Conversations with Tyler — Brendan Foody on Teaching AI and the Future of Knowledge Work
Date: January 7, 2026
Guest: Brendan Foody, CEO & Co-founder of Mercor
Host: Tyler Cowen
Episode Overview
In this episode, Tyler Cowen sits down with Brendan Foody, the 22-year-old CEO and co-founder of Mercor, an AI company transforming how advanced models are trained and evaluated in knowledge work. Their conversation covers expert involvement in teaching AI, the pace of AI’s economic impact, the future of labor and education, how evaluation data shapes AI model abilities, and what it means for expertise, hiring, and human flourishing. They also touch on Brendan's entrepreneurial background and personal experiences, offering a rich look at the near future of knowledge work alongside the realities facing today’s tech founders.
Key Discussion Points & Insights
The Mercor Model: Paying Poets $150/Hour to Teach AI
- AI Training by Domain Experts: Mercor hires leading experts—including poets, lawyers, doctors, and economists—to create evaluation rubrics and benchmarks for AI training (01:22).
- Attracting Talent: They can offer premium rates as the knowledge imparted by experts reaches billions of users via AI models (01:22–02:03).
- Evaluating AI with Human Taste: Subjective domains (like poetry) require not just consensus but thoughtful disagreement among graders to capture nuance and edge cases (02:35–03:05).
- Expertise Beyond Academia: Mercor favors experts with broad real-world perspectives, as well as academic achievement (04:14–05:21).
“When we have these phenomenal poets that teach the models… once they’re then able to apply those skills and that knowledge across billions of users, hence allowing us to pay $150 an hour for some of the best poets in the world.”
— Brendan Foody, 01:22
Measuring AI Progress and Economic Impact
- Apex: The AI Productivity Index: Developed with advisors like Larry Summers (economics), Cass Sunstein (law), and Eric Topol (medicine) to measure AI progress in “economically valuable tasks” rather than just academic benchmarks (04:14–06:21).
- Rapid Advancement: AI models improving 25–30% per year in productive tasks, with challenges in high-stakes fields demanding near-100% accuracy (06:25–08:09).
“The largest takeaway is the rate of model improvement at economically valuable tasks is incredible.”
— Brendan Foody, 05:57
- Lingering Gaps: While AIs are approaching superhuman performance in defined tasks, they still struggle with:
- Long-horizon tasks (projects taking >50 hours)
- Task-switching and integrating multiple tools
- Capturing “taste” or judgment in ambiguous domains (e.g., law, poetry)
“Models still can’t draft an email for us, can’t schedule a meeting… There’s a long way before we’re able to tell a model: ‘Go off and build a startup for 90 days.’”
— Brendan Foody, 28:42
What Data Improves AI — The Value of Rubrics and Taste
- Two Data Types:
- (1) Outputs and transcripts (e.g., classroom discussion)
- (2) Evaluation data: Rubrics defining success, test cases, ways to measure output quality (13:50–14:30)
- The latter is “the most valuable” for improving models.
- For Poetry: Optimal data includes not just poet discussion, but clear rubrics and criteria for what counts as a great poem (16:53–17:58).
- Limits of Rubrics: Some forms of taste resist reduction to checklists. RLHF (Reinforcement Learning from Human Feedback) helps, by having experts “choose between” outputs many times so models learn preferences (18:18–18:41).
- Taste Across Time: Foody predicts future models will be able to reflect aesthetic preferences of any era—“taste from every different decade”—and personalize responses accordingly (22:10–22:30).
The Future of Knowledge Work and RL as an Economic Engine
- Rise of RL (Reinforcement Learning): Much of knowledge work will shift toward teaching AI agents — training, correcting, and creating RL environments (24:46–25:25).
- Big Economic Shift: Investment bankers, customer support, and more will move from repetitive work to training AI once, then deploying at scale—akin to software’s transition from manual to automated processes (24:46–25:25).
“A huge portion of the economy will become an RL environment machine.”
— Brendan Foody, 24:46
AI, Privacy, and Individual Agency
- Personal Data & AI: Individuals could use AI layers to filter their own conversation recordings before sharing for model improvement, but privacy concerns remain real and persistent (24:25–25:49).
- Trusted Brands Will Have Edge: Companies like Apple, with reputations for privacy, might win user trust for this new wave of data-driven personalization (25:49–26:25).
Expertise and Social Perceptions
- Status of Human Experts: Despite rising AI competence, the “last 25%” of expertise remains elusive for machines—human experts still needed for edge cases and complex judgments (27:09–28:17).
- AI as a ‘Distilled’ Expert: Possibility that people might trust “impersonal” AI expertise more, but still revere human experts for the hardest cases (26:25–28:17).
Jobs Created by the Rise of AI
- Trainers of AI: Rather than replacing all jobs, new high-end roles focus on building, testing, and refining RL environments and agents across fields (29:35–30:29).
- Skills Needed: Future trainers don’t need to be technical AI experts—just domain specialists able to spot and describe mistakes in model performance (30:36–31:18).
“They just need to know about the thing. The only element of technical AI they’ll need is to find where the model makes a mistake.”
