Brendan Foody on Teaching AI and the Future of Knowledge Work
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
At just 22, Brendan Foody is the youngest unicorn founder on record and CEO of Mercor, one of the fastest-growing AI companies in history. In this conversation with Tyler Cowen, Foody explains how Mercor hires experts across diverse fields—from poets to economists to Nobel laureates—to train frontier AI models. (01:25) The company creates evaluations and rubrics that help AI labs measure and improve their models' performance on economically valuable tasks, paying premium rates like $150 per hour for poets to grade AI-generated poetry. (01:19) • The discussion spans the future of work, labor markets, education, and the rapid advancement of AI capabilities across professional domains
Speakers
Tyler Cowen
Tyler Cowen is a renowned economist, professor at George Mason University, and host of the popular Conversations with Tyler podcast. He is the author of numerous books on economics and culture, co-founded the economics blog Marginal Revolution, and serves as Chairman and General Director of the Mercatus Center at George Mason University.
Brendan Foody
At 22, Brendan Foody is the youngest unicorn founder on record and CEO of Mercor, an AI company founded in early 2023 that has become one of the fastest-growing startups in history. He's a former Thiel Fellow who dropped out of college to build Mercor, which hires experts worldwide to train frontier AI models. Foody has dyslexia and was a competitive speech and debate participant in high school, co-founding the company with his former debate teammates.
Key Takeaways
Focus on Economic Value Over Academic Metrics
The biggest disconnect in AI research is the focus on academic evaluations like PhD-level reasoning tests versus real-world economic outcomes. (04:43) Foody's team worked with experts like Larry Summers and Cass Sunstein to create the AI Productivity Index (APEC), which measures how well models perform on actual professional tasks rather than abstract academic benchmarks. They survey hundreds of experts in each field to understand how they spend their time, then create corresponding prompts and rubrics based on actual work activities. This approach reveals that frontier models like GPT-5 now score 64% on economically valuable tasks, representing a 30% year-over-year improvement rate that could transform entire industries.
Rubrics Are More Valuable Than Raw Data
While most people think about feeding AI systems more text and information, the real breakthrough comes from creating better ways to measure success. (13:25) Foody explains there are two types of data: output data (curriculum for models to read) and measurement data (rubrics, tests, and scoring systems). The measurement data is far more valuable because it allows models to attempt problems repeatedly, get scored, and learn from their mistakes. This is why Mercor pays premium rates for experts who can create sophisticated evaluation frameworks—these rubrics become the "new oil" that powers AI advancement.
The Future Workforce Will Train AI Agents
Rather than simply automating jobs away, Foody predicts a fundamental shift where knowledge workers transition from doing repetitive tasks to building reinforcement learning environments and training AI agents. (23:00) Investment bankers won't analyze data rooms repeatedly—they'll teach models how to do it once, then use those trained agents multiple times. This mirrors the software development model of building something once and using it many times. Within five years, Foody expects a majority of high-end knowledge workers will be training models either in their full-time jobs or through platforms like Mercor.
Skills Assessment Beats Cultural Fit in Hiring
Most companies make critical hiring mistakes by focusing on "vibes-based" conversations about where candidates grew up or whether interviewers would enjoy hanging out with them, rather than measuring actual job-relevant skills. (35:36) Foody advocates for giving candidates real projects that mirror the work they'll actually do, then grading their performance objectively. This approach has helped Mercor scale to over 300 employees rapidly while maintaining high talent quality. The key insight is measuring someone's actual capabilities rather than relying on cultural fit or personal connections, which often introduce bias and miss top performers.
Labor Markets Need Structural Aggregation
Current labor markets are massively inefficient because everything is disaggregated—job seekers apply to dozens of positions while companies only consider a tiny fraction of available candidates. (39:22) The solution requires both better AI for matching and a structural shift toward platforms where everyone applies to one place and every company hires from that same pool. LinkedIn has the distribution but lacks the sophisticated matching capabilities needed. As work becomes more fractional and remote, with people training AI models rather than doing full-time traditional roles, this aggregated approach becomes even more critical for enabling global talent matching and economic mobility.
Statistics & Facts
- GPT-5 currently scores 64% on economically valuable tasks measured by the AI Productivity Index (APEC), representing a 30% year-over-year improvement rate from the previous frontier model GPT-4. (06:02)
- Mercor employs over 300 full-time employees across the world as of October 2025, despite being founded in early 2023. (44:52)
- The company pays poets $150 per hour to grade AI-generated poetry and create evaluation rubrics, demonstrating the premium value placed on expert human judgment in training AI systems. (01:19)
