Inside the expert network training every frontier AI model | Garrett Lord (Handshake CEO)

Why Handshake's Student Network Became an AI Training-Data Powerhouse

When product teams and researchers talk about what actually moves the needle for large language models today, the answer increasingly points to high-quality human-created data. Garrett Lord, co-founder and CEO of Handshake, explains how a decade-old campus recruiting platform transformed overnight into one of the fastest-growing suppliers of post-training data for frontier AI labs. The story is as much about product-market fit as it is about recognizing an unfair advantage: access to millions of students, thousands of advanced-degree experts, and a trusted campus brand.

From Careers Platform to Human Data Marketplace

Handshake began as a social careers network for college students and early-career professionals. That long-term accumulation of profiles, academic signals, and university partnerships created a rare asset: a direct, targetable audience of 18+ million users that includes hundreds of thousands of PhDs and master’s students. With labs shifting from pre-training on ubiquitous internet text to post-training that requires specialized, verifiable, and often multimodal data, that audience became a new product.

What Post-Training Data Looks Like

Post-training work covers supervised fine-tuning, reinforcement learning with human feedback, trajectory capture, rubric-based evaluation, and multimodal labeling (audio, video, tool use). Garrett describes how experts—PhD researchers, domain specialists, and professional practitioners—are paid to discover model failure modes, provide ground-truth answers, and record step-by-step tool use. These units of data are often returned as structured JSON and packaged to be directly useful for post-training experiments and evaluation.

Why Experts Matter More Than Generalists

As models become more capable, the low-skill, generalist labor that once sufficed for simple labeling is less valuable. What frontier labs now need are experts who can break models in deep subdomains—mathematics, chemistry, physics, law, medicine—and produce constrained, verifiable datasets that improve reasoning, tool use, and domain-specific capabilities. Handshake's approach elevates contributors from anonymous micro-task labor to trained fellows who receive instruction, assessment, and higher compensation for hard work.

Three Priorities for Model Builders

• Quality: every unit of data must be precise and verifiable to avoid contaminating model behavior.
• Volume: labs need thousands to millions of units across focused hypothesis-driven experiments.
• Speed: rapid iteration allows researchers to test hypotheses and expand only the data pipelines that show gains.

Scaling a New Business Inside an Old One

Launching a disruptive product from within a mature company requires structure and separation. Garrett outlines how Handshake spun up a distinct organization with dedicated engineering, product, design, and operations teams focused solely on the AI data business. Early hires were entrepreneurial and comfortable with ambiguity, processes were metrics-driven, and the culture emphasized extreme ownership and rapid customer feedback.

Competitive Moat: Audience Over Ads

Many competitors buy users through expensive ads and recruiter outreach. Handshake’s decade-long relationships with 1,600 universities and high brand affinity mean near-zero acquisition cost and higher conversion and retention for expert contributors. That audience access becomes the primary moat in human-sourced training data.

The Broader Impact On Careers And Research

Rather than displacing graduates, accessible AI tools combined with paid assessment work can accelerate career outcomes. Young people who are AI-native gain outsized productivity advantages, while PhD fellows earn substantial per-hour rates by doing specialized labeling and evaluation work that simultaneously informs their research and classrooms. For employers and universities, this model promises better talent matching, improved educational design, and measurable benefits to the labor market.

Types Of Data To Expect Next

Future datasets will grow beyond text: CAD files, scientific instrument telemetry, multimodal video trajectories, annotated tool interactions, and richer audio corpora. Synthetic data has a role in verifiable domains, but domain-specific human data will remain essential for many years as labs chase narrow, high-value capability gains.

Handshake’s pivot illustrates how an established network and deep domain trust can be repurposed into a high-velocity human data engine for AI labs pursuing post-training improvements; the result is faster model progress, new work pathways for experts, and a measurable business that shows how access to audience and expert quality can define competitive advantage.