Week 4: Who Invents and Innovates

[jamesallenevans]

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

“Tapping into Talent: Coupling Education and Innovation Policies for Economic Growth,” 2024. Ufuk Akcigit, Jeremy Pearce, and Marta Prato. Review of Economic Studies, forthcoming.
“Brains and Business: How Inventors and Entrepreneurs Shape Economic Progress,” 2025. Ufuk Akcigit, Harun Alp, Jeremy Pearce, and Marta Prato. Working Paper.
“Measuring the Characteristics and Employment Dynamics of U.S. Inventors,” Ufuk Akcigit and Nathan Goldschlag, Journal of Economic Growth, forthcoming.
“Aging and the Narrowing of Scientific Innovation Download Aging and the Narrowing of Scientific Innovation”. 2025. Haochuan Cui, Yiling Lin, Lingfei Wu, James Evans. Preprint.
“Large teams develop and small teams disrupt science and technology.” 2019. Lingfei Wu, Dashun Wang, and James A. Evans. 2019. Nature 566(7744): 378-382.
“Flat Teams Drive Scientific Innovation.” 2022. Fengli Xu, Lingfei Wu, James Evans. Proceedings of the National Academy of Sciences.
“Remote collaboration fuses fewer breakthrough ideas”. 2023. Yiling Lin, Carl Benedikt Frey & Lingfei Wu. Nature 623: 987–991.
“Who do we invent for? Patents by women focus more on women’s health, but few women get to invent.” 2021. Koning, Rembrand, Sampsa Samila, and John-Paul Ferguson. Science 372(6548): 1345-1348.

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

[zararania]

Background:
A key focus of this week's readings was about how both the gender and hierarchical composition of teams contributes to the quality and focus of scientific work produced. This reminded me of "Machines of Loving Grace," where Dario Amodei outlines how artificial intelligence has the capacity to greatly accelerate progress in various fields. Most notably, there are hundreds of scientific and social problems where "a large group of really smart people [can] drastically speed up progress." This is interesting to me because in the age of hypothetical superpowered AI scientists, who work as these groups of smart people, much like the work that we read this week, there is a question of what focus and quality of scientific innovation will be accomplished. Thus I am very curious about in what manner scientific innovation will continue to grow when AI scientists begin significantly contributing to the biological and medical fields. Since I do not know what that will look like, my research question is more limited to how composition of human teams, a human and AI team, and a team of solely AI agents might differ in the quality of scientific innovation.

Research Question:
Does the integration of artificial intelligence into biological research teams enhance or inhibit the generation of novel scientific ideas? I hypothesize that hybrid (human/AI) teams will produce significantly higher novelty than either AI-only or human-only teams. This is based on the premise that while AI-only teams may converge on established, high-probability solutions (low novelty), and human-only teams are limited by data processing capacity, hybrid teams combine the computational breadth of AI with the intuitive, disruptive leaps characteristic of small human teams.

Theoretical Framework:
This study synthesizes the "small teams disrupt" framework (Wu et al., 2019) with theories of recombinant innovation. We posit that AI-only agents, trained on biological repositories, will behave like large teams: highly efficient at exploiting known pathways (e.g., refining a known drug target) but risk-averse regarding low-probability hypotheses (i.e. low novelty). Human-only agents, analogous to small teams, possess high variance and creativity but lack the bandwidth to explore the "adjacent possible" of biological literature efficiently. The hybrid model is theorized to function as an augmented small team: the human provides the novelty bias (directing the search toward unexplored areas), while the AI provides the evaluative speed to verify these novel connections instantly, resulting in the highest net novelty.

Data Visualization:

This graph illustrates the trade-off between novelty (x-axis) and scientific impact (y-axis) across three different team structures. The blue cluster (AI-only) is tightly packed on the far left, showing that while AI agents efficiently optimize known science to achieve high impact, they rarely venture into novel territory. Conversely, the red points (human-only) are widely scattered along the bottom, indicating that humans generate diverse and novel ideas but often lack the processing power to fully maximize their impact. Most importantly, the green cluster (hybrid) occupies the top-right "frontier," demonstrating that combining human intuition (to jump to new ideas) with AI optimization (to refine them) consistently produces the highest quality and most novel scientific breakthroughs.

[bheim]

How do state tax credits from R&D affect the quantity and quality of patent filings and who do they push to innovate?

Justification & Methods

Akcigit presented us with a useful model for thinking about how innovation policies should be enacted to promote growth. Since 1982, many states in the US have enacted tax credits for R&D, which reduce a company’s tax liability based on R&D spend.

R&D tax credits likely favor incumbents with strong financials more so than entrants. My analysis will use the creation of these tax credits to run a regression discontinuity design where we observe the quantity and quality (measured by citations) of patents before and after the initiation of the tax credit as well as the proportion of patents filed by new filers (a signal of new entrants).

The regression we’ll be using for this analysis is as follows:

$$ Y_{st} = \alpha + \beta \cdot [\text{RDCredit}_{st}] + \gamma_s + \delta_t + \varepsilon_{st} $$

where Y represents either patents or new entrant share (separate regressions) for state s, in year t. RDCredit is 1 if the state, s, had already or just adopted the tax credit at time t. Gamma and Delta are state and year fixed effects. Beta is the DiD estimate.

