Garrett Lord on Post-Training Data, Expert Networks, and Handshake's Second Act

Garrett Lord on Post-Training Data, Expert Networks, and Handshake's Second Act

transcriptaidata-labelingstartupshandshakecompany-building

Garrett Lord on Post-Training Data, Expert Networks, and Handshake’s Second Act

Source: Lenny’s Podcast Speaker: Garrett Lord Link: Episode

Overview

Garrett Lord, co-founder and CEO of Handshake (LinkedIn for college students, ~18 million users, 100% Fortune 500 adoption), describes how his team pivoted from student recruiting into AI data labelling in early 2025, reaching $50 million ARR in four months and on pace to exceed $100 million ARR in year one — surpassing in under two years the revenue of a business it took a decade to build. The first half of the episode is a clear tutorial on what post-training actually involves; the second half is an unusually candid account of how to incubate a new company inside a mature one.

Key ideas

  • Post-training is now where model gains come from. Pre-training — feeding models the entire internet — has asymptoted. The frontier has shifted to post-training: supervised fine-tuning (SFT), reinforcement learning with human feedback (RLHF), trajectory data (screen + mouse + voiceover of step-by-step problem-solving), and rubric-based evaluation. Labs are running experiments, testing hypotheses, and scaling whichever pipeline improves benchmarks. Experts, not generalists, are needed to find where advanced models break.
  • The only moat in human data is access to an audience. Competing platforms recruit PhDs by sending 200 LinkedIn messages and spending tens of millions on Google and Instagram ads — at high cost, low trust, and poor retention. Handshake has 18 million users, 500,000 PhDs, 3 million master’s students, and a decade of brand trust with 1,600 partner universities. Zero customer acquisition cost; high conversion; high LTV. The structural advantage is audience access, not technology.
  • Zero to $50M ARR in four months; on pace to exceed $100M in year one. Handshake began exploring AI data labelling in December 2024, signed its first frontier lab contract five months earlier, and was working with all seven major frontier labs by the time of recording. The new business is on a trajectory to exceed the revenue of the decade-old core business within two years.
  • Building a second company inside an existing one requires total isolation. Garrett ran the new business in founder mode — 80%+ of his time, separate engineering team, separate design, separate finance, separate all-hands, separate office area, separate recruiting. Compensation was tied to hurdles in the new business. DRI (directly responsible individual) ownership was assigned by capability, not organisational function. Metrics cadence from day one. Key people from the core business were asked to fully transfer — no split responsibilities.
  • “Leave nothing to chance” as an operating mode in a window of unlimited demand. When demand is structurally uncapped, execution speed and quality trust become the binding constraints. Handshake AI’s operating mantra: get on the plane, make the late-night push, check the data again, ship the extra feature. The risk of growing too fast and eroding lab trust is treated as the primary strategic threat.