Week 9: The Ends of Innovation

[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.

“Metaknowledge.” 2011. Evans, James A., and Jacob G. Foster. Science 331(6018): 721-725.
“Science of science.” 2018. Santo Fortunato, Carl T. Bergstrom, Katy Börner, James A. Evans, Dirk Helbing, Staša Milojević, Alexander M. Petersen et al. 2018. Science 359(6379).
“Toward a more scientific science.” 2018. Pierre Azoulay, Joshua Graff-Zivin, Brian Uzzi, Dashun Wang, Heidi Williams, James A. Evans, Ginger Zhe Jin et al. Science 361(6408): 1194-1197.
“The diversity–innovation paradox in science.” 2020. Bas Hofstra, Vivek V. Kulkarni, Sebastian Munoz-Najar Galvez, Bryan He, Dan Jurafsky, and Daniel A. McFarland. Proceedings of the National Academy of Sciences 117(17): 9284-9291.
“Machine science.” 2010. James Evans and Andrey Rzhetsky. Science 329(5990): 399-400.
“Machine learning as a tool for hypothesis generationL.” 2023. Jens Ludwig, Sendhil Mullainathan. The Quarterly Journal of Economics 139 (2): 751-827.
“The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies” 2025. Kyle Swanson, Wesley Wu, Nash L. Bulaong, John E. Pak & James Zou.
“After Science.” 2025. James Evans and Eamon Duede. Science.
“AI Tools Expand Scientists’ Impact but Contracts Science’s Focus.”Links to an external site. 2025. Hao, Qianyue, Fengli Xu, Yong Li, James Evans. Nature

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.

[maxzhuyt]

Hi All! I made a research discovery tool that helps scientists identify unexplored directions in their adjacent community. The app is now online at https://ai-discovery-app-production.up.railway.app.

1. Paper Upload

Submit any paper (published 2000–2025)!

2. Complementary Intelligence

Semantic Neighborhood

The pipeline's first analytical step embeds the seed paper's title and abstract into a 768-dimensional vector space using SPECTER2 (allenai/specter2_base with proximity adapter), a transformer model trained on millions of scientific papers using citation-graph supervision — papers that cite each other are pushed close together. SPECTER2 encodes the semantic content of each paper's title + abstract into a continuous space, capturing similarities that discrete labels miss. The topic embeddings (next section) is useful for detecting gaps at the community level, but SPECTER2 provides the ground-truth neighborhood of what is actually "citationally" close. Proximity in this space reflects scientific relatedness.

This embedding is searched against an index of 44.5 million papers (Social Sciences and Computer Science, 2000–2025) built from the same OpenAlex corpus. From the top 100 nearest neighbors, the pipeline filters to papers published at least 6 months after the seed that have not cited it. These are valuable signal: work that is doing closely related science but exists outside the seed's citation network. The citing community may be unaware of these papers.

Topic-Based Analysis

From the citing papers' topic assignments (drawn from OpenAlex's 4,516-topic taxonomy), the pipeline computes a Human Discoverable Distribution (HDD) — a distribution capturing where the follow-up community gravitates.

To detect structural blind spots, it loads a prebuilt 128-dimensional topic co-occurrence embedding space containing ~291,000 topic combinations. This space was constructed from a January 2026 OpenAlex snapshot of 72 million Social Sciences and Computer Science papers (2000–2025): for each paper, all topics generate unigrams and all pairwise combinations generate bigrams; co-occurrence counts across the corpus are transformed to 128 dimensions.

In this shared vector space, the pipeline identifies cultural holes — topics and topic-pairs close to the seed paper but never explored by any citing work, representing blind spots between non-communicating communities. It also identifies frontier papers ranked by Field-Weighted Citation Impact (FWCI).

Author Analysis

The pipeline loads an author embedding index (covering ~2.5M prominent authors) projected into the topic space. From these it computes seed team cohesion, citing-community diversity, and identifies adjacent unciting experts — researchers (active since 2015) nearest to the seed's topic centroid who have never cited the paper.

The entire pipeline runs in 10–20 seconds and produces a structured briefing.

3. Research Chat

The briefing and data dictionary are then handed to an LLM (GLM-5) as context for an unlimited interactive conversation. The user can chat about novel research directions falling outside the community's natural discovery horizon.

4. Multi-Agent Deliberation for Chat

Every time the user asks a research question, the system activates a multi-agent team:

• Three specialists — a Methodologist (designs and techniques), a Theorist (cross-disciplinary frameworks), a Domain Expert (state of the art, trends, key people). Each specialist can call live search WoS, OpenAlex, and Google Scholar APIs to ground their analysis.
• A Critic evaluates all three proposals for feasibility, novelty, and whether they represent true structural blind spots or merely uninteresting directions that everyone doesn't bother to try.
• A Manager synthesizes the most promising insights into a single coherent response, deprioritizing what the Critic flagged and elevating ideas that appeared across multiple specialists.

