
Explaining AI Data Labeling and Post-Training

Inside the expert network training every frontier AI model | Garrett Lord
Topic 2 of 6
Published Aug 24, 2025 · Clip 5:07–22:20 (17:13)
Garrett explains AI model training, distinguishing pre-training on existing internet data from post-training, which creates new high-quality data to improve models in areas like coding, math, and law. Post-training involves methods like RLHF, trajectories, and synthetic data, but focuses on expert-generated data to fix model weaknesses. Handshake targets advanced domains using its network of 500,000 PhDs and 3 million master's students for specialized tasks like breaking models in physics or creating educational rubrics.