition inherently creates waste; any money directly spent on competition (advertising etc.) could have been spent on research or capital. Further, a competition-based economic system creates a society that fosters violence between nations, something that gets scarier as technology progresses. AI, much better than humans, can logically evaluate the long-term future with minimal bias. Thus, an AI would be able to instantaneously make policy decisions that are, at the least, fair and viable and, once optimized, potentially perfect. As such, an AI (more likely a network of AI agents) could serve as a fantastic leader for humanity and help foster a system of global collaboration. This future is a longer-term vision than what we took in class and requires nearly all humans to have faith in AI governance. I think many people assume that this future necessitates redistribution or some kind of communist society. I return to the prisoner dilemma example; in my mind, there is a future where we are universally better off (e.g. our largest institutions are focused on the betterment of society). AI, by aligning humanity, can steer us towards this future and keep us in an optimal reality where progress is fast and happiness widespread. Exactly what this future looks like or how we reach it is impossible to predict, but I am certain Artificial Intelligence is essential to its fulfillment. Graph source: Congressional Budget Office |
|
Creative Destruction & Innovation for Economic Growth Creative destruction, is a transformative process in which new innovations disrupt existing industries, driving productivity gains and economic growth. This mechanism is particularly relevant for countries navigating the middle-income trap, where stagnation often arises from a failure to transition from imitation-based growth to innovation-led development. Drawing from endogenous growth models, the role of research and development (R&D) investment is pivotal in catalyzing creative destruction. By analyzing data on R&D intensity across countries such as the United States, Japan, China, India, Turkey, and Colombia, we can establish a clear relationship between R&D investment, innovation productivity, and economic growth. g=λ⋅Rg where g represents the growth rate of GDP, λ captures the efficiency with which R&D translates into innovation, and RRR denotes R&D expenditure as a percentage of GDP. Developed economies like the United States and Japan, with high levels of R&D investment and efficiency, exhibit robust growth driven by continuous waves of creative destruction. In contrast, middle-income countries such as India, Turkey, and Colombia allocate relatively low percentages of GDP to R&D, resulting in weaker innovation output and slower structural transformation. 
|
|
Argentina’s Escape from & Return to the Middle-Income Trap Argentina's historical economic rise and subsequent decline offer a compelling case for understanding the role of investment, infusion, and innovation in sustaining growth—a highly relevant framework in today's context of technological advancements and AI-driven innovation. Towards the end of the 19th century and the beginning of the 20th century, Argentina was one of the wealthiest countries in the world. Argentina’s economic prosperity rivaled that of the United States, Canada, Australia, New Zealand, and Western Europe. Agricultural exports, including beef, wheat, and wool, drove growth, with Argentina becoming a major global supplier. The development of infrastructure like railroads and ports was key in enabling industrialization. Argentina constructed over 30,000 kilometers of track during this period, integrating its agricultural hinterland with global markets. Furthermore, policies promoting foreign investment, open trade, and property rights laid the foundations for such growth. Thanks to these elements and the execution of an ambitious and effective education system, Argentina successfully implemented the first of the “three i’s”: investment. Complementing Argentina’s capital accumulation and participation in world trade was a massive influx of immigrants from Europe, especially Italy and Spain. This immigration wave was only smaller than that of the United States. These immigrants brought skilled labor and new technologies, enhancing Argentina’s human capital and increasing productivity. They introduced advanced farming techniques and manufacturing skills. As a result, Argentina achieved infusion, the second step required to become a wealthy nation. Finally, Argentina was also able to partially reach the third and final step of growth: innovation. New technologies in meat processing, refrigeration, and shipping revolutionized the global beef trade and made Argentina one of the biggest suppliers to Europe. Productivity in this industry also rose thanks to developments in cattle breeding. Innovations occurred in the grain industry too, with new silo designs and grain elevators being introduced to facilitate large-scale wheat exports. Further innovations took place in logistics, transportation, healthcare, and textiles, cementing Argentina’s place as a leader in economic development and creative destruction. Unfortunately, Argentina’s status as a high-income country did not last very long. Economic growth started to slow after the Great Depression, falling behind other countries with previously comparable standards of living like Australia and Canada. This stagnation was caused by political instability, populism, economic protectionism, and weakening institutions. In a unique case of a nation declining despite having all the resources to sustain existing levels of development, Argentina’s early economic advantages were squandered. Argentina’s decline can be explained by stepping back down the ladder of the three i’s. Overregulation, weakening property rights, and economic volatility stifled innovation. Then, protectionism and isolationism halted infusion. Argentina returned to step 1, still having good infrastructure and physical and human capital but lacking the infusion and innovation necessary to increase productivity and remain a high-income country. Thus, Argentina’s failure to sustain infusion and transition to innovation perfectly exemplifies the middle-income trap, where bad policies hinder the dynamic economic restructuring and creative destruction needed to reach high-income status.
|
|
Brain Drain: Retaining Talent in MIC's One of the most intriguing findings of The Rise of American Ingenuity is that innovators were much more likely to have migrated from their home state to a different state with better prospects for innovation (even more so that highly skilled workers). This speaks to one of the greatest challenges facing developing nations. MIC’s are full of untapped talent and potential, but without robust intellectual, technological, and scientific infrastructure, many talented individuals migrate in hopes of more lucrative prospects abroad. This creates a deleterious cycle of continued underdevelopment within MIC’s. Innovators are certainly the backbone of social and scientific progress. However, MIC’s inhabit a world with a myriad of existing advanced technologies that simply require infusion, rather than invention. Doing so requires an intelligent and educated body of workers. Thus, I find levels of education to be an important indicator, just as innovator status was for the American ingenuity study. Concerningly, data from the IMF reveals a largely positive correlation between level of education and migration probability. For example, in select Asian countries, those with the highest level of education were many times more likely to migrate to the United States than lesser-educated peers. Some nations, like Jamaica and Guyana, have migration rates so high that around ¾ of highly educated individuals leave for better prospects in the United States.
