Spatial intelligence
The ability to understand three dimensional space the way people do, which current text and video models largely lack.
Spatial intelligence is the term Fei-Fei Li uses for the thing AI is still bad at. You can look at one photo of a kitchen and know roughly how far the fridge is from the door, what is behind the counter, and how to walk through it without bumping into anything. Models trained on text or video mostly cannot, because they process the world as a line of words or a stack of flat frames.
It matters most for anything that has to act physically: robots, self driving, warehouse machinery, but also for design and film work where you need a scene to stay consistent as the camera moves. The research bet is that you get there by building the sense of space into the architecture rather than hoping it emerges from enough video.