All specialists have API access to OpenAlex, Google Scholar, and WoS search, enabling them to find and reference real papers and researchers during deliberation.

Try it out & Explore!

Try this app and explore your own research thoughts in detail. By interacting with the agent, you can test if the system's briefing and multi-agent deliberation can elicit truly valuable insights for your specific projects.

This is still a pilot version. If you run into technical issues or find that it isn’t working as expected for your use case, please give me your feedback so I can keep refining the pipeline!

[zararania]

Research Question
If AI systems in drug discovery are designed to be optimizers, in that they are trained to predict success within the distribution of known biology, does their adoption systematically narrow the therapeutic space that the pharmaceutical industry explores? Does this represent a structural acceleration of the industry's bias toward exploiting known targets over discovering new ones?

Background
The pharmaceutical industry has always faced a tension between exploitation and exploration: investing in well-validated biology is safer and faster, but the drugs that transform medicine tend to emerge from genuinely novel target classes. Historically, this tension was managed imperfectly by human researchers whose idiosyncratic curiosity and career incentives introduced a degree of productive randomness into the pipeline. Hofstra et al. (2020) show that exactly this kind of diversity, particularly from underrepresented scientists introducing novel conceptual combinations, is what drives disproportionate innovation in science.

The question is what happens to that productive randomness when discovery is increasingly delegated to AI. Hao et al. (2025) provide a troubling empirical precedent: across academic science, AI tool adoption expands individual researchers' output while contracting the collective diversity of questions the field pursues. If the same dynamic operates in pharma, AI reshapes what the pipeline is for. Evans and Duede (2025) frame this philosophically: AI systems that optimize within existing knowledge structures are engines of refinement, not revolution. Swanson et al.'s Virtual Lab (2025) is a genuinely impressive system that nonetheless operates entirely within a well-characterized biological problem, designing better solutions to a question humans already knew to ask. Ludwig and Mullainathan (2023) demonstrate that ML models excel at prediction within distribution but structurally cannot generate out-of-distribution hypotheses, because they have no representation of what lies outside their training data. In drug discovery, "out of distribution" is precisely where the most valuable targets live - for example, unexplored receptor families, understudied disease mechanisms, or novel modality classes.

Methodology/Visualization
I construct two parallel trend lines over 2015-2024 using figures from Citeline's Pharma R&D Annual Review (2023), the BIO Clinical Development Success Rates report (2022), and the Nature Biotechnology AI Drug Discovery Survey (2023): the share of AI-assisted programs concentrated in the top three therapeutic areas (oncology, CNS, metabolic disease), and the share of those programs pursuing first-in-class targets. If AI accelerates exploitation over exploration, these lines should form a pair of scissors - concentration rises as novelty falls.

[ShiyangLai]

Research Question.

Can the internal representations of large language models be systematically mined to generate novel, testable social-science hypotheses about cultural and semantic patterns that are not readily observable from model outputs or traditional black-box prediction systems?

Research Design.

Recent work on automated hypothesis generation demonstrates that machine learning models can identify patterns in large datasets and propose empirically testable hypotheses (e.g., Ludwig and Mullainathan 2024). However, these approaches treat models as black boxes and rely on predictive correlations in observed data. This project proposes a different strategy: instead of using model predictions to infer hypotheses, I analyze the internal representations of large language models themselves. Mechanistic interpretability research suggests that LLMs encode a large amount of latent knowledge in hidden activations and feature interactions that may not be directly expressed in generated text. These representations capture subtle semantic and cultural regularities learned from massive corpora, including patterns that the model may not reliably articulate.
The empirical workflow proceeds in three steps. First, I construct a representation dataset by passing a large corpus of social text (e.g., news articles, social media posts, or scientific abstracts) through a pretrained language model and extracting hidden-layer activations. I then train sparse feature dictionaries (e.g., sparse autoencoders) to decompose these activations into interpretable features that correspond to recurring semantic patterns. Second, I analyze the structure of these features—particularly clusters of co-activated features and interference relationships—to identify candidate latent dimensions of meaning, such as implicit social categories, narrative frames, or cultural schemas. For each discovered feature cluster, I retrieve the highest-activating text contexts and summarize the underlying concept. These summaries are translated into explicit hypotheses about how language and social meaning are organized. Third, I evaluate the hypotheses empirically using external data. For example, if an internal feature corresponds to a latent narrative frame, I test whether texts with higher activation of that feature predict downstream outcomes such as diffusion, persuasion, or audience engagement. The predictive performance and novelty of these hypotheses will be compared against hypotheses generated by standard text-based methods such as topic models or embedding similarity.

[erindner13]

Erik Rindner

Research Question

Can AI-driven delay prediction systems increase airport economic growth by enabling innovative demand management in airport retail and food services during flight disruptions?