It is imperative for MIC’s to establish infrastructure that attracts these talented individuals. I think it is interesting to consider migration like the job market. Much like someone migrates from a lesser developed country to a more developed one, employees often search for better employment opportunities. Thus, understanding the factors that keep employees happy with their current work environment may be informative for MIC’s. A study in the Harvard Business Review found three primary factors that drive worker turnover: opportunities for career advancement, pay, and company culture. These are very similar reasons for a talented individual to migrate as well. The rate of migration to the United States from MIC’s is so great because of greater career opportunities, stronger pay, and a more stable political environment. Thus, if developing nations wish to escape the middle income trap (which I argue can only be done with the retainment of local talent), these turnover factors should be a driving principle. In order to keep a well educated citizen, one must see robust infrastructure that facilitates career advancement. There must also be a competitive labor market with high wages. Lastly, there must be stable political institutions that guarantee these conditions on opportunities. |
|
One of the most fascinating aspects of deep learning is its ability to process raw, untransformed data and extract meaningful patterns and findings through representation learning. The provided image of a simple neural network highlights this concept simply yet effectively, illustrating how two sigmoid activation functions in a hidden layer transform raw input data into a final single output ( could be a classification score or probability for example). |
|
Role of the Economics of Data When Looking at AI as an Escape Route for the Middle-Income Trap AI’s potential as a transformative force for middle-income economies rests on leveraging the economics of data. As data exhibits nonrivalry—allowing simultaneous use without depletion—its strategic deployment can enable these countries to leapfrog traditional development constraints. However, disparities in data quality and infrastructure underscore the complexity of translating this potential into economic advancement. 
The economics of data offers middle-income countries a unique opportunity to redefine their developmental trajectories. However, realizing AI’s transformative potential requires deliberate action to address systemic gaps in data infrastructure, governance, and workforce capabilities. Only with strategic investments and policies tailored to these challenges, AI can be a viable pathway out of the middle-income trap. |
|
Memo: Deep Learning—Catalyst for AI-Driven Growth and Innovation Although the topic of Deep learning has yet to emerge in our lectures this week, LeCun, Bengio, and Hinton’s foundational article in Nature outlines how deep learning’s ability to discover intricate patterns in high-dimensional data is reshaping industries at a rate that outpaces conventional machine learning. Deep learning, a transformative subset of machine learning, has become a cornerstone of artificial intelligence, driving advancements in diverse fields from natural language processing to medical diagnostics. This memo focuses on how deep learning’s automated representation learning can drive economic growth by reducing the need for extensive domain expertise and enabling innovation at scale. Traditional machine learning required handcrafted feature engineering, a process demanding significant expertise and resources. Deep learning, by contrast, uses multilayered neural networks to extract features automatically from raw data. For instance, convolutional neural networks (CNNs) detect patterns in images—from edges to complex shapes—facilitating breakthroughs in autonomous vehicles and facial recognition (p. 437). Additionally, deep learning systems thrive on abundant data and computational resources, making them well-suited for modern environments. Their ability to generalize across tasks has led to their adoption in critical domains such as drug discovery, where predicting molecular activity accelerates research cycles (p. 436). By democratizing access to AI capabilities, deep learning reduces barriers for small/medium enterprises to innovate. Industries like e-commerce, finance, and logistics leverage AI-driven personalization and optimization tools, which were once exclusive to tech giants (p. 439). A key example briefly mentioned in the paper are autonomous vehicles, which exemplify how deep learning drives growth through innovation. CNNs enable vehicles to interpret real-world scenarios, such as identifying pedestrians or reading road signs, in real-time. Coupled with recurrent neural networks (RNNs) for sequential data processing (e.g., anticipating pedestrian movement), deep learning empowers self-driving vehicles to operate safely and efficiently. Economic impacts include reduced transportation costs, enhanced logistics, and job creation in adjacent fields like sensor manufacturing (p. 440). Overall, this paper highlights at a high level that deep learning is not merely a technological milestone but a pivotal driver of economic growth and societal progress. As industries continue integrating these systems, their capacity to automate, educate, and innovate will redefine global competitive landscapes and elevate inclusivity. Analytical Element: Economic Impact Channels of AI To illustrate how AI, including deep learning, impacts economic growth, we can analyze a chart from a McKinsey simulation that outlines seven key channels contributing to the net economic impact of AI by 2030. 
