Ilan Strauss

Using BEA NIPA Table 1.11 data via FRED, Penn World Table 11.0, and BLS nonfarm business sector series, I examine why different labor share measures yield such different trends—and where the missing income actually went. Key finding: labor compensation fell from 58.4% to 51.9% of Gross Domestic Income (1970–2024), while corporate profits (post-tax) rose from 7.8% to 14.4%. But a critical and underappreciated shift is the rise of depreciation from 12.8% to 16.5% of GDI—income that belongs to no one. As the capital share grows, so does the share absorbed by maintaining that capital. The economy-wide shift toward short-lived assets (software, equipment) means more output goes to replacing worn-out capital rather than to workers or capitalists as distributable income. International benchmarking shows the U.S. has the second-lowest G7 labor share at 56.8% (2023). I use GDI rather than GDP to avoid mixing income-side numerators with expenditure-side denominators.

Does rising concentration in GitHub commits reflect "rich-get-richer" dynamics from AI coding tools, or compositional changes from platform growth (40M to 100M+ users, 2019–2024)? Using GH Archive data, I analyze 19.3 million commits across 625,590 developer-year observations (January 2019 – October 2025). Methodologically, I employ power-law exponent estimation via Clauset-Shalizi-Newman (2009), attachment kernel regression measuring sublinear growth coefficients, Negative Binomial dispersion parameter analysis detecting rate heterogeneity, and top-1% composition tracking. Key finding: the attachment kernel coefficient β ≈ 0.4 indicates mean reversion rather than compounding advantage, contradicting the AI amplification hypothesis for most developers. The October 2025 endpoint reflects GitHub's API changes removing commit details, creating a natural stopping point.

An interactive simulation modeling AI platform and web content creator dynamics using Lotka-Volterra population equations with bounded gain-loss forms, combined with mechanism design theory (IR/IC constraints). Adjustable parameters include compensation rate, extraction intensity, mutualism benefit, and traffic diversion. Four pre-built policy scenarios: Status Quo, Licensing Regime, Fair Deal, and Traffic Collapse. The model is calibrated to real-world data from AI licensing deals (e.g., OpenAI-News Corp) and journalism employment trends. Early-stage work exploring whether a "content collapse" equilibrium emerges when reduced creator incentives degrade content quality, which in turn degrades AI output quality—a potential downward spiral.

ChatGPT and other conversational AI systems often end responses with suggested follow-up questions. Do these prompts actually extend conversations, or are they ignored? Using the WildChat dataset of 1M+ real ChatGPT conversations, I apply survival analysis (Cox proportional hazards, Kaplan-Meier curves) to model conversation duration and identify predictors of continuation vs. abandonment. Key variables: presence of follow-up suggestions, topic complexity, user query length, response satisfaction proxies, and time-of-day effects. Implications for understanding AI engagement patterns, designing less manipulative interfaces, and evaluating whether "engagement maximization" features serve user interests or platform metrics.