Brendan Foody, Mercor's CEO, says AI models already beat human hiring managers at assessing talent on most of Mercor's evals.
Foody argues that creating evaluations for agents is the largest barrier to automating knowledge work, and predicts eval creation may become the most common knowledge work job in the world.
Foody predicts job losses will come quickly and painfully, could fuel a populist backlash, and expects companies to use better employee analytics to justify layoffs.
AI Now Hires Better Than Humans, and Reshapes Work
Mercor argues that AI can now judge people better than human recruiters, which lets it build a global labor market and supply the experts who train frontier models. As agents spread, the scarce work becomes defining what good looks like through evals, and judgment tasks that cannot be verified stay human the longest. Displacement will still arrive quickly and painfully.
Takeaways
Models already out-assess human hiring managers. Foody says Mercor's models beat human hiring managers on most of its evals and are superhuman at judging anything measurable in text, such as interview transcripts. Multimodal signals like passion and persuasiveness will take longer. High volume hiring gets automated first.
Human data shifted from crowd labeling to vetted experts. Mercor began as a hiring business unrelated to human data. The market moved from low and medium skilled crowd labeling toward the most capable experts working directly with researchers, and top labs now use Mercor to hire thousands of trainers.
Writing evals for agents becomes the key knowledge work. Foody calls eval creation the largest barrier to automating knowledge work and expects it to take years. Work shifts from repeated variable-cost task solving to a one-time fixed cost of building evals, so it may become the most common knowledge job.
Verifiable tasks fall first, judgment stays human longer. Math and code are verifiable and will be solved quickly. Unverifiable judgment, such as assessing a founder's taste, gives a sparse signal and is much harder to automate. Valuable people will have contrarian ideas and good taste, and expert pay will grow more power law.
Job displacement will be fast and painful. Foody predicts losses in many roles, already reported in customer support and recruiting and concentrated in economic contractions. Better analytics on employee value will drive more layoffs and could cause populist backlash. Physical-world work will automate more slowly.
Mercor aims at a unified global labor market. Foody sees fragmentation as the biggest labor inefficiency. Marketplaces average 50 to one supply to demand, so Mercor plans free tools like AI mock interviews. Customer performance data feeds a flywheel, and the market may include AI agents.
Reinforcement fine tuning is very data efficient. Supervised fine tuning often needs tens of thousands of input output pairs. Foody says reinforcement fine tuning may need only hundreds to thousands of examples, which makes him bullish on it for customization at the application layer.
Hiring advice for early stage companies
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Prioritize talent density and use data driven criteria over vibes.
Use work trials and references as proxies for real outcomes, though trial data is not shared across companies.
Look at underused signals such as GitHub profiles and personal projects of top engineers.
Hosts noted Google hired well but took years to fire people, while Facebook removed underperformers early.