Augmentation: AI enhances human productivity, contributing an 11% cumulative boost by 2030. For deep learning, this includes applications like automating complex tasks such as image recognition in healthcare or optimizing supply chain logistics. Substitution: Automation replaces manual labor in repetitive tasks, offering a 2% boost. This reflects AI’s ability to streamline operations, as seen in industries like manufacturing and customer service. Product and Service Innovation: The most significant contributor, with a 24% boost, showcases AI’s role in creating new offerings, such as personalized medicine or autonomous vehicles powered by deep learning models. Global Data Flows: AI facilitates cross-border data integration, contributing a 2% boost through improved connectivity and operational efficiencies. Wealth Creation and Reinvestment: AI’s economic gains enable reinvestment, resulting in a 26% gross impact. Deep learning’s applications in finance, for instance, accelerate wealth generation through enhanced decision-making. By 2030, the simulation projects a 16% net economic boost from AI adoption, underscoring its transformative potential. Deep learning, as a cornerstone of AI, is central to driving these impacts by automating feature extraction, enabling scaling, and democratizing innovation. |
Minimizing LLM Operational Costs and Addressing Energy ImpactRecent advancements in Large Language Models (LLMs) have transformed tasks and even some industries. Nevertheless, their computational demands have led to escalating operational and environmental costs, something the authors of Power and Prediction discuss with respect to the impact on the electrical grid. AI’s transformative potential is dependent on rising computational requirements, which in turn lead to higher energy consumption. Many organizations grapple with the challenge of controlling associated operational expenses. Here, I wish to explore two primary deployment options—Software-as-a-Service (SaaS) and self-hosting and present an analytical approach to minimize overall costs and address energy considerations. SaaS expenses primarily derive from a pricing model based on the count of tokens used, along with an fixed subscription fee depending on the tier of service. This approach provides simplicity by eliminating the need for on-premise hardware and associated software maintenance costs. In contrast, self-hosting often necessitates substantial upfront investments in GPU infrastructure or cloud compute instance allocation, as well as ongoing expenses for electricity, labor, and system maintenance. In order to quantify a 'impact cost' for electrical and system demands, I propose the following set of equations that capture the dynamics of operational costs. Deriving from these and other factors that I have not captured here, we can break down the unit economics for running these models for organizations and individuals. The minimum operating cost can be expressed as the following:
A related equation could be proposed for the environmental impact but with a different metric to measure, I chose to focus on cost as the deciding factor. These models are highly simplified and inspired by easy-to-use pricing calculators available on cloud provider websites at the point of sale. Capturing complex other parts of the infrastructure are hard to standardize to one model and are highly variable for the scope I am attempting to discuss here.
As organizations expand their model usage, the power consumption (P) and usage hours (U) both increase. This is similar to the growing demands of electricity expressed in the book in regards to energy sources and sinks. The book cautions that the exponentially growing compute requirements of artificial intelligence can pose substantial challenges to power grids and organizational budgets. By aligning these cost equations with the themes in Power and Prediction, it becomes clear that conscientious resource allocation is key towards driving down our impact cost. Cloud providers such as AWS and GCP give us detailed tier-based pricing and impact analyses for these resources, something that can help both derive cost and environmental impact estimates in their respective quantifiable metrics (dollar amount, CO2 emission metrics etc.). Exploring renewable energy sources such as the flurry of recent nuclear energy deals from companies like OpenAI and Microsoft can change the way we capture these dynamics and perhaps lower our impact cost. In sum, the heightened focus on electricity usage described in Power and Prediction bolsters the argument that energy costs, if left unchecked, can undermine any economic and societal benefits LLMs promise to bring over the coming years. Incorporating a dedicated energy term into operational equations helps in both controlling costs and adhering to sustainability goals. |
|
Middle-income countries often experience a growth slowdown, referred to as the "middle-income trap," due to their inability to transition from low-value-added activities to high-value-added economic activities. This stagnation arises from limited innovation and lower returns from imitation. Still, Costa Rica and Poland were/are middle-income countries that have been able to continue growing at a faster rate than other middle-income countries. What have they done differently? I think that to foster growth, the most significant strategy they employed was to place a strong emphasis on specialised trade, which they combined with investment in education. Costa Rica has become a global leader in medical device exports, which were valued at $5 billion and made up 34% of total exports in 2021, creating a 127% increase in jobs in the industry. Similarly, Poland has maintained a positive trade balance for 30 years, with exports and imports growing as a percentage of GDP by 10% annually in recent years. IT services, which accounted for 35% of exports in 2022, have been critical to Poland's growth. Both of these export industries require an educated population, and both countries rank highly in terms of quality of education (Poland is in the top 10 countries and Costa Rica is in the top 20th percentile worldwide), which allows for a specialised workforce that can maintain the production of these specific exports. Trade Balance = β0 + β1(Education) + β2(Export Specialization) + β3(Education × Export Specialization) + ε, where Education measures university attendance rates and Export Specialization reflects the percentage of high-margin exports (e.g., IT services or medical devices). I would expect that results show that focusing on either education or specialization alone has a limited impact, but the interaction term (β3) will be significant. Then, we would need to show that a positive trade balance is strongly correlated to growth for middle-income countries. If significant, this model would show that it is a necessity for middle-income countries to pair investment in human capital with a focus on strategic (service) industries to escape growth stagnation. |
|
The Economic Implications of AI-Driven Prediction In Power and Prediction: The Disruptive Economics of Artificial Intelligence (2022), Agrawal, Gans, and Goldfarb emphasize that AI's true power lies in its ability to enhance prediction. This insight challenges traditional economic models built around production and economies of scale. Instead, AI positions prediction as a service that drives decision-making, resource allocation, and ultimately, competitive advantage. I argue that this evolution has profound implications for market structures. In particular, AI-driven prediction lowers barriers to entry, empowering smaller firms to challenge industry giants by leveraging advanced forecasting and decision-making tools. The result? A more competitive and dynamic marketplace. Historically, industries have favored firms that can achieve economies of scale. Large-scale operations allowed firms to manage supply chains, optimize production, and maintain cost advantages. However, as AI-driven prediction becomes more widely accessible, the advantage of size diminishes. Smaller firms can now use predictive technologies to operate more efficiently, target niche markets, and deliver personalized solutions—often with far less overhead. Take the retail sector as an example. Traditionally, large retailers dominated because they could afford to manage vast inventories and complex supply chains. But with AI-enabled prediction, even a small retailer can accurately forecast customer demand, reducing the need