Results

From the results, we can observe that 1) patents were trending up in quantity as time goes on (expected), but I did find about a small increase in patents not explained by the general trend. 2) Patent quality increases up until the point of the credit. Once the credit is implemented, there’s a significant jump (perhaps a rush of R&D spend). 3) Lastly, these policies actually lead to an increase in the proportion of patents filed by new entrants (about a 1.6% increase). However, there’s a clear consistent decrease in the share of patents that new entrants are filing over time.

The data present a compelling picture for tax credits. Their main immediate effect seemed to improve patent quality and even push new entrants to file patents (a very unexpected finding but perhaps reflects larger companies choosing to spend more on R&D when they hadn’t before). However, there are certainly questions of whether the long-run effects of implementing these credits. While not shown above, I found that when interacting treatment with the passing years, we see a significant decrease in patent quality not demonstrated by those states without tax credits.

[HenryPinner7]

When governments aim to strengthen long-run innovative capacity under limited budgets, which education policy/investment generates greater innovation: lowering net tuition at public flagship universities (improving access to high-quality training) or expanding enrollment capacity at nonselective colleges (broadening overall access)?

Background: Education policy is a central lever for building an innovation-driven talent pool. The week’s readings (Ackigit, 2024) emphasize that innovative capacity is highly unevenly distributed and that education policy choices shape who enters the pipeline to become inventors and transformative entrepreneurs. However, public resources are scarce: governments rarely can fully subsidize all levels of education or simultaneously expand access and increase quality everywhere, making the allocation problem essential. I have thought of two potential approaches that reflect different theories of how innovation is produced:

1. Reducing net flagship tuition (a price/quality access-margin intervention), potentially shifting high-ability students into resource-rich settings with stronger STEM training, research exposure, and peer effects – raising the quality and impact of later invention.
2. Expanding nonselective public college enrollment (a quantity/access-margin intervention), increasing the number of degree-holders and potentially broadening the base of technical talent, even if average resources per student are lower.

Comparing these margins can reveal which education investments deliver the highest “innovation return” in inventor entry, patent quality, and innovative field choice.

Research Design: To compare the causal innovation returns to two education policy levers, I will construct a state-by-cohort panel (cohorts, c, will be defined by birth year/high school graduation year for each state s) linking policy exposure at ages 17-19 to later innovation outcomes (adults 30-40). Policy variables are (a) FlagshipNetTuition: the in-state net price of attendance (e.g., tuition/fees net of typical aid) and (b) NonselectiveCapacity: an index of broad-access capacity (e.g., first-time enrollment or seats per college-age resident, student-faculty ratios, or instructional spending). Outcomes Ys,c capture the resulting size and quality of the innovative talent pool, including inventor entry (share of the cohort ever named on a patent), innovation intensity/quality (patent counts, citations to gain a sense of novelty/disruptiveness), and related proxies such as STEM degree completion or startup formation when available.

Concretely, I would use the following DiD regression:

Ys,c​ = α + β1​FlagshipNetTuitions,c ​+ β2​NonselectiveCapacitys,c​ + controls + εs,c​,

where controls include state fixed effects, cohort (year) effects, and time-varying covariates (economic conditions, industry mix, and R&D spending). I will scale estimates into policy-relevant “returns” (e.g., innovation per $1M or per additional graduate) to compare which lever more efficiently fosters innovation.

Figure:

This figure plots event-study estimates of each education policy’s effect – flagship net tuition reductions versus nonselective capacity expansions – on a common inventiveness outcome over event time. The x-axis shows years relative to policy adoption (0 marks the first exposed cohort at college-entry age), and the y-axis reports the estimated change in inventiveness (e.g., inventors per 10,000 or a standardized index); points are event-time coefficients with uncertainty bands. This figure operationalizes the memo’s core ambiguity: does innovation respond more to improving access to high-quality training or to expanding broad access at scale? If the flagship series rises more (or sooner) than the nonselective series, flagship price reductions deliver higher innovation returns; if the nonselective series dominates, capacity expansions are the more effective lever. Normalizing both series to comparable policy units makes this a direct visual comparison of which lever produces larger innovation gains and how quickly effects emerge.

[ShiyangLai]

Shiyang Lai

Research Question.
How do fine-grained, city-level taxation policies shape the migration of inventive talent and, through this reallocation, alter the spatial distribution and productivity of innovation across regions?

Research Design.
This project develops a hypothetical policy laboratory using a spatially explicit multi-agent simulation to study locational taxation at a granularity that is largely inaccessible to empirical causal inference. While existing work shows that taxes influence the migration of inventors and star scientists across countries or US states, these studies are constrained by coarse policy variation and limited counterfactual scope. I instead propose a calibrated simulation framework that enables systematic exploration of city-level tax policies and their downstream innovation effects.

The core design embeds heterogeneous inventor agents into a map-based environment implemented in the AgentSociety framework. Cities are modeled as distinct jurisdictions characterized by local income tax schedules, wage opportunities by sector, housing costs, and innovation infrastructure (e.g., firms, universities, funding availability). Inventor agents differ in productivity, sectoral specialization, peer networks, risk tolerance, family constraints, and mobility frictions. Migration decisions are modeled as discrete choices driven by after-tax income, amenities, agglomeration forces, and moving costs.