Airports are often understood primarily as transportation infrastructure, but they also function as complex commercial environments where substantial economic activity occurs. Restaurants, retail stores, and service vendors inside airport terminals generate significant revenue from passenger spending. One of the most important drivers of this activity is passenger dwell time—the amount of time travelers spend waiting in terminals before their flights. When flights are delayed, passengers remain inside the airport longer than expected, often increasing demand for food, beverages, and retail goods.
However, delays also create coordination challenges for airport vendors. Flight disruptions can cause sudden surges in passenger demand that restaurants and retailers are not prepared to meet. Because delays are difficult to predict precisely, vendors often lack the information necessary to adjust staffing levels, inventory, or service capacity in advance. As a result, airports frequently fail to capture the full economic activity generated by passengers who remain in terminals during delays.
This setting creates an opportunity for innovation through artificial intelligence. AI-based delay prediction models can analyze large datasets—including weather conditions, aircraft rotations, and air traffic congestion—to forecast flight disruptions before they occur. If airports and terminal vendors had access to these predictions, they could adjust staffing, food preparation, and inventory levels in anticipation of increased passenger dwell time. In this sense, AI would function as an innovation in demand forecasting, allowing airports to better match supply with passenger demand during delay events.
This idea also connects to this week’s reading, which shows how AI systems can coordinate complex tasks across multiple agents to improve productivity and innovation. Just as the “Virtual Lab” framework uses specialized AI agents to coordinate different stages of scientific research, AI-based delay prediction could coordinate airport operations and commercial vendors, allowing them to respond more efficiently to disruptions. In both cases, AI functions as a coordinating technology that improves how complex systems allocate resources. The growth implications of this innovation could be meaningful. When vendors are able to meet increased demand during delays, total spending per traveler may rise, increasing concession revenues within airport terminals. Because airports typically receive a share of vendor sales, this increase in commercial activity could contribute directly to airport-level economic growth. More broadly, AI-driven delay prediction could transform flight disruptions from purely negative events into opportunities for improved economic productivity within airport service markets.

Methods

To study this question, I propose analyzing airport-level data on flight delays, passenger traffic, and concession revenue. First, I would examine whether longer average delays are associated with higher passenger spending within airport terminals. This analysis would estimate the relationship between passenger dwell time and revenue generated by food and retail vendors. Next, I would simulate the potential effects of implementing AI-based delay prediction systems that forecast disruptions several hours in advance. If vendors had access to reliable delay predictions, they could scale staffing and inventory to better meet passenger demand during delay events. Comparing observed vendor capacity to predicted passenger demand would allow estimation of how much additional revenue airports could capture through improved preparation. The key outcome of interest is not simply whether delays increase spending, but whether AI innovation can improve how airports respond to those disruptions. By allowing vendors to anticipate passenger demand more accurately, AI-driven forecasting systems could increase the economic productivity of airport commercial ecosystems and contribute to sustained growth.

[eknez]

Name: Ethan Knezevich

Research Question: To what extent does digital technological adoption mitigate the economic effects of geographic and demographic structural constraints (measured using population age, University of Notre Dame ND-GAIN Country Index, and the CEPII GeoDist) on GDP per capita growth and the Gini coefficient across countries?

Progress:

This project investigates the relationship between a country's geographic and environmental resilience — as measured by the ND-GAIN Country Index — and its economic performance, proxied by GDP per capita growth and the Gini coefficient. The broader research question explores whether structural constraints such as climate vulnerability, resource scarcity, and adaptive capacity limit a nation's economic trajectory, with potential extensions toward digital adoption and AI-driven productivity as mediating variables.

My recent work has centered on data acquisition and preliminary empirical analysis. Two primary datasets were identified and successfully merged. The first is the ND-GAIN Country Index from the University of Notre Dame, a composite index measuring a country's vulnerability to climate change and other global forces alongside its readiness to leverage private and public sector investment. The dataset was downloaded and parsed in R. The second dataset, World Bank GDP per capita growth data, was retrieved directly via the WDI package in R, covering the period 2015–2023. Country-level averages were computed across the full window to smooth year-to-year volatility and capture medium-run growth trends. The two datasets were then merged by country name, yielding a cross-sectional sample of countries with complete observations on both variables.

With the merged dataset in hand, an Ordinary Least Squares regression was estimated with average GDP per capita growth as the dependent variable and the ND-GAIN score as the sole explanatory variable. This serves as an initial baseline model to assess whether geographic and environmental resilience has a statistically detectable association with economic growth at the cross-country level. Alongside the regression, a scatterplot was produced using R, with the ND-GAIN Country Index on the x-axis and average GDP per capita growth on the y-axis.

The next priority is extending the regression model to incorporate a digital adoption or AI readiness index as an additional explanatory variable, which will directly align the empirical work with the core research question. Potential interaction effects between resilience and digital capacity on growth outcomes will also be explored.

Figure:

In addition to the aforementioned figure, this generated figure illustrates the relationship between inequality and internet access, a key variable in my broader analysis of mitigation of structural constraints, which itself is primarily concerned with GDP per capita growth and inequality. However, the relationship is not "clear cut," due to a rather imperfect trend.

