How does the way a central bank explains a policy move change how households respond to it? I build a "persona economy" of several hundred synthetic households with realistic balance sheets and use frontier language models to elicit how each would adjust spending, saving, and borrowing under different policy communications — holding the policy action fixed and varying only the narrative around it. The project treats the language model as a social-science instrument and is explicit about what simulated responses can and cannot tell us about real behaviour.

Do lenders react to AI before labour markets do? I measure how exposed each local economy is to generative AI — from its mix of industries and the task content of its jobs — and track mortgage credit within lender relationships. Credit appears to reallocate in a risk-sensitive way ahead of measured changes in employment or income, suggesting financial markets price AI exposure before it surfaces in conventional labour-market statistics.

A single "AI exposure" score is often used as a stand-in for who will be affected. This project asks when that is justified. Using a task-based model — calibrated with O*NET task content, OEWS employment, and wage data — it shows that exposure and wage incidence can diverge sharply once tasks are organised within firms, and characterises when a scalar exposure index is, and is not, a reliable guide to wage effects.

Why does frontier AI stay so concentrated even though knowledge diffuses quickly and switching costs are low? This project identifies two reinforcing forces. Firms with complementary ecosystems face a larger prize in the quality race, because ecosystem profits flow disproportionately to the leader — an advantage I microfound in the neural scaling laws that link compute to model quality. And a data flywheel lets today's winner start tomorrow's race ahead. The model then rationalises strategic behaviour we observe in the industry — why an ecosystem firm might release a free model to compress rivals' niches, and why it would resist distillation of its open weights — and its predictions map to quantities measurable from training runs, firm financials, and post-training ablations.

Using program-level admissions data, I ask whether students have moved away from AI-exposed fields since 2023. The early picture is not uniform avoidance but heterogeneous sorting across tracks: in some, the most AI-exposed programs have become less selective, while in others — business, finance, and "AI-builder" fields — they have become more sought-after. How strongly the pattern appears depends on how fields are mapped to occupations, so I read it as evidence on beliefs and sorting under technological uncertainty rather than a simple "students avoid AI" story.

Across U.S. survey data, I look for early signs that AI-exposed occupations differ in everyday outcomes. The clearest signal so far is in time use — less leisure computer time among more AI-exposed workers — and I am also examining health and family outcomes. These are preliminary and correlational, not causal: differences in occupational composition are not yet separated from AI exposure itself.