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
Handshake’s breakout AI data business and the “leave nothing to chance” mindset
Garrett opens with the intensity of operating in a market with seemingly unlimited demand for high-quality human data. Lenny tees up Handshake’s new AI data business built on top of its decade-old career network and hints at the company’s structural moat: access to experts at scale.
- •A rare market moment: demand outstrips supply if you can execute
- •Handshake’s advantage: an existing trusted network (no traditional acquisition costs)
- •Human data as a core lever for continuing model improvements
- •Operating principle: obsess over quality, urgency, and customer closeness
What data labeling really is: pre-training vs. post-training (and why post-training matters now)
Garrett explains the two major phases of building frontier models: pre-training on broad internet-scale corpora and post-training to target capability gaps. As pre-training gains have started to plateau, labs have shifted much more attention (and budget) toward post-training data creation and refinement.
- •Pre-training: ingesting the corpus of recorded human knowledge
- •Why gains from pre-training have asymptoted as web-scale data saturates
- •Post-training: targeted data to improve specific capabilities (coding, math, law, etc.)
- •Research teams run hypothesis-driven experiments and scale what works
Post-training data types: SFT, preference ranking, trajectories, multimodal, and rubrics
Lenny presses for a clearer taxonomy of post-training. Garrett outlines the variety of data that labs now collect, from supervised prompt–response pairs to preference ranking and full “trajectory” recordings of tool use, plus multimodal and rubric-driven evaluation approaches.
- •SFT (supervised fine-tuning): prompt–response pairs and reasoning steps
- •Preference ranking (RLHF-style): comparing outputs and selecting better ones
- •Trajectory data: end-to-end task execution across tools and steps
- •Rubrics and model-as-judge approaches for non-verifiable domains
- •Multimodal needs: audio, video, and domain-specific formats
Why experts are the new bottleneck: breaking models and generating novel knowledge
Handshake’s core value is mobilizing deep expertise—especially PhDs and advanced students—to find where models fail and supply high-quality corrections. Garrett argues the market is shifting from generalist labeling (e.g., bounding boxes) to expert work because models have absorbed routine tasks.
- •Handshake’s supply: millions of students and professionals; hundreds of thousands of PhDs
- •Experts can reliably expose failures average users can’t find
- •Work includes producing ground truth answers and step-by-step reasoning
- •Shift from generalist, low-cost labor to expert-driven capability improvements
What the work looks like day-to-day: GPQA-style tasks, training, QA, and JSON outputs
Garrett describes a concrete workflow: experts are onboarded, trained on tools, then create data that diagnoses failures and provides correct solutions. Handshake backs this with internal research and post-training teams to evaluate quality and estimate model gain, delivering outputs in standardized formats.
- •GPQA-style example: break the model, provide ground truth and reasoning steps
- •Model non-determinism: correctness must be robust across multiple runs
- •Expert experience design: community, training cohorts, and expectations tailored to PhDs
- •Handshake’s internal infrastructure: research/post-training team and GPU resources
- •Deliverables are packaged as structured data (e.g., JSON), including multimodal work
Quality, volume, and speed: what frontier labs actually buy (and how they measure trust)
Garrett frames labs’ purchasing criteria as a three-part equation: quality first, then volume, then speed of iteration. He explains why fast turnaround enables labs to scale winning hypotheses and drop losing experiments, and why Handshake invests in assessment and reliability to maintain trust.
- •Frontier labs’ priority stack: quality → volume → speed
- •Scaling advanced-domain data is hard; Handshake targets top institutions for supply
- •Rapid iteration enables labs to expand the pipelines that show measurable gains
- •Handshake builds technology and processes to evaluate each unit of data
AI and jobs: why Garrett thinks entry-level roles won’t disappear (and may get stronger)
Lenny raises the fear that students training models could be automating their own future jobs. Garrett counters that employers are seeing productivity amplification: “AI-native” young workers can do more end-to-end work, earlier, and gain an advantage similar to early ‘Google-native’ resumes.
- •Employer signal: augmentation is dominating, not mass elimination
- •Examples: one person can now create assets, run analytics, and ship faster
- •AI-native graduates gain leverage; tools become an ‘Iron Man suit’
- •Training models can also teach fellows how to use AI to advance their research
Why Handshake entered the labeling market: cutting out middlemen and fixing a broken expert experience
Garrett recounts how intermediaries first asked to recruit Handshake’s PhDs and masters students, but the user experience on those platforms was transactional and leaky. Then frontier labs began approaching Handshake directly, revealing a chance to serve both sides better by building an experts-first platform.