Large language models are used in a constrained and interpretable role: rather than directly selecting actions, LLMs generate structured representations of agents’ perceived preferences, salient policy features, and boundedly rational interpretations of complex tax rules. These outputs parameterize an explicit random-utility migration model, ensuring that equilibrium talent allocation and innovation outcomes are governed by transparent mechanisms rather than free-form generation.

Innovation emerges endogenously from the spatial clustering and interaction of agents. City-level innovation output is modeled as a function of inventor productivity, diversity, network density, and congestion effects such as rising housing costs. This design is informed by readings on agglomeration, talent allocation, and place-based innovation policy, particularly work emphasizing superstar concentration and spillover-driven growth (e.g., Moretti; Akcigit et al.).

The simulation allows systematic policy sweeps—varying tax differentials, eligibility thresholds, and targeting strategies—to estimate migration elasticities, decompose channels (income, agglomeration, housing, salience), and trace how small fiscal differences can generate nonlinear spatial inequality in innovation. The goal is not prediction, but mechanism discovery: identifying regimes in which city-level tax incentives amplify innovation versus those in which they merely reshuffle talent.

Visualization.

[genevese]

Geneva Kirk Drayson

Based on the supplementary reading: “Too many papers? Slowed canonical progress in large fields of science Links to an external site..” 2021. Johan SG Chu and James Evans. Proceedings of the National Academy of Sciences.

Can language-model–based perplexity measures, computed relative to the existing scientific canon at a given historical moment, retroactively identify papers, topics, or subfields that later proved to be highly novel and impactful?

This project builds directly on recent evidence that language-model perplexity captures scientific surprise and predicts transformative outcomes (Zhang & Evans, 2025), while addressing the structural failure of large fields to surface novelty documented by Chu and Evans (2021). Whereas Zhang & Evans demonstrate prospective prediction—showing that high-perplexity papers published after a model’s training cutoff receive more polarized reviews, greater interdisciplinary reach, and are overrepresented among later breakthroughs—this study asks whether the same signal would have worked historically, before citation-based recognition emerged. This allows us to empirically determine whether Chu and Evans' suggestion
of integrating a widely-adopted measure of novelty into prestigious journals could promote less canonical and more innovative publications and field progress.

The empirical design is a retrospective natural experiment. For a curated set of landmark breakthroughs across disciplines (e.g., plate tectonics, PCR, backpropagation, CRISPR), we reconstruct the scientific canon as it existed immediately prior to each paper’s publication. (A synthetic dataset was used for the purpose of this memo - for details see the linked Methodology & Materials Google Doc). For each historical slice, we train or fine-tune a domain-specific language model exclusively on pre-publication literature, ensuring strict temporal separation to avoid leakage—mirroring the methodological safeguards emphasized by Zhang & Evans.

We then compute abstract-level perplexity using the same method as in Zhang & Evans for (i) the breakthrough paper, (ii) contemporaneous “control” papers in the same field, and (iii) papers that were initially dismissed or ignored. Outcomes are measured along three axes: long-run citation impact, disruption scores, and interdisciplinarity of downstream citations. The core test is whether papers that later reshape fields systematically occupy the high-perplexity tail at time of publication, even when short-run citations and reviewer consensus are low.

Crucially, the design incorporates two findings from Zhang & Evans. First, high-perplexity papers are known to generate evaluation uncertainty—longer and more variable review times, polarized judgments, and delayed citation uptake. We therefore test whether historically transformative papers display these same early-stage uncertainty signatures. Second, because perplexity produces bimodal outcomes (breakthrough vs. neglect), we condition analyses on eventual success, distinguishing productive novelty from noise rather than treating surprise as uniformly beneficial.

Finally, we stratify analyses by field size to test Chu and Evans’ prediction that canon ossification is most severe in large fields. The hypothesis is that perplexity should be most informative precisely where citation-based heuristics fail, offering a potential corrective to canon-locked evaluation regimes.

Further details on proposed methodology & materials: https://docs.google.com/document/d/1Q1Wvs6rPUTDUplg9XfXzjVQMvkcm1RK_93D6I6ZPTV8/edit?usp=sharing

[elektra773]

Elektra Makris

Research Question
Do gender imbalances in existing biomedical knowledge lead AI-driven discovery to disproportionately accelerate innovation in already well-developed, male-centered domains rather than in historically underdeveloped areas of women’s health?

Research Rationale
Innovation is not neutral: who participates and how knowledge is structured shape what gets discovered. This weeks readings showed that women inventors are more likely to produce health innovations benefiting women, yet remain underrepresented, contributing to gaps in attention to women’s health problems (Koning et al., 2021). At the same time, research on team production finds that large teams tend to develop established paradigms rather than explore neglected directions (Wu, Wang, & Evans, 2019). AI systems, trained on massive datasets of past biomedical research, can be understood as operating like a “super–large team,” learning patterns from what has already been studied. Work on the life cycle of scientific innovation further suggests that reliance on established knowledge narrows search toward dominant paradigms (Evans et al., 2025). Together, these insights imply that AI may be most effective in knowledge-dense domains and less likely to expand thin or underdeveloped areas: many of which concern women’s health. This project tests whether AI amplifies existing inequalities in the knowledge landscape rather than correcting them.

Research Design
I will define the pre-AI period as before 2015 and measure pre-AI knowledge density for each disease area using cumulative publication counts, citation depth, and clinical research activity. These measures will be normalized to capture how “mature” each domain was prior to widespread machine learning. Using MeSH disease terms and the Medical Text Indexer approach, I will classify conditions as female-specific, male-focused, or mixed, following methods similar to Koning et al. (2021).

