Data flywheel
Model quality explodes when content generation gets good enough to produce its own training data. Language reasoning models took off once we could generate massive amounts of good synthetic reasoning data in 2024. Video is next, with Seedance 2 output unlocking synthetic video at scale. Robotics and egocentric data are right behind. Get generation good enough, watch the training dataset explode, quality follows.
Connections
Parent: Data
Tangents
- Synthetic data: generation good enough becomes the dataset