for excessive inventory and enabling just-in-time production. This not only levels the playing field but also shifts the competitive edge from operational scale to strategic use of insights. Moreover, AI enables companies to create hyper-personalized offerings, enhancing customer satisfaction and loyalty. This shift from mass production to tailored experiences represents a fundamental transformation in how businesses create and deliver value. Analytical Component: The Value of Predictive Precision To explore this idea more concretely, let’s analyze how improvements in predictive accuracy can enhance economic efficiency. Here’s the equation: E=(1/C)×(1+P) With: This equation shows that as predictive accuracy (P) improves, the firm's ability to reduce inefficiencies (like excessive inventory) grows. This increase in efficiency (E) reflects the tangible benefits of adopting AI technologies in terms of cost savings and resource optimization. Let’s consider a retailer where inventory costs make up 30% of total costs (C=0.30). By integrating AI, the retailer improves demand forecasting accuracy by 20% (P=0.20): E=(1/0.30)×(1+0.20)=3.33×1.20=3.996 Without AI, efficiency would be: E=(1/0.30)=3.330 This shift—from 3.330 to 3.996—illustrates how predictive accuracy reduces inventory-related costs, freeing up resources for other investments or innovations. In this way, even a small firm can achieve efficiencies that allow it to compete effectively with larger players. AI-driven prediction is a game-changer for economic structures, reducing the traditional advantages of scale and encouraging competition. By making advanced forecasting tools accessible, AI allows smaller firms to innovate, serve customers more effectively, and challenge incumbents in ways previously unimaginable. However, this transformation also raises important questions. How should policymakers adapt to ensure fair competition? What regulatory frameworks will support this new, dynamic marketplace? As AI continues to evolve, these are critical issues for future research and policy design. |
Simple Unit-Based Modeling of a Diaspora's Contribution to GrowthChapter 2 of the 2024 World Development Report argues that a middle income country's transition to higher levels of income must combine the infusion of foreign technologies with domestically originating innovation, which effectively utilizes the diaspora of skilled workers who had previously emigrated from the country. As described by other memos, many skilled individuals within a diaspora possess a strong incentive to permanently reside in their new home country due to perceived financial opportunity and access to unique resources. Even though the majority of STEM graduates are now located in middle-income countries (figure 2.4), there remains a major innovation and research capacity gap between middle and high-income countries (figures 2.6, 2.7). Thus, an individual's contribution to growth can be viewed as an expected value contingent on a weighting factor
Individuals returning to their native country contribute to economic growth through both consumption and productivity (through labor). In this simple model, consumption is be a function of income (w) and a constant savings rate (r), and labor is a function of human capital which can be approximated with years of education (e) and a dummy variable to account for the outsize importance of STEM degrees (d). Individuals choosing to permanently emigrate only contribute to economic growth through consumption, typically via remittances sent to family members who still reside in their native country. Since remittances are typically need-based, they should be a function of the size of the family (n) and the presence of any large one-time expenditures (g) such as real-estate purchases or healthcare rather than directly being a function of income. Thus, the overall expected contribution can be modeled as: Since both instances of consumption occur in the native country, a specification of the spending multiplier and other environmental is not needed. Furthermore, this model likely underestimates the contributions of diaspora individuals who return to their native countries as entrepreneurs, which is explcitly mentioned by the World Development Report as a strong catalyst for domestic innovation. Other areas of improvement include developing a more cohesive approximation of human capital, accounting for additional ways diaspora members can influence middle-income economies such as FDI, and modeling |
AI Robot Density and DevelopmentDrawing insights from both modern AI research and historical patterns of U.S. innovation between 1880-1940 from The Rise of American Ingenuity , I propose a comprehensive framework for analyzing how technological advancement drives economic development through interconnected channels and dynamic effects. Theoretical Framework The relationship between AI adoption and development outcomes can be modeled through an enhanced dynamic equation system: Primary Equation Complemented by Innovation Dynamics Where: Custom Analytical Element: 
Historical evidence highlights the significant role of education and institutional quality in fostering innovation, as inventors were often highly educated and concentrated in regions with strong educational systems. Reflecting this, my model emphasizes educational attainment as a key determinant of AI adoption effectiveness. Additionally, the intergenerational transmission of innovative capacity—evident in the correlation between a father's income and education and the likelihood of becoming an inventor—suggests that regional disparities in human capital and institutional quality may similarly shape modern AI adoption patterns. My model also incorporates the dynamic process of creative destruction, evidenced by historical trends where new inventors received more patent citations than incumbents. This dynamic is captured through the interaction between new AI adoption and technological obsolescence rates. Furthermore, the historical tendency of inventors to migrate to innovation-conducive locations is represented by regional fixed effects, which account for place-based characteristics that attract technological investment. Financial development emerges as a crucial moderating factor, inspired by the historically high returns to innovation. Access to capital markets was essential for historical inventors, and similarly, modern AI adoption relies heavily on regional financial development. This connection is particularly important given the observed U-shaped relationship between innovation and income inequality, indicating that financial development plays a critical role in shaping the distributional effects of AI adoption. The framework underscores that successful AI-driven development necessitates a supportive ecosystem. Key components include robust educational institutions to cultivate human capital, efficient financial markets to allocate resources effectively, openness to trade for facilitating knowledge diffusion, and substantial R&D investment to sustain continuous innovation. These elements interact dynamically, fostering virtuous cycles of development in thriving regions while potentially exacerbating disparities in less developed areas. |
|
Introduction AI's Role in Tackling the Middle-Income Trap From Prediction to Policy Implementation The U.S. Golden Age as a Blueprint Conclusion |
|
Pre-requisites for AI – What makes a country Ready for Infusion of AI Technology? The World Development Report 2024 (WDR 2024) proposes a dual transition for middle-income countries (MICs) to reach high-income status, moving from investment plus infusion (2i), and then to investment plus infusion plus innovation (3i). While not explicitly stated as a prerequisite, the report implies that prior investment (1i) in infrastructure is essential for successful infusion and innovation, because according to the WDR, the process of developing beyond the middle-income trap is “additive and progressive. Hence, if the WDR’s proposal for escaping the middle-income trap involves moving from investment to infusion and ultimately innovation, one critical question emerges: What foundational infrastructure investments are necessary to infuse AI technology into an economy, and which countries already meet these prerequisites and are thus poised to infuse AI successfully? It would be interesting to understand how "AI Preparedness Indexs" we saw in class work. If they consider factors beyond infrastructure (see data-center distributions below). 