To measure AI’s impact, I will compute AI-era innovation growth as the percent change in patents filed after 2015 within each classified biomedical domain. The core empirical test examines whether domains with higher pre-AI knowledge density experience larger post-2015 innovation growth, and whether this relationship differs systematically across female, male, and mixed domains.

The key hypothesis is that AI disproportionately increases innovation in already knowledge-dense areas. If women’s health domains tend to have lower pre-AI knowledge density, they should exhibit flatter relationships between prior knowledge and AI-driven growth. This would suggest that AI behaves like a paradigm-developing, large-team technology: highly effective at extending established lines of research but less likely to open up new directions in thinner knowledge areas (Wu et al., 2019; Xu, Wu, & Evans, 2022). In this case, AI could unintentionally widen gender gaps in biomedical innovation by reinforcing historical imbalances in what has been studied.

Figure 1. Simulated relationship between pre-AI knowledge density and AI-era innovation growth across biomedical domains. Each point represents a disease area categorized as female-specific (red), male-focused (blue), or mixed (purple). Slopes are steeper for male and mixed domains, illustrating the hypothesis that AI disproportionately accelerates innovation in already knowledge-dense areas.

[sarahkarn]

Sarah Karn

Research Question: Does artificial intelligence relax or reinforce the talent bottleneck in innovation?

A key insight of the existing endogenous growth literature is that innovation is constrained not only by financial incentives, but also by scarcity and allocation of talent. In Tapping into Talent, Akcigit, Pearce, and Prato show that expansions in educational opportunities does not lead to a proportional increase in innovation. Rather, the ability to create inventive output is bottlenecked by training capacity, the time-consuming process of developing high-skilled human capital, and the synergistic interactions within research teams. Thus, innovation scales nonlinearly with increases in high-skill labor.

The rapid advancements of artificial intelligence naturally motivate questions regarding this framework. Today, AI tools can perform tasks that typically require experience and coordination of multiple workers: code generation, experimental design, and even testing hypotheses. I pose the following question: does AI relax the talent bottleneck in innovation by providing access to inventive capacities, or does it instead reinforce existing concentration by disproportionately complementing institutions that already brim with elite talent? If AI lowers the cost of invention, it is possible that the relationship between talent and innovation could flatten. Alternatively, AI could intensify concentration of talent resources if it serves as a complement to scale, capital, and absorptive capacity.

The attached plot exhibits AI research intensity versus aggregate innovation output across countries.
X-axis: Number of AI-related academic publications (Stanford AI Index Report, ~2019–2021). Captures the depth of AI-specific human capital and research activity.
Y-axis: Annual patent applications (WIPO). Serves as a proxy for innovative output.

Interpretation of the Results:
The figure reveals a strong, approximately linear relationship between AI research intensity and aggregate innovation across countries. It is important to note, however, that this relationship is anchored by a small number of frontier countries (US and China), which are disproportionately responsible for global AI research and patenting activity. If we examine nations at the moderate end of the AI research spectrum (Germany, France, UK), we see that they are nonetheless clustered toward the lower end of innovation output. This suggests that AI primarily functions as an amplifier of existing innovative capacity. By the framework of Tapping into Talent, AI doesn’t seem to relax bottlenecks arising from the time required to cultivate skilled human capital and from complementaries within research teams. Thus, the relative ranking of innovation systems seems to be largely unchanged due to AI.

[TarunSethi13]

Tarun Sethi

Research Question: Do political donations distort the allocation of public R&D subsidies toward incumbent firms and established innovation hubs, thereby pushing innovation policy beyond the Schumpeterian optimum and reducing the rate of disruptive innovation?

Background:

Recent work in the economics of innovation emphasizes that technological progress depends not only on the level of R&D investment, but on who conducts innovation, where it occurs, and how research is organized. This week’s readings show that large, hierarchical teams tend to produce developmental rather than disruptive innovation, that aging scientists increasingly act as gatekeepers rather than paradigm shifters, and that innovative activity is increasingly concentrated in large incumbent firms and dense geographic hubs. Together, these patterns suggest that innovation systems may become less dynamically efficient even as aggregate R&D spending rises.

At the same time, the Schumpeterian growth model highlights that R&D subsidies are welfare-improving only up to an optimal level. While subsidies can correct underinvestment driven by knowledge spillovers and encourage creative destruction, poorly targeted or excessive subsidies may weaken competitive pressure by entrenching incumbents.

This memo explores a political-economy mechanism that may push innovation policy away from this optimum. Political donations by firms may influence not only regulation, but also the design and allocation of public R&D subsidies. Large incumbent firms are particularly well-positioned to capture such subsidies due to brand recognition, legal and compliance capacity, grant-writing expertise, and dense political networks. These advantages allow incumbents to absorb large subsidy allocations and navigate complex application processes more effectively than entrants. As a result, public R&D funding may flow disproportionately toward established firms and regions, increasing total R&D while unintentionally suppressing disruptive, entry-driven innovation.

Theoretical Sketch:

When public R&D subsidies are disproportionately captured by politically connected incumbent firms, entry and experimentation by smaller firms and peripheral regions decline. Because entrants are a key source of disruptive innovation in Schumpeterian models, this weakens creative destruction even as aggregate R&D spending rises, altering the selection mechanism through which new technologies replace incumbent ones.

