The video compression is very cool. And the small tricks like binning the mouse movements.

Wonder how much data is generalizable across different UIs? ie how good will the model be at using Figma if it’s never seen it before but has seen a lot of Photoshop