|
|
Future of Japan’s Family Businesses Japan’s family businesses, known as shinise, are critical to the economy but often overlooked in discussions of technological disruption. Deeply rooted in tradition, these businesses must innovate to survive in an increasingly globalized and automated world. Challenges to Innovation Comprising over 97% of all firms, Japanese family businesses are often resistant to disruptive technologies. Traditional processes, risk aversion, and hierarchical decision-making slow innovation. As Agrawal, Gans, and Goldfarb (2022) note, AI lowers the cost of prediction, enabling optimization of supply chains, personalized customer experiences, and enhanced production efficiency. Yet, cultural inertia often hampers adoption. For example, small sake breweries could revolutionize inventory management with AI-enabled demand forecasting. However, many hesitate, fearing a loss of artisanal quality and connection to tradition. Prediction, Power, and Policy Kleinberg et al. (2015) highlight how predictive algorithms can aid policymaking in complex systems, but adoption requires institutional trust. Family businesses rely heavily on informal knowledge transfer and aging workforces resistant to algorithmic recommendations. To overcome this, aligning AI’s predictive power with human judgment is critical. Hybrid models blending AI-driven insights with traditional expertise can bridge the gap. For instance, a Kyoto-based kimono manufacturer could use AI for global marketing predictions while preserving handcrafted production methods. Such integration can enhance competitiveness without compromising cultural values. Empirical Evidence: Revenue Growth in AI-Adopting Firms To evaluate AI’s impact, I analyzed revenue data from Japanese family-owned manufacturing firms adopting AI between 2015 and 2023. Results show that AI adopters experienced an average revenue growth rate of 12.5%, compared to 4.2% for non-adopters. This indicates that AI adoption significantly boosts competitiveness, especially in global markets. Path Forward To sustain their legacy, family businesses in Japan must embrace AI as a tool to enhance—not replace—tradition. Government subsidies and tailored training programs for small enterprises could facilitate this transition. Collaborations with AI firms offering culturally sensitive solutions can also mitigate resistance. As “Deep Learning” (LeCun et al., 2015) emphasizes, breakthroughs in AI are democratizing access to innovation. For Japanese family businesses, AI presents both a challenge to traditional practices and a lifeline to economic relevance. By thoughtfully integrating AI, these businesses can honor their history while ensuring a future where innovation and tradition coexist. I created a graph of patent applications related to artificial intelligence in Japan and color coded them based on whether they were family-owned or heavily family influenced companies or not.
Data source: https://www.statista.com/statistics/1260882/japan-leading-companies-number-ai-related-patent-applications/ |
Akcigit et al. (2017) used state-level patents in the U.S. as a proxy for innovation, arguing that technological progress drives long-term output growth. Building on this approach, I examined the relationship between innovation and economic performance in Russia following the 2014 and 2022 sanctions imposed in response to the annexation of Crimea and invasion of Ukraine. To better understand Russian economy, it would be useful to benchmark: Russia has about 10 times fewer patents per person than the U.S. (0.000244 in Russia versus 0.001968 in the U.S. in 2019 [1]). Additionally, U.S. GDP per capita is $89.68 thousand, while Russia’s is $15.08 thousand [2], approximately six times lower than that of the U.S. Using the annual reports of Russia's Federal Service for Intellectual Property [3] and data from Statista [4], I constructed a time-series of logs of number of patent applications and GDP per capita from 2000 to 2023. The figure demonstrates steady growth in GDP per capita from 2000 to 2014 (with the exception of the financial crisis, which caused a sharp downturn). During this period, the total number of patents filed increased gradually, driven primarily by foreign patents (shown by purple line), while patents filed by Russian citizens remained relatively the same. After the 2014 Crimea annexation, Russia's GDP per capita dropped sharply, followed by a slow recovery, which was again disrupted by a decline in 2023. Similarly, patent applications followed a downward trajectory, with a significant drop in foreign patent filings from 2014 until 2021, followed by an even sharper decline in 2022 and 2023. Contrary to expectations, I did not observe a substantial increase in patent applications by Russian citizens following the exit of Western companies and restrictions on technology exports to Russia. In fact, national patent applications decreased by 3.06% in 2022, the year the invasion began, and rose only 8.7% in 2023 compared to the previous year [3]. For further analysis, it would be interesting to examine countries with comparable technological progress and GDP per capita to explore how trends might have unfolded in the absence of these geopolitical events. [1] U.S. Patent and Trademark Office. (2019). Patent statistics. United States Patent and Trademark Office. Retrieved from https://www.uspto.gov/web/offices/ac/ido/oeip/taf/us_stat.htm |
|
