Anthropic's Co-Founder and Top Economist on Doing Research at the AI Frontier

Episode Summary

Executive Summary: Odd Lots interviews Jack Clark and Peter McCrory of Anthropic about AI’s rapid progress, its still-muted macroeconomic footprint, and the policy/safety frameworks needed as models become more capable. The conversation spans recursive self-improvement, labor-market effects, hiring changes, productivity measurement, national-security risks, and how Anthropic’s public-benefit mission shapes its research and disclosure.

Main Topics: AI’s pace of progress and recursive self-improvement (Priority: 5/5): Jack Clark describes AI as advancing exponentially across vision, sound, code, and other domains, with Anthropic seeing its own production function improve as models help engineers write far more code. Macroeconomic impact and productivity measurement (Priority: 5/5): Peter McCrory argues AI’s economic effects are real but still diffusing; Anthropic is trying to measure time savings and productivity gains to estimate labor-productivity effects and inform policymakers. Labor-market and hiring shifts (Priority: 5/5): The guests discuss a barbell hiring pattern: more demand for senior staff and AI-native newcomers, while some early-career tasks are increasingly automated or delegated to models. Safety, alignment, and existential risk (Priority: 5/5): The discussion covers alignment failures, model misbehavior in tests, and the need for the option to slow or pause deployment if radical misalignment ever emerged. Policy, regulation, and national security (Priority: 4/5): They advocate for technocratic oversight, third-party testing, transparency rules, and KYC-style controls for frontier AI, especially where bio/cyber risks intersect with commercial deployment. Anthropic’s public-benefit model and institutional design (Priority: 4/5): The company’s economics and social-science work is presented as a way to make AI impacts legible to the public and to guide both product decisions and policy responses. Enterprise adoption, organizational bottlenecks, and trust (Priority: 4/5): Even when model capability is strong, deployment depends on contextual data, workflow redesign, and trust; large firms may benefit, but bureaucracy and tacit knowledge remain obstacles.

Key Arguments: AI progress looks exponential across multiple modalities, suggesting a general-purpose technology rather than a niche tool. The economic effect is lagging because capabilities must diffuse through organizations, data systems, and workflows before showing up in aggregate statistics. Anthropic’s internal data suggest engineers now write about eight times as much code as in 2021-2024, indicating significant productivity gains inside the company. AI is currently more of a labor-augmenting, skill-biased technology than a full substitute for cognitive labor, which helps explain why the labor market still looks relatively normal. Hiring is shifting toward senior experts and AI-native younger workers, while middle layers and routine implementation tasks are under pressure. Frontier AI requires new governance tools: systems for evaluating model properties, third-party audits, transparency, and potentially KYC-like controls for sensitive deployments. The largest risk may be misuse, weak oversight, or poorly designed deployment rules rather than immediate extinction, but the field still treats existential risk as a real tail risk. Safety can be a competitive advantage because trust, reliability, and serviceability matter to customers; being safe does not necessarily mean sacrificing performance.

Data Points: Code output at Anthropic: 8x more code - Jack Clark says engineers in 2026 are writing about eight times the amount of code they did in 2021-2024. Research time horizon for productivity impact: Next decade - Anthropic’s growth-accounting estimate discusses labor-productivity gains diffusing over roughly ten years. Estimated labor-productivity increase: 1.8 percentage points per year - Peter McCrory says Anthropic’s estimates point to labor productivity growth rising by 1.8 percentage points annually over the next decade. Survey sample size: 81,000 people - Peter references a large-scale qualitative survey on hopes and fears about AI conducted by Anthropic’s Societal Impacts team. AI adoption timing in the report: March report - Peter notes recent findings on young workers in high-AI-exposed roles using data from a March report. Macro context: Largest non-recessionary labor market slowdown on record - Peter uses this to explain why early-career job-finding trends are hard to interpret. Potential diffusion lag: 1-2 years - Jack says he expects to point policymakers to a steeper graph in a year or two as diffusion accelerates.

Pivotal Quotes: "AI is more important than anything else, so I felt best to sort of optimize for that above all else." — Jack Clark: Clark recalls his 2016 decision to leave journalism and focus on AI. "I can't reconcile that with the world staying normal for long, but it's going to take a while for that to diffuse into the world and change it." — Jack Clark: Clark reacts to Anthropic engineers writing far more code with model assistance. "The bulk of the risk is us messing it up, like through misuse or ignoring risks or not setting up the right policy environment." — Jack Clark: Clark explains where he sees the greatest near-term danger in AI.

Implications: AI may reshape productivity, hiring, and regulation faster than official statistics reveal. Firms that combine capability with trust and measurement may win, but policymakers will need faster, more technocratic oversight to manage bio/cyber risk and labor disruption.