Inside the expert network training every frontier AI model

Garrett Lord, co-founder and CEO of Handshake, discusses how his career-networking platform built a fast-growing AI data labeling business by leveraging its network of PhDs and students to create expert training data for frontier AI labs. The conversation covers what post-training data work actually involves, why experts have replaced generalists, and how he incubated the new business inside an existing company.

01Key learnings

• Post-training, not pre-training, now drives most model gains, so high-quality expert data targeting specific capability gaps is where value concentrates.
• The labeling market has shifted from cheap generalist labor to domain experts, so the right supply is credentialed specialists like PhDs and master's students.
• Access to a trusted audience is the real moat in human data; owning an audience removes customer acquisition costs compared with competitors buying ads and recruiting on LinkedIn.
• Model builders care most about quality, then volume, then speed; fast turnaround on data that measurably improves a model is what wins repeat business.
• When incubating a new business inside an existing one, dedicate separate teams, offices, onboarding, and compensation so people aren't split between responsibilities.
• Use a metrics-driven operating cadence from day one and make one directly responsible individual per initiative, chosen for capability rather than function.
• Look at what your company already owns that could become a new opportunity when a technology shift changes what's valuable, as Handshake did with its student network.

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.