十年校园人脉变成AI数据工厂：Handshake四个月做到5000万美元年经常性收入

Lord explains that training a model has two phases. Pre-training means absorbing the entire corpus of written human knowledge, including books, videos and everything else on the internet. Around 18 to 24 months ago, he says, the gains from that approach began to flatten because the labs had essentially consumed all of it.

Since then, most progress has come from post-training, which improves a model in specific capability areas such as coding, mathematics, law or finance. Research teams inside the labs work like scientists. They form a hypothesis, collect a small batch of data to test it, and scale up collection only if the hypothesis holds.

The data comes in several forms: reinforcement learning environments, trajectories, audio and multimodal material, prompt-and-response pairs for supervised fine-tuning, and preference rankings for reinforcement learning from human feedback. Lord cautions that the split is not clean, because labs have their own views and the field changes quickly. Reasoning models in particular need step-by-step instructions, since frontier systems still struggle in very advanced domains.

Handshake's pitch is an engaged audience: 18 million professionals, 500,000 PhDs and 3 million master's students. A PhD, Lord notes, means someone has defended a thesis showing they extended human knowledge in a domain, so targeting this group reaches knowledge that never made it onto the internet. He says labeling used to be generalist work done by lower-cost international labor, but models are now good enough that generalists are no longer needed.

Labs concentrate on the most economically valuable areas, which Lord describes as advanced STEM and its derivatives such as accounting, law, medicine and finance. Handshake also uses millions of bachelor's students for tasks like audio work and tuning a model's voice and tone. The common thread is that an average person can hardly make a frontier model give a wrong answer, while a physics PhD can show exactly where its reasoning or ground truth fails.

Lord points to the public GPQA paper as a template. The expert breaks the model, supplies the correct answer and writes out the reasoning steps, because a model can reach the right answer while getting steps six through ten wrong. He also describes a trajectory as a full record of an expert's screen, mouse and voiceover while they work through a problem, usually stored as JSON. Rubrics let a model act as a judge in domains such as education or medical imaging where no single correct answer exists.

As an example, he cites a PhD in education who spent two decades as a teacher and now probes frontier models on how to teach well. In his telling, the people doing this work are treated as experts, with an instructional design team and an assessments team helping them learn the tools before they produce data.

Lord says model builders care first about quality, since bad data is very hard to undo, like teaching a student the wrong material. Second is volume: thousands of high-quality items in chemistry, math and physics. He says Handshake gets there by reaching top-GPA students at schools such as Stanford, Berkeley and MIT.

The third demand is speed. Researchers run several bets at once, and when one shows a gain they want to scale that pipeline within days and drop the others. To keep up, Handshake has built technology to assess each unit of data and runs its own post-training team on rented GPUs, so it can sit with researchers and show how the data might improve their model.

Asked whether the human role will fade, Lord jokes that it is the MBAs who think it will all go away. He expects data types to evolve, with demand growing far beyond academic domains into the step-by-step work of knowledge workers, but he says humans will be needed for the next decade or so until full ASI. His rule of thumb is that when a model fails at an advanced task, a top expert is probably already working with a frontier lab to fix it.

Lenny Rachitsky raises the tension that students training models may find entry-level jobs scarcer. Lord says he sits in the camp of GDP growth over universal basic income, and notes that Handshake sees the hiring side directly because 100 percent of the Fortune 500 use it. Employers, he reports, are not saying young people will go jobless.

His example is social media marketing, which once took separate people for design, video and analytics and can now be done by one young person with AI tools. He also cites an intern who had a first pull request up the afternoon they started. Lord concedes that jobs will evolve and some people will be displaced and need to reskill, and he sees Handshake as having a role in that transition.

He compares AI fluency with the old habit of listing Google Search as a resume skill, arguing that growing up with the tools is a big advantage. For the experts themselves, he says PhDs can earn $100 to $200 an hour in their field, versus about $25 as a teaching assistant. They also bring insights back to their classrooms and learn to use AI in their own research.

Asked about Scale AI's acquisition and what it means for the market, Lord says labs probably do not want a peer invested in their latest research, and that Handshake has seen a surge in demand. His thesis is that the only moat in human data is access to an audience. Competitors, he says, run TikTok, Instagram and Google ads and staff large recruiting teams to find, for example, 200 physics PhDs.

