observation · 18 Mar 2024 · 38:39

A substantial portion of the data used for training models like Sora is self-supervised and not human-labeled.

it's a lot of data that's not human label. It's self-supervised in that way.

Watch at 38:39