- •Middlemen demand surfaced first; user feedback highlighted poor training and payments
- •Frontier labs started reaching out directly to bypass intermediaries
- •Thesis: the moat in human data is access to a trusted audience
- •Launch timeline: exploration over holidays, team built in January, rapid monetization
- •Now working with most major frontier labs
Handshake’s competitive advantage: trusted audience, targeting, and lower CAC than recruiters + ads
Garrett explains why many labeling companies rely on expensive recruiting and performance marketing to find scarce experts—and why that breaks unit economics. Handshake can precisely target qualified experts from an existing, trusted platform, improving conversion, retention, and overall LTV.
- •Competitors: heavy recruiter headcount + high ad spend to source experts
- •Expert mismatch: professionals don’t want ‘low-cost labor’ workflows
- •Handshake: brand trust, rich academic signals, and built-in distribution via universities
- •Better retention increases LTV; low acquisition costs strengthen margins
- •Access to audience is positioned as the only durable moat in human data
Incubating NewCo inside a 10-year company: separation, ownership, and founder-mode execution
Garrett details the operating system that allowed a new business to move at startup speed inside a mature organization. The key was radical separation—teams, onboarding, cadence, incentives—paired with high-ownership hiring and clear DRIs to avoid the drag of legacy prioritization.
- •Separate engineering/design/ops/finance functions; minimal shared responsibilities
- •Physical and cultural separation: different office area, separate all-hands and onboarding
- •80%+ CEO focus on NewCo; core execs continued running the legacy business
- •Hiring for early-stage comfort with ambiguity and intense pace
- •Clear DRIs, flat structure, and explicit expectations about chaos and workload
Scaling under ‘unlimited demand’: growing headcount, protecting quality, and deepening customer trust
With demand exceeding capacity, Handshake’s main constraint becomes execution: hiring fast without compromising quality and reliability. Garrett emphasizes saying “no” early, proving quality with initial customers, and scaling only once trust is established in a research-sensitive environment.
- •Rapid scaling: from a handful of people to ~75+ and accelerating hiring
- •Bottleneck: meeting demand while keeping data quality consistently high
- •Strategy: deliver for one top customer first, then expand carefully
- •Cultural mantra: “leave nothing to chance” to avoid future regret
How this ties back to Handshake’s core mission: reinventing job matching with AI
Garrett connects the AI data business to Handshake’s long-term ambition: dramatically better labor-market matching. He argues AI will eliminate tedious hiring mechanics (resume screening, cover letters) and enable richer skill capture, work simulations, and more equitable access to opportunity.
- •Vision: build the best job-matching marketplace on the internet
- •AI will reshape applications and screening (resumes and cover letters won’t persist)
- •Opportunities: AI interviewing, skills extraction, and work-simulation assessments
- •Personal motivation: democratizing access regardless of school or network
- •Human data learnings feed back into the core product to improve outcomes
Will we run out of data? The next bottlenecks and evolving formats (tools, audio, multimodal, synthetic)
Lenny asks whether model progress will plateau due to data scarcity. Garrett argues data will keep evolving into new formats and domains—tool use in science, CAD files, audio, and multimodal workflows—while synthetic data helps in some verifiable areas but won’t replace human expertise broadly.
- •Data won’t ‘run out’; the valuable types will shift toward new workflows and tools
- •Future data: scientific tool use, CAD files, domain-specific operating systems
- •Growing demand for audio and multimodal data
- •Synthetic data may help in verifiable domains but won’t dominate frontier progress
Lightning round: books, shows, products, and the Princeton pool shower hustle story
In quick-fire format, Garrett shares formative books, what he’s watching, and a favorite new-parent product. He closes with an early Handshake founder hustle story: sleeping in a car, showering at university pools, and nearly getting arrested at Princeton—then using it as a memorable conversation starter.
- •Book picks: Zero to One, Shoe Dog, The Hard Thing About Hard Things
- •TV: starting Game of Thrones for the first time
- •Product: SNOO for newborn sleep support
- •Motto: “leave nothing to chance”
- •Origin hustle: sleeping in the car, campus pool showers, Princeton security incident
Where to find Garrett and how listeners can help: hiring engineers and scaling the team
Garrett shares how to contact him and underscores that hiring is the primary constraint to meeting demand. He calls out engineering as the top need and invites people interested in either the core Handshake product or the new AI data business to reach out.
- •Contact: Handshake message, X/Twitter, and email
- •Primary need: hiring (especially engineers) to keep up with demand
- •Global footprint: New York, San Francisco, London, Berlin
- •Roles span consumer product, employer product, and the AI data business
