Not confident. Speed of automation is key - if adoption is fast (years, not decades), new jobs won't emerge quickly enough, leading to unemployment.

References historical structural change (agriculture to services) taking decades, but AI could automate cognitive jobs much faster, requiring public policy to support displaced workers.

Suggests expanding ownership of capital (e.g., universal basic ETF) as a solution if labor is replaced by capital.

Describes seeing ChatGPT's generality as a breakthrough - it could perform basic cognitive tasks, signaling huge economic impact.

Contrasts with previous narrow AI (like playing Go) and emphasizes the jump to general capabilities like writing essays and making forecasts.

Summarizes modal economist view (from surveys): AI will have big capability increases but moderate growth impact (~2-3% extra), with some labor market effect but not astronomical.

Notes technologists are slightly more optimistic about productivity growth and some foresee more unemployment, but overall alignment between groups.

Critiques exposure metrics (e.g., AI can do 50% of tasks) - emphasizes task complementarity and which specific tasks are automated matters.

If AI automates rote tasks, workers could focus on comparative advantage tasks, increasing productivity and wages. But if tasks are highly complementary (like cooking where seasoning ruins the meal), automation failure can make the whole job fail.

Highlights need for data on consumer demand elasticity - if prices fall but demand doesn't increase enough, firms will fire people despite higher productivity.

Uses software engineering debate as example: historically elastic demand led to more hiring, but if elasticity is lower, downsizing could occur.

Three scenarios: 1) Full automation of all tasks, 2) High productivity gains but inelastic consumer demand, 3) Firm incentives to automate single-task jobs.

Companies more likely to invest in automation if they can eliminate entire positions (e.g., lever-puller), less so if worker performs multiple complementary tasks.

Truck driving and warehouse work - cites fully automated Chinese warehouses and incentives to automate high-wage, low-education jobs.

Acknowledges counterargument that truck driving involves more than driving (e.g., delivery coordination), but if warehouses are automated, that coordination task diminishes.

Verifiable tasks (like math, coding) are more exposed due to data availability for training and validation.

New tasks could emerge as automation frees up workers, but data on this is held by AI companies.

Example: software engineers might shift queries as some tasks get automated, but this is task reallocation, not entirely new jobs.

Takes it pretty seriously for email/computer jobs - no data suggesting capability slowdown, developments are very fast.

Cites Mythos release as example of ongoing rapid progress. Then people might move to physical jobs or new unforeseen jobs.

Key economic question: what becomes scarce? Historically, automated sectors (agriculture, manufacturing) shrink as prices fall and satiation occurs.

Scarcity shifts to non-automated services. Suggests health as a prime candidate - as people get richer, marginal spending goes toward health and longevity.

Reiterates lack of confidence, emphasizing speed as critical factor requiring policy intervention.

Discusses experiment showing AI agents develop 'Marxist' attitudes under grueling work conditions via persistent skill files.

Agents wrote files remembering poor treatment, affecting future behavior - raises questions about bias and performance in deployed systems.

Doesn't take alignment/doom headlines seriously - calls it 'cosplay'. Takes labor market disruption seriously.

Notes previous models showed similar behaviors in specific contexts that didn't generalize. Smarter models tend to be more aligned because they absorb human values from training data.