On training data, his framing is clear:
These models are not photocopiers.
They “watch” movies, images, text and learn patterns, the same way humans watch and learn.
The goal is understanding how the world works, then generating new sequences that follow those rules.