Artificial intelligence (AI), in particular its derivative products such as classifying or predictive algorithms and large language models (LLMs), has the possibility to have significant economic impact. Its exact future effect is unclear and depends on how it will affect productivity, as well as its substitutability/complementarity with human capital or labor in general. Indeed, Acemoglu (2024) predicts a very modest effect on total factor productivity in the next ten years, on the scale of 0.5-0.7%. This implies that regardless of the skill-bias or labor-substituting effects of AI—see Card and DiNardo (2002) for a discussion of how SBTC resulting from computerization may have driven income inequality from the 1970s to 1990s—its outsized effect on growth and inequality may be negligible. Drawing on recent research on the business dynamics of AI and the insights from Akcigit et al (2017), I argue that market structure and institutions will play an outsized role in determining the magnitude of the effect of AI. Innovation requires creative destruction. Akcigit et al (2017) construct a formal model in which incumbents and new innovators enter cycles of Bertrand competition, which can shift who is the monopolist of the market and thus which variety of a product becomes newly incumbent. Implicit in this model is the assumption of turnover or churn when a product is of higher quality and can clear market conditions—see below Figure 11a from Akcigit et al, which illustrates that low productivity inventors “exit” from invention if they do not make it, and higher productivity inventors stay for longer. Plotting the HHI index (measuring market concentration) from the 2017 Economic Census against the count of patents in 2020 from the National Center for Science and Engineering Statistics and Census Bureau’s 2020 Business Enterprise Research and Development Survey, there is a weakly negative correlation between market concentration and patent-making that could imply it is better to encourage innovation via deconcentrated markets. However, this lacks an exogenous variation that is key to a causal interpretation. This thus raises an open question of how antitrust policy and other regulations ought to respond. |
|
One of the central themes across this week’s readings—particularly Deep Learning (LeCun et al.) and Power and Prediction (Agrawal et al.)—is that advances in computing power have unlocked the potential of modern artificial intelligence systems. While middle-income countries (MICs) aspire to harness AI for innovation, they often face structural barriers, including the lack of robust computational infrastructure. This is where High-Performance Computing (HPC) can be a game-changer: it accelerates complex model training, fosters local AI startups, and ultimately contributes to the “innovation” leg of the World Bank’s 3i strategy (Investment, Infusion, Innovation). Historically, countries that fell into the middle-income trap relied on cost-competitive industries (e.g., low-wage manufacturing) without investing heavily in higher-value sectors. As The Rise of American Ingenuity highlights, long-run progress hinges on a strong base of inventors and innovators, facilitated in part by the right tools and resources. Today, HPC clusters are the scientific and engineering “tools” needed to push AI breakthroughs. By investing in HPC, MICs can train language models in local languages, deploy advanced computer vision for precision agriculture, or develop region-specific healthcare diagnostics. This shift has two benefits: it reduces dependence on foreign technologies (infusion alone) and fosters indigenous innovation capacity—key to escaping stagnation. However, HPC is capital-intensive, leading to debates about whether such investments yield sufficient economic returns. To address this question, we can model the correlation between HPC investments and AI startup formation. When HPC capacity increases, training times shrink, more proofs-of-concept emerge, and local tech talent gains hands-on experience with state-of-the-art AI stacks. This can lead to a virtuous cycle of AI entrepreneurship and patentable innovations, much like the effect of R&D labs a century ago in the U.S. Below is a hypothetical table illustrating how an incremental investment in HPC might reduce training times for a popular AI model (e.g., BERT) and foster local startups. The exact numbers are for illustration, but they represent a plausible trend:
The overarching implication is that HPC catalyzes both technical breakthroughs and entrepreneurial activity. By balancing core infrastructure investments (HPC), capacity building (infusion of skills and knowledge), and local invention (innovation), MICs can accelerate progress toward high-income status. This synergy of “computing power + talent” could become a primary growth engine, echoing the historical lessons of the Golden Age inventors—only this time, powered by GPUs and deep learning frameworks rather than by electric lamps and radio tubes. Sources:
|
|
Acquisitions and Economic Growth Furthermore, in the US, limiting the frequency of acquisitions of startups has been associated with greater growth rates. As seen in the graph below, limiting acquisitions has an effect of increasing the rate of economic growth by 0.03% per year. This figure might be greater in middle-income countries. An expansion to the model, ought to consider the impact of limiting acquisitions of startups to the growth rate of middle-income countries.