Sources: University of Notre Dame ND-GAIN Country Index and World Bank Digital Adoption Index

[nbcodes57]

Nick Baumgartner
Research Question: Does AI-augmented research disproportionately bypass marginalized populations, amplifying existing innovation gaps beyond what purely human-driven science produces?

Motivation: Ludwig and Mullainathan (2024) show that ML can generate novel hypotheses, but only within the boundary of available data. Yin et al. (2025) find that AI-augmented science, while boosting individual productivity, contracts the collective topical coverage of research. Previous literature from this course on innovation bias establishes that inventors disproportionately create solutions for people like themselves. This memo aims to explore the intersection of these works: if human innovation is biased toward well-represented groups because innovators draw on lived experience, and AI innovation is biased toward data-rich groups because algorithms draw on training data, then AI reinforces rather than corrects the human bias. Because data availability correlates with existing representation, AI may amplify the gap. The scale of the underlying disparity is stark: as of 2023, roughly 86% of GWAS participants are of European ancestry despite representing approximately 16% of the global population.

Research Design: The proposed design is a cross-population OLS regression using data on roughly 20–30 demographic or geographic groups over 2015–2023. The dependent variable is the share of publications targeting each population group, normalized by that group's disease burden, and identified via method keywords in OpenAlex or PubMed. The key independent variable is how many entries exist for each population in major repositories like the GWAS Catalog. The core comparison is the slope of dataset availability on research output for AI-assisted papers versus traditional papers. A steeper slope for AI-assisted work would support the hypothesis that AI amplifies existing data representation gaps. The mechanism is straightforward: traditional research can draw on fieldwork or case studies from data-sparse populations, whereas ML methods require large structured datasets, mechanically tightening the link between data availability and output. Controls include research funding, disease burden, and English-language publication norms. The design is correlational, dataset availability and research interest are likely jointly shaped by funding and institutional capacity, though the controls partially address this. Robustness checks include alternative population groupings and varying the AI-keyword classification threshold.

[genevese]

Geneva Kirk Drayson

Does anonymized peer evaluation (double-blind review) reduce the “discounting” of novel contributions from underrepresented scientists, and does it increase downstream uptake (citations / idea diffusion) of those contributions?

Underrepresented scientists appear to generate scientifically novel ideas at elevated rates, yet those novel contributions diffuse less and translate into weaker career outcomes—i.e., novelty is discounted. (Hofstra et al., 2022)
This pattern is consequential because novelty is a key input into long-run scientific progress, and systematic under-uptake implies a measurable loss of potential discoveries and applied spillovers. The metaknowledge / science-of-science agenda argues that science is sufficiently instrumented (publications, citations, text, and networks) to detect such regularities and then redesign institutions to reduce waste and bias. (Evans et al., 2011)

A plausible mechanism for discounting is evaluative bias operating at the peer-review / selection stage. In randomized work embedded in real review processes, reviewers scored more novel proposals worse (even controlling for proposal quality), indicating that evaluation systems can directly penalize novelty . If novelty is already penalized on average, and if underrepresented scholars are disproportionately producing novel combinations, then evaluative processes may create an interaction effect: novelty penalties may be larger when novelty comes from underrepresented groups, producing exactly the paradoxical pattern observed in large-scale career data.

This project tests a concrete, policy-relevant intervention: anonymization in peer evaluation. If anonymized review attenuates novelty discounting and increases subsequent uptake, it supports double-blind review (or stronger anonymization tooling) as a scalable lever to improve allocative efficiency in science—shifting attention toward high-novelty work regardless of the author’s inferred demographics. If anonymization does not help, the evidence points to downstream mechanisms (network access, prestige, or citation practices) rather than gatekeeping, narrowing where to target reforms.

Datasets:
OpenReview platform data (ICLR and other venues using OpenReview): submissions, review texts, scores, decisions, author lists, timestamps (publicly accessible via OpenReview API).
OpenAlex (open bibliographic graph: works, citations, authors, institutions, concepts).

[HenryPinner7]

Research Question: Do inventors from companies acquired by large tech companies (Apple, Microsoft, Google, and Amazon) become more or less innovative after the acquisition (as measured by patent output and patent impact)?

Motivation: See previous memos.

Progress: Building directly on last week’s acquisition–inventor dataset and our pre-acquisition baseline, this week we shifted the analysis to the around-acquisition window and generated our first post-acquisition results for inventor patent output. Using each deal’s acquisition date as an “event,” we assigned every inventor’s patent filings into 1-year event-time bins from T = −5 to −1 and T = +1 to +5, omitting T = 0 to avoid mixing pre- and post-timing in the transition year. We then aggregated to the bin level by calculating the average number of patents per inventor per event-time bin, and plotted these averages with 95% confidence intervals to visualize how output evolves leading into and following acquisition; this event-time analysis draws on 16,328 patents, 2,725 inventors, and 140 acquired companies. To understand whether the overall pattern is driven by inventors who stop patenting entirely after acquisition, we re-ran the same bin-averaging and plotting exercise on a restricted sample of inventors with Post > 0 (i.e., at least one post-acquisition filing), which we will show in our final report.

Finally, we also looked at patenting output pre vs. post acquisition in relation to the size of the acquisition deal itself by introducing a heterogeneity cut: we sorted acquisitions into deal-value quartiles for the 34 companies where we could reliably recover deal values (equal number of deals per bin) and estimated the Post − Pre change in inventor patent output within each quartile. This deal-value analysis is based on 7,866 patents across 1,320 inventors, and produces the point estimates and confidence intervals shown in the quartile figure. Together, these steps let us pair a clean event-time visualization with an initial test of whether “what happens after acquisition” differs systematically by deal size.

Figures:

Analysis: Looking at the figures we collected from this analysis, average patenting rises steadily in the five years leading up to acquisition, peaks at T = +1, and then declines sharply from T = +2 through T = +5. The pre-acquisition run-up is consistent with acquisitions selecting “high-momentum” targets, while the +1 spike likely reflects short-run transition dynamics (e.g., pipeline completion or filing timing). The sustained post-+1 decline, which is measured on the full sample including inventors with zero post-acquisition patents, suggests acquisition is associated with a meaningful reduction in observable inventor patenting in the medium run.

Overall, the quartile split doesn’t give us a clean, decisive story yet, as precision varies a lot across bins. The lowest-value deals (Q1) show a positive estimated change of +1.52 patents per inventor (although the value is insignificant at p = 0.377), which is imprecise, but suggests smaller acquisitions may be the least harmed and could even benefit in output post-acquisition. The other bins show that Q2 increases output by +1.84 (p = 0.008, statistically significant), Q3 decreases by −1.09 (p = 0.051, borderline significant), and Q4 increases slightly by +0.31 (p = 0.436, not significant). Overall, the quartile split suggests mixed evidence of heterogeneity in how acquisitions affect inventor patent output, where any potential “boost” looks most plausible for smaller deals (even if noisy), while larger deals do not show a consistent positive or negative effect on average.

[nicolaspark1]

Nicolas Park
Research Question:
Does the co-occurrence of concept pairs along an indirect pathway, specifically GLP-1 drugs with appetite mediators (A→B) and appetite mediators with snack categories (B→C), predict the emergence of direct co-occurrence between GLP-1 drugs and snack categories (A→C) in consumer discourse?

Motivation and Link to the Readings:
This memo adapts the "ABC discovery" framework from computational scientific discovery to the study of market structure. In that framework, new hypotheses emerge when two separate bodies of knowledge share an intermediate concept: if A appears with B in one context and B appears with C in another, a previously unrecognized A–C relationship may exist. I apply this logic to consumer discourse about products.

Here, “A” represents GLP-1 weight-loss drugs (e.g., Ozempic, Wegovy), “B” represents behavioral mediators such as appetite suppression or reduced cravings, and “C” represents snack or food categories such as chips, soda, and desserts. If consumers discuss GLP-1 drugs together with appetite suppression (A→B), and appetite suppression together with reduced snack consumption (B→C), this structure implies a latent competitive relationship between pharmaceuticals and snack products (A→C) before consumers explicitly articulate it. The empirical question is whether the indirect A→B→C pathway anticipates the emergence of direct A→C language, and if so, how far in advance.

Data and Research Design:
I collect public English-language posts from X (Twitter) using Grok's search interface, spanning January 2020 through March 2026. Posts are retrieved through keyword co-occurrence queries capturing three edge types: A–B posts linking drug terms with appetite mediators (N=105), B–C posts linking mediators with snack categories (N=67), and A–C posts linking drugs directly with snack terms (N=97). After removing duplicates, advertisements, and finance-only posts, the final dataset contains 348 unique posts.

Because early GLP-1 discourse is sparse and temporal coverage is uneven, posts are aggregated into semi-annual periods, yielding 11 periods with observed activity. For each period, I compute the A–B rate and B–C rate as proportions of sampled posts, then define the ABC implied substitution score as their product: ABC_implied = (A–B rate) × (B–C rate). This captures the strength of the indirect drug→snack relationship mediated through appetite language. I compare this to the direct A–C rate. Both series are normalized to their respective peaks and fitted with degree-3 polynomial trends to reveal underlying trajectories.

Visual Evidence:

Figure 1 plots the fitted trends alongside observed data points. The ABC implied signal rises earlier, reaching elevated levels by 2022H1 and peaking around 2023H1. The direct substitution signal remains low until mid-2023 before rising steadily. The trend lines cross at approximately mid-2023, after which direct language dominates. The estimated lead time between peaks is approximately 1.8 years.

Contribution:
These results provide evidence that the indirect A→B→C pathway predicts the emergence of direct A→C substitution language by nearly two years. This suggests consumer discourse contains latent signals of cross-industry competitive pressure before those relationships become explicit, which is consistent with the broader argument that AI-assisted analysis of public language can function as an early warning system for innovation-driven market reallocation.

[TarunSethi13]

Research Question

Did North Carolina’s 2021 corporate tax reform increase local innovation activity relative to neighboring states?

Research Design

To examine the potential impact of North Carolina’s corporate tax reform on innovation activity, I construct a border-county comparison using employment in research and development services (NAICS 5417) from the Bureau of Labor Statistics’ Quarterly Census of Employment and Wages (QCEW). Following recent work emphasizing the geographic concentration of innovation, I focus on the Charlotte metropolitan corridor, comparing counties on the North Carolina side of the border (Mecklenburg, Gaston, Union, and Cabarrus) to nearby South Carolina counties (York and Lancaster). These counties share a common regional labor market and economic environment but operate under different state tax policies. This cross-border design provides a natural framework for a difference-in-differences approach.

The primary outcome variable is annual average employment in the scientific research and development sector (NAICS 5417), which serves as a proxy for local innovation activity. Figure 1 presents indexed employment trends normalized to 2019, the final pre-COVID year prior to North Carolina’s 2021 tax reform. Indexing to 2019 allows for clearer interpretation of relative growth while avoiding distortions from small baseline employment levels earlier in the sample.

Figure 1:

Several patterns emerge from the figure. First, employment trends in the North Carolina and South Carolina counties move relatively similarly prior to the policy period, suggesting that these regions are subject to similar regional economic conditions. Second, the data exhibit substantial volatility around 2020 and 2021, likely reflecting disruptions associated with the COVID-19 pandemic. These pandemic-related shocks complicate simple comparisons of employment levels across states during this period.

Despite this volatility, the post-2021 period shows evidence of stronger recovery in the North Carolina counties relative to the South Carolina counties within the Charlotte corridor. Following the sharp decline observed in 2021, R&D employment on the North Carolina side of the border increases more rapidly through 2022 and 2023. While this descriptive pattern is not sufficient to establish causality, it suggests that policy differences across the border may contribute to divergent post-pandemic recovery trajectories.

Future analysis will formally test this relationship using a difference-in-differences regression framework that controls for county fixed effects and year-specific shocks, including those associated with the COVID-19 pandemic. Such a specification will allow for more rigorous identification of whether North Carolina’s corporate tax policy changes are causally associated with increased innovation activity relative to neighboring states.

[hcarpenter4]

Hudson Carpenter

Research Question:

Do inventors from companies acquired by large tech companies (Apple, Microsoft, Google, and Amazon) become more or less innovative after the acquisition (as measured by patent output and patent impact)?

(See previous memos for motivation)

Progress:

This memo focuses on our progress made in relation to changes in patent impact by inventors before and after their company is acquired by a large incumbent technology company. That is, we tracked changes in average citations per patent within the 5 years before and after acquisition. In addition to these overall results, we performed two heterogeneity tests to better understand the particular mechanisms that create our results. In this memo, I will discuss our industry-level analysis, in which we conducted a difference-in-difference regression to isolate the effects of same-industry acquisitions compared to cross-industry acquisitions.

Data:

We collected acquisition data from LSEG Workspace on all completed mergers, and majority acquisitions by Google, Apple, Microsoft, and Amazon. We used Google BigQuery to isolate all the inventors producing patents for these acquired companies, and filtered for distinct company-name pairs to track post-acquisition patenting activity either at the acquired company, or the incumbent firm. Our final dataset consisted of:

• 140 acquired companies
• 2,725 distinct inventors
• 16,328 combined patents

Analysis 1

These results show the average forward citations per patent issued by inventors in the years surrounding an acquisition. Across all acquiring firms, patent impact declines following acquisition. On average, inventors produce their most influential patents prior to being acquired. Notably, there are spikes in citations per patent approximately five years and two years before acquisition. This pattern suggests that inventors often experience multiple bursts of high-impact innovation before their firms are acquired, which may reflect a degree of selection bias in acquisitions, with incumbent firms targeting companies during periods of strong innovative output. Overall, the results indicate a clear decline in patent impact both immediately before and in the years following acquisition.

Analysis 2

We ran this DiD regression to estimate the difference in patent impact trends between same-industry acquisitions and different-industry acquisitions. The regression results show a substantial decline in patent impact following acquisition. For acquisitions by firms in different industries, the estimated coefficient on the post-acquisition indicator implies that forward citations per patent fall by roughly 18.8 citations after the acquisition. Firms acquired by companies in the same industry have slightly higher average citations per patent prior to acquisition, as indicated by the positive coefficient on the same-industry indicator. The interaction between the post-acquisition period and same-industry acquisitions suggests that the decline in patent impact after acquisition is larger for same-industry deals than for cross-industry acquisitions. That said, the coefficients on the same-industry indicator and the interaction term are not statistically significant, indicating that there is no strong evidence in our data that patent impact differs meaningfully between same-industry and cross-industry acquisitions either before or after the acquisition.

[clsun2004]

Chloe Sun

Research Question

Did faster adoption of India's Unified Payments Interface (UPI) causally increase firm entry and local economic activity across districts?

The central claim is that districts with more rapid post-2016 growth in UPI transactions saw larger increases in new firm registrations relative to slower-adopting districts. Engaging this week's readings, we push beyond a standard transaction-cost framing: drawing on Evans & Foster (2011), UPI functions as epistemic infrastructure — not merely cheaper exchange, but a system that restructures who becomes legible in economic knowledge space. A vendor entering the UPI graph gains a transaction history, a discoverable network, credit signals. The empirical question is whether this legibility translates into durable entrepreneurship, and whether its benefits are distributed uniformly.

Motivation

UPI, developed by the National Payments Corporation of India and launched in 2016, enables instant, low-cost digital payments directly between bank accounts. Its rapid expansion dramatically reduced transaction costs for consumers and firms. Lower transaction costs can reduce entry barriers, increase experimentation, and accelerate reallocation toward more productive firms. If digital payments make it easier for new firms to transact, reach customers, or operate formally, we should observe increased firm formation in regions where adoption is stronger. The project also speaks to when digital public goods actually translate into economic dynamism — digital infrastructure often requires complementary investments like internet connectivity, banking access, and digital literacy to generate growth effects, so examining heterogeneous effects across districts sheds light on where these conditions are met.

Research Design

We construct a district-year panel covering roughly 740 districts from 2014–2024, spanning pre- and post-UPI periods. The key explanatory variable is district-level UPI adoption intensity measured as transactions per capita, supplemented by merchant-facing proxies like QR-code penetration and merchant enrollment. Primary outcomes are new firm registration rates from the Ministry of Corporate Affairs — including the share of new firms in sectors more likely to benefit from digital payments such as retail trade and platform-based services — and within-district nighttime radiance growth (VIIRS) as a broader proxy for local economic activity.
Identification relies on staggered bank integration into UPI, coordinated by NPCI based on technical readiness rather than district-level entrepreneurial demand, providing plausibly exogenous variation in adoption timing. An event-study framework with district and year fixed effects tests parallel pre-trends and estimates post-integration divergence between faster and slower adopting districts. Controls include pre-existing banking penetration, internet access, smartphone usage, and pre-2016 economic trends. Heterogeneity analysis across district types sheds light on when and where digital payments infrastructure translates into measurable entrepreneurial outcomes.

Visualization

The two charts below illustrate the hypothesized mechanism. The scatter plots district-level UPI adoption intensity against firm entry rates, with dot size reflecting total new registrations and color indicating urban tier. The event-study chart shows hypothesized pre-trend parallelism supporting causal identification, and post-integration divergence between high- and low-adoption districts isolating the UPI effect on firm entry.

[ukaishan]

Urjita Kaishan
Does the gender gap in life expectancy across Illinois counties predict differences in innovation capacity, and can AI-enabled healthcare technologies reduce this gap over time?

Do counties with larger gender gaps in life expectancy exhibit systematically lower long-run growth potential, and can the diffusion of AI-enabled scientific innovation mitigate these disparities? Specifically, I ask whether variation in the female-male life expectancy gap across Illinois counties reflects differences in local capacity to participate in and benefit from AI-driven scientific progress. The goal is to determine whether AI’s transformation of science (particularly in healthcare) acts as a convergence force that compresses regional health inequality, or instead reinforces existing disparities in human capital and growth.

This project draws inspiration from Evans and Duede’s After science (Science, 2025), which argues that AI is reshaping the structure of scientific discovery itself. The authors suggest we are entering a new epistemic regime in which AI systems help generate hypotheses, prioritize research directions, and influence which problems receive attention. In endogenous growth terms, this implies a structural shift in the knowledge production function. If AI alters how and where scientific knowledge is generated and deployed, then local differences in health outcomes may reflect unequal integration into this evolving innovation system. I extend this logic by arguing that gender life expectancy gaps capture underlying disparities in effective labor supply, human capital investment incentives, and institutional access to AI-enabled healthcare innovation. Counties with persistently lower male life expectancy may experience reduced expected returns to skill accumulation and weaker innovation ecosystems, reinforcing divergence.

To study this question, I propose a county-level panel analysis across Illinois. The primary outcome variables will be (1) male life expectancy, (2) female life expectancy, and (3) the gender life expectancy gap by county and year. Key explanatory variables will include measures of local economic dynamism (income growth, employment growth), healthcare access, education levels, and an index of AI exposure based on occupational composition or proximity to AI-intensive industries. The baseline regression will estimate the relationship between AI exposure and changes in the gender life expectancy gap, controlling for unemployment, median income, demographic structure, insurance coverage, and county fixed effects. To mirror the theoretical insight from After science, I will test whether counties more exposed to AI-intensive sectors exhibit faster convergence in male life expectancy, consistent with AI-driven scientific advances diffusing into local healthcare systems. I will also examine heterogeneous effects across counties with different baseline health infrastructure to determine whether AI amplifies or compresses inequality.

