Frontier AI labs have largely exhausted the gains from pre-training models on the public internet. The majority of performance improvements now come from post-training processes like RLHF and fine-tuning, which require high-quality, specialized data to enhance model capabilities in specific domains.
Handshake successfully identified that its decade-old network of students and alumni was a uniquely valuable asset for the burgeoning AI data market. They built a new, hyper-growth business on top of their existing platform, demonstrating how established companies can pivot to capitalize on technological shifts.
As AI models become more sophisticated, the need for data providers has evolved from low-cost generalists to highly-skilled subject matter experts. These experts provide the nuanced data required to improve models in complex fields, creating new, high-paying gig work opportunities ($100-$200/hr for PhDs).
The speaker argues that the only sustainable moat in the human-generated AI data market is access to a large, engaged, and proprietary audience. While competitors rely on expensive acquisition channels, Handshake leverages its existing network at near-zero cost, creating a powerful and defensible business model.
To ensure focus and agility, Handshake launched its AI business as a separate, incubated unit with dedicated leadership. The CEO devoted over 80% of his time to the new venture, treating it like a new startup to navigate the zero-to-one phase effectively within a larger organization.
Keep pulling the thread on Garrett Lord.