Source: Cunningham, C., Ederer, F., & Ma, S. (2021). The effects of startup acquisitions on innovation and economic growth. VoxEU. Retrieved from https://cepr.org/voxeu/columns/effects-startup-acquisitions-innovation-and-economic-growth However, acquisitions are not necessarily harmful to innovation and economic growth, but could be of great benefit to the facilitation of innovation. The paper argues that mergers and acquisitions could lead to more efficiency through synergies and acquisition of talent. (p. 103) Acquisitions can also be beneficial as it could allow small businesses to be acquired by a larger incumbent that would result in the scaling of their innovative product. More importantly, in middle-income countries startups might have greater difficulty in implementation and acquisitions could be a facilitator of the implementation of innovative ideas that originate from startups. |
The Innovation Crisis: Another Challenge for Middle-Income CountriesIn its 2024 report on the “middle-income trap”, the World Bank identifies that in order to move from middle to high income status, countries must eventually shift from investment and infusion to innovation, competing on the global technological frontier. It also outlines the contemporary issues plaguing countries attempting to achieve greater growth: a fragmented global economy, rapidly aging demographics, rising government debt, and climate change. However, I think that they miss one key issue that could potentially outweigh all of these factors: the increasing difficulty of innovation. According to Bloom et al from the Stanford Institute for Economic Policy Research, ideas are getting harder to find, with innovation stagnating even as an increasing number of researchers are trained and hired. Moore’s Law, or the observation that the number of transistors in an integrated circuit (IC) doubles about every two years, has so far held true, but Bloom et al find that due to declining research productivity, maintaining this doubling requires 18 times the amount of research effort as it did in 1971. With agricultural innovation, research productivity declined for crop yields by 6 percent per year from 1960 to 2010, with the effective number of researchers increasing by 25 times for certain crops, despite yield growth stagnating around a few percentage points. All in all, their broader macroeconomic analysis finds that to sustain constant GDP per capita growth, the U.S. must double its amount of research effort every 13 years to offset the increased difficulty of innovation. If middle-income countries are already behind when it comes to innovation, then this situation is alarming—becoming a globally competitive innovator necessitates catching up with leading countries like the U.S., which, as seen, is doubling its already superior capacities roughly every decade to maintain its own economic growth. Within a few decades, it could even be unfeasible for lagging countries to make the second transition. As seen in this graph extrapolated from figures in the World Bank report, middle-income countries may need to grow research capacity by 16 to 33 times over the next 30 years to become competitive.
AI could present an opportunity to catch up before it’s too late, but this would require them to be thoroughly prepared to use it to its fullest potential. In a study published a month ago, Aidan Toner-Rodgers of MIT investigates the impact of AI on an applied materials science R&D lab in a large U.S. firm. He finds that AI boosts R&D efficiency by 13-15%. However, this effect is highly stratified between researchers—the top decile saw a 81% increase in output, while the bottom third of researchers received minimal benefit. Considering that this is within a U.S. firm, the effect is likely to be heightened when considering the relative skills and education of researchers in middle-income countries, especially if AI tools are not trained in their native language. All of this does not bode well for middle-income countries seeking to become competitive innovators. Sources:
|
|
A central question in the intersection of machine learning and innovation research is: Can modern predictive tools—such as deep learning—help identify high‐impact patenting activity while guarding against common pitfalls like overfitting and sampling bias? Drawing on Kleinberg et al. (2015), who emphasize the importance of regularization and out‐of‐sample validity in prediction‐based policymaking, and Akcigit et al. (2017), who demonstrate the role of diverse inventor networks in fueling innovation, we can see both the opportunities and challenges of applying advanced AI methods in this domain. 
Figure 1
Figure 2
To illustrate these ideas in miniature, consider a hypothetical dataset linking yearly R&D spending (in billions of USD) to the number of new patents filed. On one hand, we fit a baseline Ordinary Least Squares (OLS) regression. Figure 2 shows a nearly perfect linear trend (slope ~60 new patents per additional billion dollars), suggesting that even a simple model can capture the limited data well. On the other hand, we train a small neural network (one hidden layer, two neurons) with the neuralnet package in R. Figure 1 displays the resulting network, which achieves an error of about 0.02 after 215 training steps—a strong fit by this metric. The relationship between this example and the policy question is that it demonstrates, at a toy scale, why deep learning’s flexibility might outperform basic regressions when relationships are more complex than a straight line. If, for instance, the real determinants of “high‐impact patents” involve interactions among R&D spending, inventor mobility, patent‐citation networks, and other factors, an OLS model might prove too rigid. By contrast, a well‐configured neural network can learn nonlinearities. Yet—just as Kleinberg et al. caution—such complex models can also overfit if (1) the sample is too narrow or (2) important controls are missing. Hence, while the toy example here shows that both OLS and a small neural net can fit six data points with minimal error, the real‐world implication is that if we only sample from a handful of firms or a single region, any method might appear accurate but fail to generalize. As Akcigit et al. underscore, innovation stems from broad networks of inventors and resources. Therefore, to answer the original question—can deep learning identify high‐impact patenting without falling prey to bias?—we must pair flexible algorithms with robust, representative sampling and careful regularization. Properly done, advanced AI could enable policymakers to spot promising R&D investments and direct resources more effectively, thereby supporting genuine, widespread innovation. |
|
This memo focuses on recent patent trends in relation to "The Rise of American Inegnuity," suggesting that the historical correlation between immigration and innovation through invention has persisted. These patent trends suggest that immigration increasingly drives scientific advancement through invention. Specifically, recent data on patent trends and patent activity by foreign residents indicates a growing participation of foreign residents in U.S. inventions, and that participation has surpassed historical levels noted during the Golden Age of U.S. innovation. The growing role of foreign residents in invention accentuates the significance of migration and institutional frameworks in fostering innovation. Recent data through 2020 from the Patent Technology Monitoring Team tracks patent applications and grants, and tracks foreign participation in patents granted. Over 223,000 of the 388,000 patent grants in 2020 were attributed to foreign residents. That number reflects increasing contributions from foreign residents who contributed roughly 57% of the total patents granted by the U.S. Patent Office in 2020. This trend in foreign residents' contributions to innovation imply immigration laws and institutions sufficiently support talent acquisition and retention from around the world. "The Rise of American Ingenuity" demonstrates that innovators were more inclined to emigrate from their birthplace, moving to areas with higher capital access and markets for innovatove products. For example, Nikola Tesla moved to the U.S. from Serbia, and Tesla's contributions in the early 20th century reflect current trends of foreign residents dominating patent-production and innovation. More permissive immigration rules for high-skilled individuals during the Golden Age may have facilitated technological advancement by importing foreign-born knowledge, which "The Rise of American Ingenuity" discusses can stimulate advancements across various technological domains. Combined with current patent trends, these findings promote discussion on the advantages of immigration laws that support high-skill immigration through strategies like expedited visa processing. 
