Inside the expert network training every frontier AI model | Garrett Lord (Handshake CEO)

Garrett Lord describes Handshake AI as a new business built on Handshake's decade-old student and early-career network. Lenny frames the opportunity as a company using an existing distribution advantage to create new data for model improvement, while Lord argues that access to an engaged expert audience is the central moat in human data. He says Handshake reaches 18 million professionals, including 500,000 PhDs and 3 million master's students, and that this audience lets the company target expertise rather than recruit anonymous workers one by one. These figures and the business trajectory are claims made in the conversation, not independently verified evidence in the episode.

The interview's main technical distinction is between pre-training and post-training. Lord says pre-training absorbed enormous collections of human-created material, including books, videos, and other internet data, while recent gains increasingly come from post-training. In his account, post-training means collecting targeted, high-quality examples in capabilities such as coding, mathematics, law, and finance, then testing hypotheses about which data improves a model. The examples can include reinforcement-learning environments, tool-use trajectories, multimodal material, prompt-response pairs, and preference rankings from human feedback. He presents this as an evolving experimental process rather than a single labeling task, and does not claim that all model progress has stopped on the pre-training side.

Lord explains Handshake's role as supplying people who can expose failures that ordinary users may not find. A physics PhD can probe several subfields, identify an error in a model's reasoning, distinguish a wrong final answer from a wrong path to that answer, and test whether the model can use tools or follow a procedure. The value is therefore not only producing a correct answer. It is finding difficult failure cases and creating the evidence needed to improve a specific capability. Lord also says the market has shifted from generalist labor doing basic tasks toward specialists who can work in advanced science, mathematics, medicine, law, finance, and other domains. That claim describes his view of current demand; the discussion does not quantify how much of the wider labeling market has actually made that transition.