— Brendan Foody, 30:36
Elasticity and the Future of Labor Markets
- High Price Elasticity in Software: Increased AI productivity could increase the number of engineers and software produced (31:21–31:55).
- Aggregation and Matching Problem: The big inefficiency in labor markets is disaggregated matching. In the future, AI “aggregators” may optimize matching on a global scale, moving beyond resumes to deep performance data (39:20–41:03).
- Potential for Nepotism’s Return: As credentials and AI-prepped applications proliferate, referrals and insider recommendations may gain renewed importance, but AI may help counterbalance this with better data (41:31–42:06).
Education and Teaching
- AI Tutors for All: Envisioned world where everyone has a 24/7 AI tutor, like “personal Sal Khan,” reshaping motivation and information access (33:18–33:47).
- Teachers Still Matter: While core instruction can be automated, personal relationships and mentorship-type teaching will remain essential (34:03–34:39).
Company-Building and Hiring at Mercor
- AI in Hiring: Mercor’s own hiring processes are automated, focusing on skills and project-based evaluations rather than “vibe”-based interviews (34:52–36:08).
- Project-Based Tests vs. Background Checks: For ambiguous roles, drilling into candidates’ prior similar experiences can be effective (36:41–37:26).
- Measurement over Social Cues: Body language and articulateness can mislead; focus should be on direct measures of work-relevant skills (37:30–37:59).
Notable Quotes & Memorable Moments
-
On Taste and Rubrics:
“Immanuel Kant… said, in essence, taste is that which cannot be captured in a rubric. And if the data you want is a rubric, and taste is really important, maybe Kant was wrong, but how do I square that?”
— Tyler Cowen, 17:55 -
On Reinforcement Learning Society:
“Instead of the investment banker redundantly analyzing a data room… they’ll teach the model how to do that once… Similar to building software once, they’ll be able to use that many times as they use their agent… That’s why I believe a huge portion of the economy will become an RL environment machine.”
— Brendan Foody, 24:46 -
On AI’s Next Leap:
“I would be shocked if we don’t have enormously capable models across those dimensions… long horizon tasks, in the next six to twelve months.”
— Brendan Foody, 09:04 -
On Future Data Needs:
“If people have tests that models are bad at, that map to a meaningful amount of economic value… that’s super exciting for us.”
— Brendan Foody, 14:43 -
On Human-Machine Comparison in Expertise:
“My read on the market is that models are advancing very quickly at automating 50–75% of what humans and experts are able to do, but will really struggle with that last 25%.”
— Brendan Foody, 27:09
Timestamps for Major Segments
| Timestamp | Segment/Topic | |:---:|:---| | 01:01–03:15 | Why pay $150/hr for poets? Mercor’s expert-driven AI model | | 04:14–06:36 | AI Productivity Index (Apex), expert selection, measuring economic impact | | 09:45–12:05 | Limits of AI: Taste, long-horizon tasks, when AIs can rival experts | | 13:50–16:58 | The value of evaluation data, rubrics vs. “taste,” ideal data for social science/poetry | | 18:18–19:22 | RLHF and subjective domains, optimizing for user vs. expert preferences | | 21:32–22:30 | Teaching AI taste from different eras, personalization of style | | 24:46–25:25 | Society as a RL engine, shift from repetitive work to AI agent training | | 27:09–28:17 | Status of experts as AI gets better—“last 25%” challenge | | 29:35–31:18 | New jobs: AI trainers, requirements for expertise | | 33:18–34:39 | AI tutors, future of teaching and teacher roles | | 34:52–37:26 | AI-driven hiring, skills over vibes, challenges for non-technical roles | | 39:20–41:03 | Making labor markets efficient, global matching/aggregation | | 41:31–42:06 | Nepotism, mentors, and the AI-driven job market | | 49:43–51:19 | Brendan’s 8th grade donut company—entrepreneurial lessons | | 51:19–52:59 | Extemporaneous speaking, debate, and cognitive styles | | 53:22–55:17 | Dyslexia and entrepreneurship: creativity and strengths | | 56:12–56:42 | Cultural/dating issues among young tech founders | | 57:10–57:46 | Favorite food spots in San Francisco | | 58:00–58:10 | Why the company is named Mercor; roots in Latin | | 59:08–60:43 | Next goals for the company, intersection of labor markets and AI |
Closing Reflections
- On What’s Coming:
Foody sees the next wave of AI not as the end of work, but as a force shifting high-end knowledge workers into model trainers and RL environment builders, with human expertise essential for the toughest domains. - On Data and Expertise:
The future hinges on curating the right diagnostic and evaluative data, especially in fields like law and poetry where subjective judgment remains paramount. - On Talent:
Skills-based hiring, project-driven evaluations, and recognizing unique strengths (like those developed from dyslexia) lay the foundation for building fast-growing, effective organizations—even as the nature of work is transformed by AI.
For Further Listening
Find more episodes and full transcripts at conversationswithtyler.com. For research and info on Mercor’s work, visit mercatus.org.
Summary compiled using the episode transcript. All quotes are attributed and timestamped per guidance.