At the same time, subsidy capture reinforces organizational and geographic concentration in innovation. Large incumbents tend to allocate subsidized R&D toward hierarchical research teams operating in established hubs, environments that favor incremental, developmental innovation over exploratory search. Political influence may therefore push subsidy policy beyond the Schumpeterian optimum: innovation inputs and citations increase, but disruptive breakthroughs decline, potentially lowering long-run productivity growth despite higher R&D investment.

Empirical Sketch:

Future work could combine firm-level political donation data with information on public R&D subsidy allocation and patent-based measures of innovation type. Donations could be linked to subsequent subsidy receipt, while innovation outcomes would be measured using patent disruption indices, team size, inventor seniority, and geographic concentration. The core question is whether politically connected firms or regions receive disproportionately more subsidies and whether this exposure is associated with lower rates of disruptive innovation, conditional on total R&D spending.

[erindner13]

Erik Rindner

How can AI-driven personalization enhance and increase in-stadium revenue for professional baseball teams beyond ticket sales?

Recent advances in artificial intelligence have expanded its role in professional sports beyond player performance and team strategy and into the business side of fan engagement and revenue generation. In Major League Baseball, teams increasingly view stadiums not just as venues for games, but as complex consumption environments where food, merchandise, and experiences drive significant revenue beyond ticket sales. AI-powered systems—such as dynamic pricing, personalized promotions through team apps, and demand forecasting for concessions—promise to help teams better understand fan behavior and optimize spending once fans are already inside the stadium. At first glance, this shift appears highly beneficial: AI allows teams to tailor offers to individual fans, reduce waste, and maximize per-capita revenue in ways that were previously impossible.

However, this innovation raises deeper questions about how value is created in live sports environments. Traditionally, in-stadium spending was driven by shared experiences, social norms, and spontaneous decision-making. Fans bought food or merchandise based on convenience, atmosphere, or emotional attachment rather than targeted incentives. AI-driven personalization may fundamentally reshape this experience by nudging fans toward specific purchases at specific moments, transforming the stadium from a communal space into a data-optimized marketplace. While this may increase efficiency and revenue, it also alters how fans interact with the game, with each other, and with the team’s brand. In this sense, AI does not simply enhance existing revenue streams but redefines the problem teams are trying to solve—from selling an experience to extracting maximum value from each attendee.

This concern aligns closely with Arthur’s argument that technologies evolve by recombining existing components and, in doing so, redefine what organizations consider success. Once AI-driven personalization becomes embedded in stadium operations, future revenue strategies are constrained by these systems, creating path dependence in how teams engage fans. Rather than experimenting broadly with atmosphere or pricing, teams may increasingly rely on algorithmic predictions of individual behavior. This mirrors the logic in Simulating Subjects, which emphasizes how systems trained on past behavior can standardize future outcomes, narrowing the range of experiences users are exposed to. Applied to baseball, AI-driven revenue optimization may increase spending on average while simultaneously homogenizing the fan experience.

Methods

To study this question, I propose a comparative research design examining in-stadium revenue patterns before and after the adoption of AI-driven personalization tools. Using publicly available team financial data, attendance figures, and case studies of MLB teams that have implemented advanced mobile apps or dynamic concession pricing, teams could be categorized by their level of AI integration. Key outcomes would include average in-stadium spending per fan, revenue growth independent of attendance, and variation in spending across different fan segments and game contexts. These outcomes could be compared across teams and over time to assess whether AI-driven personalization leads to sustained revenue growth or diminishing returns.

The key outcome of interest is not simply whether AI increases revenue, but how it reshapes the fan experience and the mechanisms through which value is created. AI may be highly effective at optimizing short-term spending, yet incomplete in capturing the long-term importance of atmosphere, loyalty, and communal enjoyment. In an ideal setting, AI would enhance fan experiences while supporting sustainable growth, rather than reducing live sports to a purely transactional interaction.

[johnab20]

Research Question

Does the commercial scaling of "See & Spray" AI technology represent a force of creative destruction that threatens the revenue models of incumbent herbicide manufacturers, and do these incumbents utilize non-productive strategies to mitigate the shift?

Research Design

This design explores the underlying human capital and organizational architecture that fuels the competitive struggle between AI-native entrants and chemical incumbents. Informed by the theories of Akcigit et al. (2024, 2025), I focus on the "Brains and Business" match—the idea that growth is driven by how effectively high-ability inventors are matched with firms that can scale their breakthroughs. By integrating the findings of Wu et al. (2019), which demonstrate that small teams are significantly more likely to disrupt existing technology while large teams perform incremental development, I propose a longitudinal assessment of "team density" and inventor backgrounds. This study will utilize patent metadata to compare the average inventor team sizes for AI-related patents at John Deere with the legacy chemical patent teams at Bayer. The intent is to test whether Deere’s "See & Spray" success is a product of lean, agile software squads that remain unburdened by the bureaucratic "large team" drag and "scientific narrowing" typically observed in aging firms with massive, multi-divisional R&D networks like Bayer (Cui et al., 2025).

