Hi, I'm Caleb! I'm a research scientist at Upstart working on causal machine learning methods for real-world applications. I recieved a PhD in Economics from New York University and I use this site to share my research and writing. I might also start writing occational Substack articles (are you also bothered by the woeful statistical misunderstandings by those who claim that fine-tuning is evidence for theism?). If I do, I will link to them here.
Standard models predict that expectations of artificial general intelligence (AGI) should elevate long-term interest rates. I show that this prediction need not hold. I develop a heterogeneous-agent asset pricing model in which AGI capable of automating most human labor, can lower interest rates even as it dramatically accelerates growth. This result suggests caution when interpreting long-term bond yields as a signal of market expectations of transformative AI.
The rapid progress of development in the field of artificial intelligence may profoundly reshape the global economy by both increasing productivity and automating away many jobs. This paper explores how households adjust their economic behavior today in anticipation of transformative AI (TAI). Building on previous research, I introduce a novel mechanism where the future reallocation of labor from humans to AI systems owned by wealthy households creates a zero-sum contest for control over AI resources—driving changes in current savings decisions and asset prices.
Prediction markets suffer from reduced liquidity and price accuracy for long-horizon events due to the opportunity cost of committed capital. Recently, major platforms have introduced interest-bearing positions to mitigate this "long-horizon problem." I evaluate this policy using agent-based simulations with large language model (LLM) traders in a 2 x 2 factorial design, varying time horizon (4 days vs. 2 years) and the presence of interest. While long horizons degrade accuracy, the observed pricing bias (0.72 percentage points) is significantly smaller than theoretical and prior empirical estimates. Paying interest eliminates approximately 83% of the horizon effect on accuracy and more than triples market participation (from 17% to 62% of wealth). These findings suggest the long-horizon problem may be overstated in existing literature and that interest-bearing positions are a highly effective intervention, primarily by incentivizing participation rather than correcting bias.
MONET is a novel neural network architecture using extreme expert specialization (~250k experts/layer) to improve interpretability without sacrificing performance. I explain how it works, present a new interpretation of the architecture, and propose efficiency improvements.