Research

I model two channels by which AI progress can feed back into the automation of AI research and lead to a self-sustaining acceleration. Capability improvements allow AIs to perform a larger share of AI research tasks, while efficiency improvements raise AI research output per unit of compute on tasks already automated. I separate these margins in a task-based R&D model of AI progress. A software intelligence explosion (SIE)—a self-sustaining acceleration in AI progress with other inputs held fixed—requires both margins to improve sufficiently fast. Efficiency must improve fast enough to overcome the rivalry of compute, while capability must improve fast enough to prevent human researchers from bottlenecking AI progress. Neither margin can substitute for a shortfall in the other. I argue that this distinction is essential for empirical studies of an SIE, because residual measures of algorithmic efficiency include capability improvements and thus are biased toward overestimating the likelihood of an SIE. I conclude with some extensions that this model is uniquely suitable for investigating.

How do firms navigate the tension between securing proprietary knowledge and remaining open to knowledge flows? We study this tension in the context of economic espionage, compiling a comprehensive dataset of incidents where foreign competitors steal technology from US firms. In an event study design, revenues and R&D at targeted firms decline by 40% within five years, with mechanism analysis suggesting that this loss is driven by obsolescence of the firm's valuable technology. These effects do not appear for firms unsuccessfully targeted for espionage, supporting a causal interpretation. However, we find no evidence that targeted firms become more restrictive on knowledge flows after espionage; firms do not reduce hiring of foreign scientists, patenting with foreign inventors, or international business functions. Overall, espionage has clear economic harms to targeted firms and US industry, but firms appear to value the benefits of inward knowledge flows more than these costs.

Statistical methods typically lack power to detect p-hacking and publication bias among individual researchers, making it difficult to detect how prevalent p-hacking is. We propose a novel empirical Bayes method to estimate researcher-level p-hacking prevalence. Using data from top medical journals, we find strong evidence these practices are widespread: at least 85% of researchers over-reject null hypotheses, with a conservative lower bound of 73%. Our approach identifies 20% of researchers as over-rejecting, whereas conventional tests detect none. These findings demonstrate that p-hacking and publication bias are systemic rather than isolated misconduct, underscoring the need for structural reforms in scientific practice.