Lenny's Podcast

Garrett Lord: How Handshake feeds every frontier AI lab now

How expert trajectories from chemists, coders, and teachers feed frontier labs; Lord on post-training, audience as the only moat, and a new Handshake unit.

Garrett LordguestLenny Rachitskyhost

Aug 23, 20251h 9mWatch on YouTube ↗

WHAT IT’S REALLY ABOUT

Handshake turns student network into explosive AI training data powerhouse

Handshake CEO Garrett Lord explains how the company leveraged its massive network of students, graduates, and experts to build a hypergrowth AI data-labeling business on top of its decade-old core marketplace. As frontier labs shifted from broad internet pre‑training to post‑training with highly specialized human data, Handshake realized its 18M+ users—especially 500K PhDs and 3M master’s students—were an ideal expert supply. In under a year, the new Handshake AI unit scaled from zero to over $50M ARR, on track to surpass $100M and rival the $200M ARR of the original jobs platform. The conversation breaks down how post‑training works, why expert data is now the bottleneck, how Handshake structurally outcompetes legacy data-labeling firms, and how they successfully incubated a startup‑style business inside a mature company.

IDEAS WORTH REMEMBERING

5 ideas

AI gains have shifted from scraping the internet to post-training with expert data.

Frontier labs have largely exhausted the public web for pre‑training, so most performance improvements now come from post‑training—supervised fine‑tuning, preference ranking (RLHF), trajectories, and multimodal data built by domain experts.

The bottleneck is no longer cheap generalist labor but high-quality experts.

Models are now strong enough that low-cost generalist labelers add limited value; labs need top physicists, chemists, teachers, coders, and other specialists to locate failure modes and produce ground-truth reasoning paths in advanced domains.

Owning an engaged audience is the core moat in human data.

Where competitors spend heavily on recruiters and ads to find experts, Handshake already has trusted relationships and rich profiles for millions of students and professionals, enabling near-zero CAC, better targeting, and higher retention on projects.

Expert work is curated, trained, and measured like a scientific process.

Handshake builds training cohorts, instructional design, internal post-training teams, and GPU-backed evals so each unit of data is checked for quality and actual model gain before it’s scaled or sold across multiple labs.

Building a new AI business inside an existing company requires separation and founder-level focus.

Handshake AI was spun up with its own org, metrics cadence, hiring bar, compensation, and physical space, with Lord spending ~80% of his time on it and redeploying top talent solely to the new unit to avoid legacy drag.

WORDS WORTH SAVING

5 quotes

The only moat in human data is access to an audience.

— Garrett Lord

The models have gotten so good that the generalists are no longer needed. What they really need is experts.

— Garrett Lord

There will never be a time like this. I've never seen anything like it… where there's unlimited demand.

— Garrett Lord

For as long as models are improving, humans will be needed in this process.

— Garrett Lord

Being AI-native, young people are at a huge advantage.

— Garrett Lord

Pre-training vs. post-training in modern AI model developmentRise of expert-driven data labeling and RLHF/SFTHandshake’s pivot from student job marketplace to AI data providerStructural advantages of owning a trusted, at-scale expert networkBuilding and operating a zero-to-one startup inside a 10-year-old companyImpact of AI on early-career jobs and “AI-native” workersFuture of human-in-the-loop data, synthetic data, and evolving data types

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.