Handshake instead spent a decade building trust with 18 million people who have active profiles and academic histories. That lets it target very precisely and reach high-quality volume faster than anyone else. The business grew out of inbound interest: middleman companies asked to recruit Handshake's PhDs, users complained that those platforms were transactional and had heavy drop-off, and then the frontier labs began reaching out directly.

Lord says he started flying to meet lab leaders over Christmas and New Year's, assembled a team from the human data world and began building the platform in January. Revenue started about five months before the recording. He says they now work with seven frontier labs, and Lenny notes the pace is zero to $50 million in four months with a path past $100 million in the first year.

The original business, a student-focused network about 10 years old, brings in roughly $200 million a year, and Lord describes it as an "unconnected graph" that starts from interests rather than from your existing job and connections. He argues it is hard to incubate something new inside a mature company, which makes running a second zero-to-one especially fun.

Lord describes a competitor with about 200 recruiters sending individual LinkedIn messages and spending tens of millions of dollars a month on performance ads. He argues that experts do not want to be treated like low-cost workers drawing bounding boxes, and that other platforms drop highly skilled people into projects with no training.

Handshake onboards experts into a community with groups and a feed, then into a cohort of peers with similar backgrounds, and teaches them to use the tools. Some projects are pre-built by Handshake itself, a bit like a movie production: it pays for a unit of data, post-trains on it to check quality, and sells that package to multiple labs. People who do well can then move onto customer projects that want only the best.

On economics, Lord says customer acquisition cost is effectively zero because Handshake partners with 1,600 universities and powers 92 percent of the top 500 schools. Lifetime value depends on retention and the number of projects someone joins. Because universities would not tolerate poor treatment of their students and alumni, the company has strong incentives to treat experts well, which drives high conversion and repeat engagement.

Lord says Handshake began by researching the space in December, hiring expert firms such as AlphaSights and GLG for calls with researchers, which a $200 million business could afford. The AI team grew from four or five people to more than 75, with 12 starting in a single Monday. He says the constraint is capacity: demand is effectively unlimited, so the main risks are picking the right long-term strategy and not growing so fast that trust with labs erodes.

His earlier experience helped. Building Handshake meant signing up universities, then selling to employers like Goldman Sachs, then building a social network, so marketplaces felt like a series of zero-to-ones. For the new unit, people reported directly to him, hires carried no duties in the old business, and one directly responsible individual was named per area. The team sat in a separate part of the office, worked a far more demanding schedule, and had compensation tied to hurdles in the new business.

Lord says the approach was to make everything separate: engineering, design, finance, operations, all-hands meetings, onboarding and recruiting. He spent more than 80 percent of his time on it, hired leaders with start-up experience, and was upfront that the work would be chaotic. Some of the core business's best senior engineers were asked to move over, and some declined because of the weekend and late-night load.

The team's motto, he says, was to leave nothing to chance: take the flight to see a customer, make the late-night push, check the data six times. He recalls drawing the remaining days of the year on a whiteboard and telling the team that this kind of moment would not come again. Rachitsky notes that the motto fits a business where trust in data is everything.

Lord says the larger opportunity is building the best job-matching marketplace on the internet, and that searching for and applying to jobs will be reinvented by AI. He mentions an AI interviewer that collects skills and asks about experience, plus work simulations that help employers find strong candidates. Reviewing 200 resumes by hand, he says, will not survive the next five years.

He ties this to his own story: he went to community college, attended a little-known school in Michigan's Upper Peninsula and interned at Palantir. He started Handshake so that anyone could find a good opportunity regardless of connections or school. Work from the human data business is already being folded into the core product, which he expects to improve outcomes for students and employers.

On whether data will run out, Lord says the kind of data needed will keep evolving: CAD files, scientific tool-use data for lab automation and drug discovery, video and especially audio. He thinks synthetic data has a role in verifiable domains but will not dominate, and that there are billions of dollars of value in following the frontier over the next decade.

In the lightning round he names Zero to One, Shoe Dog and The Hard Thing About Hard Things as favorite books. He also recounts showering in a university pool during early sales trips and being questioned by Princeton campus security, which he says gave a meeting with the career center a memorable start. His closing ask is for engineering hires across offices in New York, San Francisco, London and Berlin.