The people building AI's brain are getting paid to render their own expertise obsolete—and they're doing it between video games and PhD dissertations.
The Summary
- AI training companies like Mercor, Scale AI's Outlier, and Turing are recruiting PhDs and domain experts to probe, test, and grade frontier AI models for what amounts to full-time gig work at "competitive" salaries
- A PhD candidate in game theory juggling EY, LSE teaching, and gaming marathons now spends the equivalent of another full-time job teaching AI systems his discipline—the same systems that will eventually do his job
- This is the white-collar gig economy's strangest iteration: experts paid to transfer their expertise into the machines that will replace them
The Signal
The labor market for AI training has exploded in three years from nothing into a parallel economy of credentialed knowledge workers. Lazaros-Antonios Chatzilazarou represents the archetype: PhD in progress, elite institution, top-tier firm on the resume, recruited via LinkedIn at 1:30 AM because someone's algorithm flagged his game theory expertise. He takes an online skills assessment, passes, and within a week he's developing prompts to test how AI models apply his specialty, then grading their responses. No interview. No office. No manager hovering. Just expertise commoditized into training data.
This isn't Mechanical Turk labor. These aren't people clicking "dog" or "not dog" on image sets for pennies. Companies like Mercor, Scale AI's Outlier platform, and Turing are building networks of domain experts—academics, professionals with specialized knowledge in game theory, advanced mathematics, video production, technical writing—to do RLHF (reinforcement learning from human feedback) work at scale. The compensation is "competitive," a deliberately vague descriptor that likely means low six figures for PhD-level work done as contract labor with no benefits.
"The smartest people in the world are teaching AI to think like them, one prompt at a time."
The irony is structural, not accidental. Frontier labs need human expertise to make AI models better at everything from reasoning to domain-specific problem-solving. But the better these models get, the less they'll need humans to do what those humans are currently teaching them to do. Chatzilazarou is grading AI on game theory applications today. In 18 months, that same AI will likely handle game theory problems better than most humans with his credentials. He's building the thing that makes his expertise less valuable.
The economics here are strange but predictable:
- AI labs need scarce expertise fast and can't hire everyone full-time
- Experts need income and flexibility while pursuing other career tracks
- Intermediary platforms capture margin by matching supply and demand at scale
- Everyone involved knows this work has an expiration date, but the payday is good enough to ignore it
Scale AI's Outlier network doesn't hide what it is. The website literally says "become the expert that AI learns from"—a pitch that assumes you know you're helping something else become you. Turing wants "walk-and-talk" video creators, probably to train multimodal models on how humans speak, move, and present information in public spaces. Every task is a small piece of a larger transfer: human knowledge moving into silicon.
The Implication
This is what the middle of the agent transition looks like. Not mass unemployment yet, but mass participation in your own replacement. If you have specialized knowledge right now, you can get paid well to encode it into models that will eventually do your job faster and cheaper. The gig is real. The money is real. The countdown is also real.
Watch two things: how long these platforms keep recruiting at current pay rates, and what percentage of workers doing this training are doing it because they can't find full-time work in their fields. When "teaching AI" becomes the fallback instead of the side hustle, we'll know the transition accelerated past the point most people expected.