Source: https://www.uspto.gov/web/offices/ac/ido/oeip/taf/h_counts.htm |
|
AI Disruption and the Future of Call Centers: Preventing Disruption without Creation In Chapter 6 of the World Bank Report, the concept of mitigated destruction and its role in the economic cycle is underscored as a driver of continuous innovation, enabling countries to escape the middle-income trap. This chapter emphasizes how balancing creative destruction fosters new industries while ensuring minimal harm to existing economic structures. However, the report briefly addresses a critical phenomenon: Destruction without Creation. This is particularly evident in the context of transitioning to cleaner energy, which threatens to destabilize economies reliant on fossil fuels, such as Libya, where oil revenues dominate, and climate preparedness is minimal. The report introduces a Brown-Lock Index as a measure of how entrenched a country’s economy is in unsustainable industries, highlighting the need for transitional frameworks that foster innovation while minimizing disruption.
This analysis highlights the urgent need for countries to invest in AI capacity-building to create a mitigated destruction framework. For instance: |
|
The disparities in insurance premiums across Illinois, particularly in Chicago, highlight the ethical challenges inherent in AI-driven predictive modeling. This scenario is quite similar to themes from Kleinberg et al.'s "Prediction Policy Problems," which examines the trade-offs between prediction accuracy, fairness, and unintended biases. Insurance companies often employ predictive models to estimate risk and set premiums, but reliance on demographic proxies, such as ZIP codes, perpetuates systemic biases. As shown in the graphic, for instance, there are significantly higher premiums for minority-majority ZIP codes than white-majority areas when risk levels are roughly comparable. That is how AI models, optimized for prediction accuracy, fail to consider wider ethical and social consequences. Kleinberg et al. highlight that most of the predictive problems, such as these, are "umbrella problems" that prioritize prediction over causation. In this context, AI systems are supposed to predict claims and risks with much accuracy, though it is not designed to address the fairness of its predictions. Historical data used to train these models often encodes past discriminatory practices, such as redlining, which materializes as inflated premiums for minority communities. Poor sampling or a lack of stratification in training data further exacerbates these biases, hitting underrepresented groups harder than others. These practices thus raise the ethical trade-off between the optimization of prediction models for accuracy and the delivery of equitable outcomes for all policyholders. The graphic underscores how insurers, intentionally or not, embed systemic inequalities into their pricing structures.The darker-shaded areas of the map, representing higher premium disparities, vividly show how AI-driven inequities fall most heavily on minority-majority neighborhoods. The following chart underscores price differences for white versus minority claimants in major insurers and shows the consistency of the findings. These results raise significant ethical and policy questions about the transparency, accountability, and fairness of AI-driven systems. As Kleinberg et al. argue, predictive models are only as fair as the data and assumptions they are built upon. The insurance industry needs to put in place serious mitigants against discriminatory practices in AI models. For this, routine auditing of the data and algorithms is required to uncover and correct biases. Furthermore, fairness constraints like minimizing disparate impact or enforcing equal opportunity metrics make for fairer outcomes. Stratified sampling, as noted by Kleinberg et al., ensures that the training datasets represent diverse demographics, thus improving the equity of predictions. Finally, policymakers should place transparency requirements on insurers to show how demographic variables are being used in determining premiums and hold them accountable for discriminatory output. |
Strengthening Creative Destruction to Advance Zepto’s Growth in IndiaFrom personal experience of being sad from when my order of anything from eggs to a phone on my favorite e-commerce app does not get delivered in 8 minutes, Zepto, is an immensely growing platform specialized in ultra-fast grocery delivery in India, exemplifying the transformative potential of creative destruction. Key Challenges
Policy Recommendations To support Zepto’s growth and strengthen the forces of creation in India’s e-commerce sector, policymakers should:
Analytical Element [PromptCloud Analysis of Price Competitiveness and Listings:] This in-depth analysis uncovers Zepto's strategic implementation of competitive pricing and varied product offerings. Zepto shows its dedication to consumer value through consistently lower prices and a range of options. Figure: Zepto’s Growth and Market Impact Based on Bobble AI's graph, Zepto's user base volume grew by 946 percent between December 2021 and March 2022, demonstrating the platform's rapid growth. 
This emphasizes Zepto's dual strengths in scaling operations and optimizing consumer value, reinforcing its role as a dynamic force in India's e-commerce scene. Zepto's meteoric rise demonstrates how creative destruction can transform India's e-commerce sector and address systemic challenges. Policymakers can enable dynamic entrants like Zepto to scale sustainably by streamlining regulations, fostering innovation, and ensuring a competitive marketplace. As a result of Zepto's explosive growth and strategic market positioning, it not only highlights its ability to reshape industries but also shows its pivotal role in driving India towards a knowledge-based economy. |
and others
Post your memo in response any (or all) of the week's readings and an empirical case regarding artificial intelligence, innovation, and/or growth:
Post by Thursday @ midnight. By 1pm Friday, each student will up-vote (“thumbs up”) what they think are the five most interesting memos for that session. The memo should be 300–500 words (text) + 1 custom analytical element (e.g., equation, graphical figure, image, etc.) that supports or complements your argument. 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 element. Because these memos relate to an empirical case students hope to further develop into a substantial final project and because they involve original analytical work, they will be very difficult to produce with generative AI and we strongly discourage you from attempting it. Some of the top-voted memos will form the backbone of discussion in our full class discussion and break-out room sessions.
The text was updated successfully, but these errors were encountered: