Papers and projects

Better Keep the Twenty Dollars: Incentivizing Innovation in Open Source

with Annamaria Conti (IE University), Jorge Guzman (Columbia) and Maria Roche (Harvard Business School) · Revision submitted, Management Science · NBER Working Paper 31668

Open-source code is inside almost every phone and website, but much of it is maintained by volunteers, and there is no settled way to pay for it. We followed about 100,000 GitHub users around the launch of GitHub Sponsors, which lets anyone send money to a developer. Signing up to receive sponsorship raised output even before any money came in. Actually receiving money had a lasting negative effect on innovation, whatever the amount, and the drop came from the community and service work that open source relies on rather than from coding.

NBER working paper SSRN NBER Bulletin HBS Working Knowledge

Artificial intelligences and human scientists exhibit complementary strengths in theory building

with Ke Li (INSEAD) and coauthors · under review · arXiv 2609.32562

This study tests whether AI can do one of the most creative parts of science, coming up with theories. Twenty-five AI language models, 13 senior researchers and 60 PhD students each built theories about gender and racial inequality, used them to predict new findings, and revised them when new evidence came in. The models beat most individual people on most tasks. They also wrote more elaborate theories, which judges rated higher without knowing who wrote them, but the extra detail did not make their predictions more accurate. The human scientists got more out of simpler theories, and because their theories varied more, pooling many people's predictions helped more. People also changed their minds selectively, mostly where they had been wrong, while the models revised more readily.

arXiv

Local Government State Capacity: Evidence from Brazil

with Michael Best (Columbia), Renata Lemos (World Bank) and Daniela Scur (Cornell) · in progress

Economists and political scientists talk about state capacity, meaning how well a government can actually get things done, but it is hard to measure. Every Brazilian municipality publishes an official gazette recording its daily business. We are reading essentially all of them, 16.4 million paragraphs so far, sorted by a series of language models, to build the first daily measure of what local governments do, such as hiring staff or responding to emergencies. About three-quarters of that work turns out to be buying goods and services or making personnel decisions, and the mix is almost the same in poor and rich municipalities.

EconPol Forum

Municipal Responses to Natural Disasters: Evidence from Brazilian Municipalities

with Michael Best, Renata Lemos and Daniela Scur · in progress

Brazil floods often, and fairly predictably. Putting the gazette data together with flood records, we look at what a disaster does to local government: which work stops, what gets protected, and how long it takes for normal business to return.

Persistence is Selection: How Serving Lowers the Supply of Experts

sole author · in progress

Internet standards bodies, journals and university hiring committees all depend on experts who serve for little or no pay. People who have served come back more often than people who haven't, and it is tempting to read that as commitment. Three institutions let me test this, because they fill their committees by public lottery and publish enough to redo every draw. One is the committee that chooses the leaders of the Internet Engineering Task Force, the volunteer body that writes the internet's technical standards. The other two are Italian committees that qualify and hire professors. The pattern turns out to be selection. The people who put themselves forward were more likely to come back to begin with, and serving itself makes them less willing to serve again.