The figure below provides a comparative conceptualization of the R&D workforce structure at both firms using 2025–2026 financial and patent data. It contrasts the total scientific scale of the incumbent against the disruptive intensity of the entrant to visualize the "structural disadvantage" of large teams. While Bayer operates a massive global network of 15,900 scientists, the visual suggests that this volume may lead to the developmental focus described by Wu et al., whereas John Deere’s software-centric engineering workforce, though smaller in raw scientific headcount, achieves a higher concentration of "innovation density" per inventor. This architecture allows Deere to leverage its 17,900 granted patents through a model of elite "Fellows" and specialized engineering hubs, illustrating a "flat team" advantage that facilitates the fusion of breakthrough ideas.

The comparison is anchored in the reported 2024–2025 R&D metrics for both firms, showcasing a stark divergence in human capital deployment. Bayer AG currently employs approximately 15,900 research scientists supported by a massive R&D budget of €5,860 million, a scale that aligns with the "Large Team" model of incremental development and defensive product maintenance. In contrast, John Deere manages a leaner, engineering-heavy workforce of approximately 5,000 specialists whose output is reflected in over 17,900 granted patents as of late 2025. By analyzing "Investment Density"—the R&D spend allocated per specialist—the visual illustrates that Deere invests more aggressively in its individual "brains," supporting the hypothesis that smaller, high-density teams are the primary engines of the creative destruction currently threatening legacy chemical revenue streams.

John Butka

[nbcodes57]

Nick Baumgartner
Research Question:
Do state tax rate changes affect utility patent activity in neighboring border counties?

Motivation:
Prior research demonstrates that inventors respond to personal income taxes with an elasticity of 0.6-0.7 (Akcigit et al., 2018), and that out-of-state mobility responses exceed in-state responses by more than tenfold. However, the literature has largely focused on own-state tax changes, leaving neighbor state externalities—the pull effects of tax competition—underexplored. I propose a difference-in-differences design using border counties to isolate these cross-state spillovers.

If state income taxes significantly impact innovation location decisions, we should observe measurable shifts in patenting activity when one state raises taxes relative to its neighbors. Specifically, counties in lower-tax states bordering higher-tax states should experience relative increases in utility patent filings after tax hikes.

Data and Design:
I use USPTO utility patent counts by Metropolitan Statistical Area from 2000 to 2015. I will use this data to analyze neighboring counties after significant tax events. I aim to flesh out how to incorporate many different case studies across the US into a single analysis but to start I chose New Jersey's 2004 "millionaire tax" (which raised the top rate from 6.37% to 8.97%) to test whether patent activity shifted toward neighboring Pennsylvania counties. New Jersey's June 2004 tax increase created a substantial tax differential with Pennsylvania (flat 3.07% rate). I focus on cross-border MSAs spanning both states: Philadelphia-Camden-Wilmington (PA-NJ-DE-MD) and Allentown-Bethlehem-Easton (PA-NJ). These regions share labor markets and infrastructure, making them ideal for isolating tax effects from other confounders.

If taxes drive innovation location, we expect to see: (1) declining patent shares in NJ portions of border MSAs post-2004, (2) rising patent shares in PA portions, and (3) no such divergence in control regions without tax changes. The attached visualization plots total utility patents in NJ-PA border MSAs from 2000-2015, with a vertical line marking the 2004 NJ tax increase.

Results:

Pennsylvania utility patents show marginal increase but remain relatively flat, while the NJ counties show a post-tax-hike decrease albeit seemingly part of a previously existing trend. While neighboring counties may be compelling, this case study makes me hypothesize that inventors are not heavily geographically bound, and instead of moving to nearby counties they would opt for innovation heavy hubs across the country (agglomeration). However, I look to continue to explore this analysis as this is just one small case study and I hope to flesh out more across both geography and time span.

[verachaudhry]

Vera Chaudhry

Research Question:

How does combining a PhD slot expansion and education subsidy impact the optimal policy bundle? With a combined policy, what is the optimal number of PhD slots to add and what is the optimal tax policy?

Motivation [“Tapping Into Talent”]: First, we observe that high IQ individuals who pursue PhDs are likely to become inventors. I will focus on two of the policies in the paper. (1) Education subidy – the paper shows that the education subsidy increases the quality of researchers because it enables financially unable children with high IQs to pursue research. (2) PhD Slot Policy – increasing the number of PhD slots reduces the quality of the marginal researcher, presumably because now, children who were smart but not previously financially able can now get a PhD. This model ignores a pivotal question – if PhD spots increase, but educational subsidies are also provided such that the new students admitted to the pool have a high IQ, what are the implications?

Research Design
In the original state of the world, the ability threshold decreases with new PhD spots (graph 1) and this becomes more drastic if the tax increases dramatically and the returns to becoming a PhD decrease significantly. Similarly, education increases the ability IQ only to the extent that the returns to getting the PhD do not become de-incentivizing (Graph 2).
Given a fixed Ns you want to add, I solved the existing model for the education subsidy you would need to provide to not decrease the ability threshold. The solved model [typed in LateX] can be seen here:

Note that the subsidy this model solves for is an underestimate because with taxes to fund this program (this current solution assumes that taxes are a negligible part of the GDP and don’t need to be included), the ability threshold declines.

