Lenny's PodcastGarrett 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.
CHAPTERS
- 0:00 – 5:55
Handshake’s surprise breakout: building the expert data engine behind frontier AI
Garrett Lord and Lenny frame the episode around Handshake’s new AI-data business, built on top of its 10-year college recruiting network. They preview the core thesis: the new moat is access to expert audiences and the ability to generate high-quality post-training data at speed and volume.
- •Handshake’s new business creates post-training data that advances frontier models
- •Explosive growth context: “unlimited demand” and the urgency to execute
- •Handshake’s strategic advantage: a massive, trusted network with low/no CAC
- •Shift from generalist labeling to expert-driven data creation
- 5:55 – 9:48
Data labeling 101: pre-training vs. post-training and why the market shifted
Garrett breaks down model training into pre-training (ingesting broad internet-scale data) and post-training (targeted data to improve specific capabilities). He explains why pre-training gains have begun to asymptote, pushing labs to seek expert-generated data to keep improving benchmarks.
- •Pre-training: absorbing the corpus of human knowledge (text, video, books)
- •Post-training: targeted capability improvements (coding, math, law, finance, etc.)
- •Frontier labs run hypothesis-driven experiments; data collection follows what works
- •Most current gains in model performance come from post-training
- 9:48 – 13:22
Why experts matter now: finding and fixing the model’s hardest failure modes
The conversation shifts to the role of deep domain experts in identifying where models fail—something general users can’t reliably do anymore. Garrett explains how PhDs can ‘break’ models, isolate failure points, and provide ground-truth answers and reasoning that make models more robust.
- •Experts can surface subtle reasoning and correctness failures in advanced domains
- •Breaking a model involves locating where reasoning steps go wrong, not just final answers
- •As models improve, generalist labeling becomes less valuable; expert work becomes essential
- •Handshake can ‘hyper-target’ niche expertise not represented on the public internet
- 13:22 – 19:52
What experts actually do: GPQA-style tasks, trajectories, and creating usable training artifacts
Garrett gives concrete examples of expert workflows: generating hard questions, validating ground truth, and producing step-by-step reasoning. They introduce trajectory data—capturing full tool-use behavior (screen, mouse, narration)—and discuss how this data is packaged and delivered to labs.
- •Example workflow: create challenge questions, provide correct answers + reasoning steps
- •Trajectory data captures end-to-end human problem solving (tools, actions, narration)
- •Outputs are structured artifacts (e.g., JSON) usable in training pipelines
- •Rubrics help evaluate non-verifiable domains; models can act as judges using rubric criteria
- 19:52 – 24:17
Quality, volume, speed: what frontier labs demand from post-training partners
Garrett outlines the three core constraints for labs: data quality, the ability to generate volume in highly specialized areas, and speed for rapid experimentation. He describes Handshake’s investment in training, assessment, post-training teams, and even GPUs to validate and iterate on data effectiveness.
- •Lab priorities: quality first, then volume, then speed of iteration
- •Hard problem: scaling thousands of high-quality expert items in esoteric fields
- •Handshake builds training/assessment to ensure experts produce consistent, usable data
- •Internal research/post-training and infrastructure help estimate data impact and improve pipelines
- 24:17 – 33:40
Will AI kill entry-level jobs? Why AI-native grads may be advantaged
Lenny raises the concern that students helping train AI may reduce their own job prospects. Garrett argues the opposite: employers see AI as an ‘Iron Man suit’ that increases productivity, and younger workers who grow up AI-native will outperform by leveraging tools earlier and more effectively.
- •Employers report AI increases productivity rather than eliminating junior hiring en masse
- •Examples: one person can now do creative + analytics + multi-platform execution
- •AI-native skill becomes analogous to early ‘Google search’ fluency on resumes
- •Model-training experience can pay well and teach cutting-edge tool usage
- 33:40 – 35:41
From middlemen to direct: how Handshake discovered the AI data opportunity
Handshake initially sent talent to other ‘middleman’ platforms, then observed poor experiences for experts and increasing direct demand from labs. Garrett explains the insight: Handshake could treat experts better, serve labs directly, and build an experts-first platform that becomes core infrastructure for AI progress.