As shown in this figure, life expectancy in Illinois counties increased from 1990 to 2010, but substantial dispersion persists across counties and between men and women. This cross-county heterogeneity provides a natural setting to examine whether AI-driven changes in the structure of scientific innovation translate into more equitable health improvements, or whether the benefits of “after science” accrue primarily to already advantaged regions.

[bheim]

How disruptive have publications been in CS, Psych, and Math across the past 5 years?

In building towards our final project, I wanted to collect data on disruptiveness across a few fields as a proof of concept. To that end, I collected data from the prior 5 years from CS, Psychology, and Mathematics to better understand how disruptive their publications are. As a reminder, disruptiveness is computed as follows

$$ \frac{n_i - n_j}{n_i+n_j+n_k} $$

where ni is the number of citations for the focal publication that do not cite the references the focal publication rates, nj is the number of citations that cite both the focal publication and the references the focal publication cites. Lastly, nk is the number of citations that do not cite the focal paper but do cite the references the focal paper cites.

Since we’re looking at such recent papers, we found that throwing out nk so that our analysis was simply $\frac{n_i - n_j}{n_i+n_j}$ was necessary as nk would mog (as the kids are saying these days) the other numbers.

Below are the results from my analysis. You’ll notice the standard deviations are huge, which is driven by the small sample I chose (only 50 papers per year). In our final project, we’ll scale this up to 1000 papers per time period (or more if we are feeling ambitious!).

One important feature we wanted to validate is that as papers become more recent, their distance from 0 rapidly deteriorates. This is expected since these papers simply haven’t had time to get cited much. We’ll have to incorporate this fact into our model. An interesting fact that appears from a quick glance is that mathematics papers are remarkably developmental as compared to CS and psychology. For a more developed field, this may make sense. Computer Psychology seems to be the most disruptive field. I’m excited to expand this analysis to more fields and incorporate an AI exposure term in our regression.

It’s been a pleasure writing these memos and I hope others have found them to be interesting.

[vkurakata]

Can AI agents autonomously generate and empirically validate novel hypotheses about hiring decisions from high-dimensional resume data?

In Machine Learning as a Tool for Hypothesis Generation, Ludwig and Mullainathan show that machine learning models trained on high-dimensional data can detect behavioral patterns that human researchers might not naturally consider. Their framework uses algorithms to uncover predictive structure in complex data, after which humans interpret these patterns and translate them into novel, testable hypotheses. This proposal extends that idea by replacing the human “interpretation bottleneck” with an AI agent that proposes and tests hypotheses in a labor market setting.

Specifically, the project asks whether an AI agent can perform the latter stages of this pipeline itself, transforming black-box predictive patterns in resume data into interpretable hypotheses about hiring decisions. The empirical setting is hiring or callback outcomes using high-dimensional resumes (including text and formatting) combined with structured applicant characteristics. Existing labor economics research typically focuses on hypotheses that are straightforward to formulate ex ante, such as race- or gender-related signals. However, resumes contain many high-dimensional cues, such as tone, narrative structure, action verbs, perceived “polish,” formatting choices, and other stylistic features, that may systematically influence employer decisions in ways that have not yet been formalized.

Methodology:

The empirical strategy proceeds in several stages. First, a predictive model of callback outcomes is trained using structured applicant variables and high-dimensional resume features, and its predictive performance is evaluated out of sample.

Second, candidate influential patterns are extracted using counterfactual edits: minimal modifications to a resume that shift the model’s predicted callback probability. These counterfactual pairs reveal which features of resumes appear to influence hiring predictions.

Third, an AI agent analyzes these counterfactual resume pairs and generates interpretable hypotheses in the form of measurable features, such as a leadership-language index, a concision score, a credential-signaling measure, or a layout readability proxy.

Fourth, each hypothesis is evaluated using standard econometric specifications with appropriate controls and job-posting fixed effects. Hypotheses are validated through out-of-sample predictive contribution and robustness checks, while accounting for multiple hypothesis testing.

The key outcome of interest is whether AI-generated hypotheses explain incremental variation in callback outcomes beyond a baseline set of “human” predictors and whether these relationships generalize across jobs and time. As a stronger validation, the project could incorporate an audit experiment in which the AI agent generates paired resumes that differ only in one discovered feature, allowing the causal effect of that feature on callback rates to be estimated. The broader contribution is a replicable discovery pipeline that identifies novel, interpretable determinants of hiring decisions from high-dimensional application data.

This figureillustrates simulated hiring outcomes as a function of an AI-discovered resume feature. Higher values of the feature increase callback probability.