Handshake's PhD Network Beats Crowdwork
Joe Kim, CEO of Office Hours, on the end of crowdwork
This points to a structural shift in AI training, where the scarce input is no longer cheap labor, but verified expertise. Handshake already had a dense supply of graduate students, postdocs, and PhDs inside its university network, so when labs needed people who could solve math proofs, judge legal reasoning, or evaluate scientific answers, it could convert a recruiting graph into an expert work marketplace much faster than crowdwork platforms could.
-
Handshake’s core business was a campus recruiting network that scaled to 18M students and alumni, 1M companies, and about 1,400 colleges. That gave it direct distribution into the exact population frontier labs wanted for post training work, especially academically credentialed workers who could be screened through school affiliation and field of study.
-
The product shift is concrete. Instead of paying thousands of generalist annotators to click labels, labs now pay smaller pools of specialists to write step by step solutions, rank competing model answers, and run domain specific evaluations in areas like math, physics, and computer science. Handshake AI monetized this by paying experts roughly $100 to $125 per hour for specialized tasks.
-
The comparison set shows why access matters more than workflow software alone. Mercor built AI screening and matching for experts across medicine, law, and research, while traditional labeling firms like Invisible and Scale grew from operational infrastructure. Handshake’s advantage was starting with a pre assembled supply of credentialed academics, which helped Handshake AI reach $80M annualized revenue within 8 months.
Going forward, the winners in model training will look less like generic labor marketplaces and more like owners of hard to replicate expert supply. As models push deeper into professional work, networks with trusted access to PhDs, clinicians, lawyers, and engineers will capture the highest value layers of evaluation, fine tuning, and benchmarking.