Brendan Foody on Teaching AI and the Future of Knowledge Work

What It Covers

Brendan Foody, 22-year-old CEO of Mercor, discusses building the fastest-growing AI company by hiring experts to train frontier models, achieving 64% automation of economically valuable tasks and creating new job categories for knowledge workers.

Key Questions Answered

• •AI Performance Metrics: Frontier models improved 25-30% annually on economically valuable tasks, with GPT-5 scoring 64% on real-world professional work compared to GPT-4's baseline. This measures actual job automation potential rather than academic benchmarks, surveying experts on time allocation and creating corresponding evaluation rubrics.
• •Data Collection Strategy: The most valuable training data includes rubrics for grading model outputs and test question-answer pairs with scoring criteria, not just raw content. Models learn by attempting problems repeatedly and receiving scored feedback, similar to how professors grade essays with structured evaluation frameworks.
• •Labor Market Transformation: Within five years, majority of high-end knowledge workers will transition from performing tasks to training AI agents and building evaluation environments. This creates new job categories where investment bankers build testing frameworks instead of conducting analyses, applying fixed-cost knowledge investments across unlimited agent deployments.
• •Hiring Methodology: Effective talent assessment measures actual job skills through concrete projects rather than vibe-based conversations about background and cultural fit. Companies over-index on personal similarity instead of testing candidates' ability to perform specific deliverables like analyzing data rooms or drafting legal documents under realistic conditions.
• •Model Limitations Timeline: AI excels at tasks completable in chat windows but cannot yet draft emails or schedule meetings. Long-horizon tasks spanning 50-100 hours remain challenging. Foody predicts models will struggle to find mistakes when questioned by domain experts within six months, though automating the final 25% of expert capabilities will take significantly longer.