Inside the Expert Network Training Every Frontier AI Model
Garrett Lord
Co-founder & CEO, Handshake
AUG 24 2025
The Hidden Layer
Expert Networks Are the Invisible AI Infrastructure
"The only moat in human data is access to an audience."
Most model gains now come from post-training, not pre-training, since the labs already sucked up the internet
Handshake's network: 500,000 PhDs, 3 million master's students, 18 million professionals
Expert pay: PhDs earn $100–$200 an hour for AI work versus $25 an hour as a teaching assistant
The type of data keeps evolving: CAD files, scientific tool use, trajectories, multimodal, audio
Framework
The Expert Network Model
500K
PhDs on Handshake
$50M
ARR reached in 4 months
Recruited from top schools: PhDs at Stanford, Berkeley, MIT in physics, math, chemistry, biology
Community, cohorts, and an instructional design and assessments team teach experts the tools
Handshake runs its own post-training team to check the quality of each unit of data
Direct-to-lab relationships, cutting out the middleman recruiting platforms
The network moatA decade of trust with 18 million people means no customer acquisition cost. Rivals run 100–200 recruiters on LinkedIn and buy Google, YouTube and Instagram ads to find experts.
Invest in expert onboarding: an instructional design and assessments team, cohorts, and iterative training
The expert career opportunityDomain experts who understand AI training workflows will command premium rates for the next decade. This is a career skill worth investing in.
Contrarian
Expert Knowledge Myths
✗AI will replace domain expertsINSTEAD →✓ AI trained by domain experts will replace the tasks domain experts hate. It will amplify what domain experts do uniquely.
✗Crowdsourcing works for specialized domainsINSTEAD →✓ Crowdsourcing produces average judgments. Specialized domains require specialized expertise.
✗The best models don't need expert feedbackINSTEAD →✓ The best models require the best expert feedback. The bar rises with model sophistication.
✗Expert networks are too slow to scaleINSTEAD →✓ Expert networks scale slowly and compound over time. That's the point, it's a defensible moat.