Episode Summary
Executive Summary: Brendan Foote argues that AI-driven hiring and evaluation will become a core infrastructure layer for the economy. He explains how Mercor uses LLMs to predict job performance, why eval creation is now the bottleneck for model progress, and how labor markets may shift toward global, more efficient matching. He also warns that knowledge-work displacement will be rapid and politically disruptive.
Main Topics: Mercor’s business: AI-based talent prediction (Priority: 5/5): Mercor automates resume review, interviewing, and candidate ranking using LLMs to predict job performance better than humans, serving top AI labs and other employers. Evals as the bottleneck for AI progress (Priority: 5/5): As models become stronger at reinforcement learning, the limiting factor becomes creating evaluations for economically valuable tasks across industries, not just benchmark tests. Talent signals, power laws, and hidden indicators (Priority: 4/5): Foote says knowledge work is often power-law distributed, so identifying outliers matters more than average performance. He emphasizes online artifacts, interview transcripts, and domain-specific signals. Labor market displacement and reskilling (Priority: 5/5): He predicts rapid displacement in knowledge work, with many workers moving toward physical-world jobs, human interaction roles, or niche tasks, creating major political and economic pressure. Reinforcement fine-tuning and model customization (Priority: 4/5): RFT is presented as a data-efficient way to train models on business-specific outcomes, enabling enterprise agents to learn what good performance looks like with relatively few examples. The future of hiring and marketplaces (Priority: 4/5): Mercor’s long-term vision is a global unified labor market where software automates matching, assessment, and even management, with humans and agents working together. What humans should learn next (Priority: 3/5): Foote advises children and workers to focus less on narrow coding skills and more on taste, general reasoning, adaptability, and contrarian problem-finding.
Key Arguments: AI systems are already often better than human hiring managers at evaluating candidates from text-based inputs and interviews. The most valuable data for AI training is shifting from low-skill crowdsourcing to expert vetting and real-world performance prediction. Knowledge work is highly power-law distributed, so accurately identifying top performers has outsized economic value. Models are strongest where tasks are text-based, high-volume, and verifiable; they are slower on multimodal, motivational, and taste-based judgments. The real bottleneck for enterprise AI is not raw model reasoning but building domain-specific evals, workflows, and tool-use feedback loops. RFT can customize models much more efficiently than traditional supervised fine-tuning, making enterprise adaptation practical. Job displacement will likely be fast in digital work and much slower in physical-world roles, causing social and political friction. Future labor markets will be more global, software-mediated, and potentially hybridized with AI agents competing alongside humans.
Data Points: Mercor founding year: 2023 - The company was founded by three college dropouts and Teal fellows. Funding raised: $100 million - Mercor has raised this amount since founding. Revenue run rate: over $100 million - The company surpassed this milestone during its rapid growth phase. Labor marketplace supply-demand ratio: 50:1 - Foote cites this as the average ratio of supply side to demand side in labor marketplaces. Candidate pool considered by a San Francisco company: a fraction of a percent of people in the world - He uses this to illustrate how fragmented and manual hiring is today. Typical SFT scale: a few hundred to tens/hundreds of thousands of examples - Used to contrast with RFT, which he says is far more data efficient. RFT sample efficiency: hundreds to thousands of examples - Foote says this is enough to customize models for some enterprise use cases. Model evaluation scope: customer support, consulting, software engineering, physicians, lawyers, hobbyists, video games - Examples of the broad range of roles and tasks Mercor/AI labs are evaluating. Time horizon for some eval buildouts: years-long - He says even verifiable domains like software engineering require long buildouts to evaluate properly.
Pivotal Quotes: "it'll almost be irrational to not listen to the model" — Brendan Foote: On the future of AI-based hiring decisions and trust in model recommendations. "the largest barrier to automating most knowledge work in the economy" — Brendan Foote: Describing why agent eval creation matters more than benchmark performance. "it's almost just this structural part of building labor marketplaces" — Brendan Foote: Explaining why many applicants never convert to jobs and why Mercor builds free tools to improve marketplace liquidity.
Implications: Expect hiring, performance management, and labor matching to become increasingly AI-mediated. Enterprises may adopt specialized evals and RFT to automate more knowledge work, while workers will need adaptability, taste, and transferable reasoning to stay valuable.