- •Middleman platforms sought to recruit Handshake’s PhDs/masters students
- •User feedback revealed transactional workflows, poor training, and payment friction
- •Frontier labs began approaching Handshake directly to cut out intermediaries
- •Strategic bet: build an experts-first platform to monetize knowledge at global scale
- 35:41 – 37:18
Hypergrowth execution: zero-to-$50M ARR and scaling to serve ‘unlimited demand’
Garrett recounts the fast build: exploration over the holidays, platform build in January, monetization starting a few months later, and rapid expansion to multiple frontier labs. The core operational challenge becomes scaling supply and delivery without sacrificing trust and data quality.
- •Rapid timeline from exploration → team build → commercialization
- •Working with seven frontier labs; growth driven by massive unmet demand
- •Primary constraint: operational scale while maintaining quality and trust
- •Saying no and sequencing matters: deliver for one customer before expanding
- 37:18 – 40:43
Handshake’s original network: a college-to-early-career social marketplace as the underlying asset
Garrett explains the core Handshake product—an ‘unconnected graph’ for young people who don’t yet have networks or job histories. This decade of trust, profiles, and engagement becomes the unique leverage point enabling expert sourcing and high conversion for the new AI data business.
- •Handshake is ‘LinkedIn for students’ with discovery, messaging, groups, and feeds
- •Designed for users without prior jobs or social graphs; focuses on interests and exploration
- •10-year network effects: millions of students, alumni, and advanced degree holders
- •This trust and profile data enables precise targeting and high-quality expert matching
- 40:43 – 48:38
Why access to an audience is the moat: outperforming recruiter-heavy, ad-driven competitors
They contrast Handshake’s distribution advantage with competitors who rely on recruiters and performance marketing to find experts. Garrett argues that as the market shifts to expert data, brand trust, community experience, and retention become structural advantages that determine unit economics and quality.
- •Competitors spend heavily on ads and recruiters to source scarce experts
- •Experts expect expert-grade onboarding/training—not ‘low-cost labeling’ treatment
- •Handshake’s low CAC + high conversion + high retention drives better LTV economics
- •Thesis: the only real moat in human data is access to a trusted audience
- 48:38 – 57:44
Incubating NewCo inside OldCo: separation, ownership, cadence, and ‘founder mode’ intensity
Garrett details how they made an internal startup work: separate teams, separate operating rhythms, explicit expectations, and compensation tied to the new business. He emphasizes clear DRIs, flat structure, metrics rigor, and a culture anchored in urgency—‘leave nothing to chance.’
- •Radical separation: dedicated engineering/design/ops/finance, separate onboarding and all-hands
- •Clear ownership: single-job focus, DRIs chosen by capability not hierarchy
- •Hiring for early-stage comfort with ambiguity; explicit ‘24/7’ pace expectations
- •Culture: celebrate impact, move fast, and preserve quality to sustain customer trust
- 57:44 – 1:00:19
The long game: AI-powered job matching and reinventing hiring workflows
Garrett connects the AI data business back to Handshake’s mission: transforming labor-market matching. He predicts AI will replace tedious steps like mass resume review and manual cover letters, enabling richer skill capture, simulations, and better employer-candidate fit at scale.
- •Vision: build the best job-matching marketplace on the internet
- •AI interviewer, skill extraction, and work simulations improve signal and fairness
- •Hiring managers won’t manually review hundreds of resumes in the near future
- •Human-data capabilities from the AI business feed improvements into the core marketplace
- 1:00:19 – 1:04:02
What limits model progress next: evolving data types, multimodal demand, and the role of synthetic data
Lenny asks whether the world will ‘run out of data’ and what bottlenecks remain. Garrett argues data needs will evolve toward tool-use, CAD files, scientific systems, audio and video—while synthetic data helps in some verifiable domains but won’t replace human expertise at the frontier.
- •Data won’t run out; it will shift toward new modalities and tool-use contexts
- •High demand areas include audio, multimodal, and specialized scientific workflows
- •Synthetic data can help in verifiable domains but won’t dominate frontier progress
- •The bottleneck is matching evolving data needs with high-quality human expertise
- 1:04:02 – 1:09:50
Lightning round + early hustle stories: books, Game of Thrones, SNOO, and the Princeton pool shower
The episode closes with quick personal questions and a memorable founder-hustle anecdote. Garrett shares influential books, a show he’s watching, a newborn parenting product, his motto, and how showering at campus pools helped them sell early Handshake deals.
- •Book recs: Zero to One, Shoe Dog, The Hard Thing About Hard Things
- •Personal: first-time Game of Thrones watch; newborn life + SNOO
- •Motto: ‘leave nothing to chance’
- •Origin story hustle: traveling campus-to-campus, sleeping in a car, Princeton pool shower incident
