Inside AI with Anthropic's Peter McCrory

Episode Summary

Executive Summary: The episode opens with a fast-moving macro update on a contentious Fed meeting, sticky inflation, and weaker-than-expected growth data, then pivots to a wide-ranging interview with Anthropic’s head of economics, Peter McCrory, on how Claude is used in economic research and enterprise workflows. The discussion argues AI is already boosting productivity and widening task scope, but its biggest effects may depend on business adoption, complementary investments, and labor-market churn rather than immediate job destruction.

Main Topics: Fed hold, dissent, and policy direction (Priority: 5/5): The Fed left rates unchanged, but the meeting was described as unusually contentious, with several officials pushing for a less dovish stance and one governor advocating a cut. Powell’s decision to remain on the board was framed as a sign of concern over Fed independence. Inflation remains sticky and may worsen (Priority: 5/5): The hosts highlighted core and headline PCE inflation staying well above target, plus rising gasoline prices and elevated inflation expectations. This was used to argue that the next Fed move could be less likely to be a cut and could even become a hike over time. GDP and consumer spending slowed (Priority: 5/5): Second-quarter growth tracking was revised lower after a batch of data, and the hosts said the overall GDP number was mediocre. Consumption softened, the saving rate fell, and the week’s data were judged concerning given war-related energy shocks. How economists use Claude at Anthropic and elsewhere (Priority: 4/5): McCrory explained that Claude helps with classic economist tasks like data downloads, regressions, charts, and coding, but also expands what economists can do by enabling dashboards, rapid iteration, and better communication across teams. AI productivity gains may be large but are mediated by adoption (Priority: 5/5): McCrory argued current-model usage suggests substantial task-level time savings and potentially large economy-wide productivity effects, but actual measured gains will depend on business adoption, data modernization, and whether tasks are substitutable or only augmentative. Labor-market risks are uneven and may hit young workers first (Priority: 4/5): The conversation explored whether AI will displace jobs. McCrory said broad unemployment effects are not yet visible, but exposure is higher in roles like data entry, technical writing, and some programming, with early-career workers showing more anxiety and some signs of weaker hiring. Broader benefits and under-discussed risks of AI (Priority: 3/5): Beyond productivity, McCrory emphasized AI’s potential to accelerate innovation, health, science, and human flourishing. A less-discussed risk he highlighted is that AI adoption could amplify downturns if recessions encourage firms to automate faster.

Key Arguments: The Fed’s reluctance to signal cuts is driven by high inflation, war-related energy pressures, and changing internal politics at the Board. Powell’s decision to stay on the Fed board underscores concern about institutional independence and possible future legal pressure. GDP growth around 2% is weak relative to expectations, especially with consumer spending slowing and the saving rate low. Claude is most valuable not only as a speed tool but as a scope-expanding assistant that helps users do more kinds of work. Human expertise remains essential because the model can make subtle analytical mistakes and still needs oversight, prompting iterative human-AI collaboration. Task-level time savings can be enormous—especially for information synthesis and analysis—implying meaningful potential for aggregate productivity growth. The realization of AI productivity gains depends on complementary investments, organized data, and business adoption; otherwise effects may stay hidden in measured GDP. AI’s labor-market impact is likely to be uneven: augmenting skilled workers while posing more risk to routine tasks such as customer support and data entry. Evidence of displacement is limited so far, but younger workers in highly exposed occupations may already be seeing weaker hiring. A major upside of AI is not just higher GDP but faster innovation in science, health, and quality of life.

Data Points: Fed policy rate: unchanged - The Fed meeting ended with no change in interest rates. Probability of a June rate cut: 7% - Co-host Marissa cited market-implied odds for the next meeting. Probability of a June rate hike: 0% - The panel said markets do not expect a near-term hike. Core PCE inflation: 3.2% y/y - Discussed as the Fed’s preferred inflation gauge, excluding food and energy. Headline PCE inflation: 3.5% y/y - Used to show inflation remains above target. Five-year breakeven inflation: high end of the 4-5 year range - Mentioned as signaling elevated inflation expectations. GDP growth tracking estimate: 2.6% - The host’s nowcast was revised down after trade data. Alternative GDP model estimate: 2.0% - Justin’s current-quarter model estimate was described as exact. Q4 GDP growth: 0.5% - Referenced as a weak prior quarter, likely affected by shutdown effects. Private consumption plus investment growth: 2.5% - Identified as the stronger component within GDP composition. Saving rate: 3.6% - Noted as near historical lows and a concern for consumer resilience. Gasoline price: $4.40 per gallon - Cited as an example of energy-driven inflation pressure. Anthropic survey sample: 81,000 users - Used to study how Claude users perceive AI’s economic impact. Task-level time savings: 80-90% - McCrory described estimated time savings for some Claude-assisted tasks. Projected labor productivity lift: 1.8 percentage points per year - Anthropic’s estimate if current task-level gains diffuse over 10 years. Alternative productivity lift including task failure/rework: about 1.0 percentage point per year - Adjusted estimate after accounting for cases where Claude requires more human effort. Economists’ AI productivity consensus: about 0.5 percentage point TFP / 0.7-0.8 percentage point labor productivity - Mark Zandi contrasted Anthropic’s estimate with broader forecaster assumptions. Current labor market exposure gap: 95% potential vs about one-third observed - For computer-mathematical jobs, large language models could handle most tasks in principle, but observed use is much lower. Model capability horizon doubling time: 4-7 months - McCrory cited rapid improvement in the task horizon models can reliably complete. Diffusion speed vs past technologies: 5-10 times faster - Anthropic data suggested AI adoption across regions is much faster than 20th-century technologies.

Pivotal Quotes: "The next move might not be a cut. It might be an actual increase." — Mark Zandi: Used to summarize the Fed’s more hawkish tone amid persistent inflation. "It’s not just about doing what you’re already doing faster, but also doing more in a different range of tasks." — Peter McCrory: McCrory’s central explanation of AI’s productivity effect. "The albatross theory." — Mark Zandi: Referring to the paper’s finding that older-worker organizational effects can impede productivity and tech adoption.

Implications: AI appears likely to raise productivity materially, but the near-term story may be uneven labor churn, not mass unemployment. For firms, adoption quality and data readiness will matter as much as model capability; for policymakers, inflation, rates, and labor-market adjustment remain tightly linked.