I propose a new theoretical model that can be solved for the optimal tax rate if new PhD slots were fully funded. I.e. the government raises money through a proportional tax to fund new PhD slots. I use the model to solve for (a) optimal tax rate T and (b) the number of PhD slots with the following design:

Tax revenue:
T(τ) = τ · π_I · Ā
Spending to fully fund ΔN = N − N₀ new slots:
S_slots = I^PhD · Ā · (N − N₀)
The balanced budget T(τ) = S_slots shows that slots financed by τ follow:
N(τ) = N₀ + (τ · π_I) / I^PhD
Solve for optimal tax rate by plugging in slots financed equation into tax revenue eq gives you:
max over τ ∈ [0,1):
g(τ) = G · (1 − τ)^a · [ N₀ + (τπ_I)/I^PhD ]^(1−b)
Can then get that FOC’s give you the optimal tax rate:
τ* = [ (1−b)π_I − a I^PhD N₀ ] / [ (a + 1 − b) π_I ]
Which can be used to solve for the optimal fully funded slots:
N* = N(τ*) = N₀ + (τ*π_I)/I^PhD

[csardinas]

Research Question

Do countries with higher supplies of educated labor experience more creative destruction, or do institutional frictions prevent talent from translating into firm entry? More specifically, how does firm entry intensity vary across countries after normalizing for both labor force size and tertiary educational attainment?

Background and Motivation

In Tapping into Talent: Coupling Education and Innovation Policies for Economic Growth, Akcigit, Pearce, and Prato (2020) argue that talent is a scarce input into innovation and growth, and that frictions in education, finance, and occupational choice generate misallocation of talent with first-order consequences for productivity. Their core insight is that innovation policy alone is insufficient if high-ability individuals are prevented—by financial constraints or institutional barriers—from acquiring the human capital needed to enter innovative occupations. As a result, societies may fail to “tap into” their available talent even when educational attainment rises.

While their analysis relies on rich Danish microdata, an open question is whether similar patterns appear in cross-country data at the macro level. In particular, if talent supply increases (e.g., via higher tertiary attainment), does creative destruction—measured through firm entry—scale proportionally across countries? Or do institutional differences lead some countries to convert talent into entrepreneurial activity more effectively than others?

Research Design and Methodology

To examine this question, I construct a cross-country panel using OECD data on (i) business dynamism and (ii) educational attainment. Business entry rates are drawn from the OECD Timely Indicators of Entrepreneurship (TIE), focusing on firm entry counts by country and year. Talent supply is proxied by the share of the working-age population (ages 25–64) with tertiary education, using OECD education statistics.

The empirical strategy proceeds in three steps. First, I compute creative destruction intensity, defined as firm entries per worker, to normalize entrepreneurial activity by country size. Second, I estimate the cross-sectional relationship between tertiary attainment and entry intensity in the most recent year, capturing how business dynamism scales with talent supply. Third, I compute residual entry performance—the difference between actual entry intensity and that predicted by talent alone. These residuals identify countries that under- or over-perform in creative destruction conditional on education. This approach mirrors Akcigit et al.’s conceptual framework: education expands the potential talent pool, but realized innovation depends on how effectively institutions and incentives channel talent into high-impact activities. Countries with negative residuals exhibit patterns consistent with talent misallocation, while positive residuals suggest environments that better translate human capital into entrepreneurial dynamism. In this sense, the analysis provides a macro-level complement to the paper’s micro-level evidence on talent, education, and innovation.

Visualizations

[clsun2004]

Chloe Sun

Research Question:

Does early adoption of generative AI tools by firms shift their subsequent innovation output away from disruptive breakthroughs and toward incremental development?

Motivation:

This memo engages Wu, Wang, and Evans (2019), Large Teams Develop and Small Teams Disrupt Science and Technology, which shows that large teams disproportionately engage in developmental, incremental innovation, while small teams are more likely to generate disruptive, paradigm-shifting ideas. Their core insight is that team structure shapes not only the quantity but the direction of scientific and technological progress. I propose to apply this framework to the diffusion of generative AI tools such as GitHub Copilot, ChatGPT, and other LLMs inside firms and research teams. Generative AI plausibly lowers the marginal cost of incremental development, code completion, and documentation, which may tilt innovation toward “developmental” improvements rather than disruptive breakthroughs. If AI tools are adopted first and most intensively by large teams focused on scaling products and optimizing workflows, then AI diffusion may systematically bias the innovation process toward exploitation rather than exploration. This creates a theoretically grounded and testable extension of Wu et al. (2019)’s development–disruption distinction.

A closely related sub-question is whether any such shift is stronger in firms with larger inventor teams, consistent with Wu et al.’s core mechanism. Conceptually, disruption is measured using citation-based novelty metrics similar to those in Wu et al. (2019), where a patent is more disruptive if it eclipses the future citations of the prior art it builds on. The first hypothesis is that firms that adopt generative AI earlier will, on average, produce patents that are less disruptive in the post-adoption period. The second hypothesis is that this reduction in disruptive innovation following AI adoption will be stronger for firms with larger inventor teams. The intuition is that generative AI amplifies the productivity of routine idea recombination and scaling, which disproportionately benefits large teams engaged in incremental development.

