Mecka AI's Josh Gao on building robot training data at scale

Josh Gao, founder of Mecka AI, runs a robot data factory that manufactures training data for AI robotics—a crucial input that unlike large language models, cannot draw from free internet data. In a conversation with Sequoia Capital, Gao describes the infrastructure required: custom sensors like force-sensing gloves and egocentric cameras worn by humans, 100-gigabit upload lines to handle petabytes of footage, and quality control gates to filter out errors and fraud.

Mecka collects egocentric video of human work (roughly one million hours), combines it with robot experience and small amounts of direct robot control data, and uses simulation to fill gaps. Gao argues robotics remains orders of magnitude behind Tesla's autonomous driving progress—which focused on two degrees of freedom over a decade, while a robot hand has dozens. He predicts no single breakthrough moment like GPT-3, but rather a steady unlocking of commercial tasks: first constrained industrial work, eventually extending to autonomous systems like asteroid mining robots.