How 1 Trillion Video Tokens Are Changing AI
A new paper by the GMAE team dives into the power of pre-training video models on a massive 1 trillion video tokens. The takeaway? Bigger really is better. They uncovered scaling laws that hold across many design choices, showing consistent improvement as models grow. Interestingly, autoregressive training performs just as well as diffusion and flow-based methods—different roads, same destination. This research could reshape how we build AI for video understanding.