Research Design:

The empirical design measures firm-level AI adoption timing using either text analysis of SEC 10-K filings or AI-related job postings. For 10-Ks, adoption is defined as the first year in which a firm’s filing includes terms such as “artificial intelligence,” “machine learning,” “generative AI,” “large language model,” or related phrases, using standard parsed corpora from EDGAR or WRDS. Alternatively, adoption can be defined as the first year in which a firm posts job advertisements requiring machine learning, NLP, or generative AI skills, using data from Burning Glass or Lightcast. Innovation outcomes are measured using USPTO patent data from PatentsView or the NBER Patent Data Project, with firm–patent links based on harmonized assignee identifiers. Patent-level disruption scores are constructed following Wu et al. (2019), based on whether future patents citing a focal patent do not cite the same prior art. Team size is measured as the number of inventors listed on each patent. Standard controls include firm R&D spending from Compustat, firm size, industry fixed effects, and year fixed effects.

The empirical strategy is an event-study difference-in-differences design centered on each firm’s first AI adoption year. Let Post_it indicate years after firm i’s adoption. The main regression relates a patent’s disruption score to Post_it, team size, and their interaction, with firm and year fixed effects and standard controls. The key coefficients of interest test whether disruption falls after AI adoption and whether this decline is larger for patents produced by larger teams. This design allows a transparent visualization of dynamic effects and pre-trend checks.

Disruption_it ​= β_0 ​+ β_1​Post_it ​+ β_2​LargeTeam_i ​+ β_3​(Post_it​ × LargeTeam_i​) + α_i ​+ δ_t ​+ ε_it

The core visualization is an event-study plot of patent disruption around AI adoption, which can be implemented directly in R. Time is normalized to event time k = t − Adoption_i, and mean disruption scores are computed for bins from five years before to five years after adoption. Two lines are plotted: one for patents produced by large teams (above-median team size) and one for small teams (below-median team size). The x-axis shows years relative to adoption, and the y-axis shows the mean disruption score. If the hypotheses are correct, disruption should fall after adoption, with a visibly larger decline for large-team patents.

[lianduann]

Lian Duan

Research Question
Does the adoption of Generative AI (GenAI) systematically shift the nature of innovation from "Disruption" (creating new paradigms) to "Development" (refining existing ones)? Specifically, does GenAI induce "premature creative destruction" by accelerating the displacement of incumbents via low-cost recombination, before genuine high-productivity frontiers are discovered?

Motivation
Schumpeterian growth relies on the assumption that resources displaced by "destruction" are reallocated to higher-productivity "creation." However, recent literature suggests that the structure of innovation is as critical as its rate. Wu et al. (2019) distinguish between "Developmental" science (which builds on the status quo) and "Disruptive" science (which renders the status quo obsolete).

GenAI, as a prediction engine trained on historical data, fundamentally lowers the cost of combining existing knowledge. This creates a risk of a "Recombination Trap":

1. The "Large Team" Proxy: Wu et al. (2019) find that large teams excel at development but fail at disruption. By allowing individuals to process vast amounts of codified information, GenAI may inadvertently replicate the "large team" dynamic—favoring safe, consensus-based incrementalism over risky breakthroughs.
2. Artificial Hierarchies: Xu et al. (2022) demonstrate that "flat teams" (high Lead-ratio) drive novelty, while "tall teams" (low Lead-ratio) focus on development. If firms use GenAI to scale output without scaling leadership attention, they effectively lower their $L$-ratio, creating "artificial hierarchies" that suppress novelty.
3. The Remote Execution Bias: GenAI facilitates remote workflows. However, Lin et al. (2023) show that remote collaboration produces fewer breakthroughs and focuses on late-stage technical execution. GenAI may therefore accelerate a shift toward efficient execution of old ideas rather than the conception of new ones.

Proposed Research Design
The empirical strategy moves beyond simple productivity measures to quantify the quality of innovation.

1. Key Metric: I will construct the Disruption Index ($D$-score) for patents and papers (following Wu et al., 2019). The primary test is whether industries with higher GenAI exposure exhibit a systematic decline in average $D$-scores (indicating a shift toward "Development").
2. Mechanism Tests: I will interact GenAI exposure with organizational characteristics, specifically the $L$-ratio (team flatness) and Remote Intensity. This will test if the "innovation quality penalty" is concentrated in firms that use AI to reinforce hierarchy and distance, rather than to empower flat, exploratory teams.

Visualization

Figure 1: The "Recombination Trap" — Hypothesized Shift in Innovation Quality ($D$-Scores)This density plot illustrates the hypothesized impact of Generative AI on the character of innovation, measured by the Disruption Index ($D$-score) ranging from -1 (Development) to 1 (Disruption).
The X-Axis (Innovation Character): Represents the nature of the output.Values to the left (negative) indicate "Development"—refining existing knowledge.Values to the right (positive) indicate "Disruption"—creating new paradigms that render old knowledge obsolete.

The Blue Curve (Baseline / Pre-GenAI): Represents the historical norm. It features a wide variance and a "fat right tail," indicating that while most science is incremental, the system reliably produces rare, radical breakthroughs ($D \approx 1$).

The Red Curve (GenAI Exposure): Represents the hypothesized shift under high GenAI usage. The curve becomes taller and narrower (lower variance) and shifts to the left.Taller Peak: Indicates a flood of low-cost, derivative content.Vanishing Tail: Indicates the loss of the "disruptive tail," as algorithmic prediction favors likely combinations over risky, high-reward anomalies.

Interpretation: The visual shift demonstrates the "Recombination Trap": GenAI increases the volume of average ideas (the peak rises) but effectively decapitates the potential for exceptional ideas (the tail disappears).