I find video models very hard to steer. They can output amazing "shiny" things quickly, but if you want something very specific, you may need to spend days on it. It is not like sitting down with an artist and seeing how they make the exact changes you need in real time while also ensuring artistic cohesion.

AI is obviously very useful, but perhaps not in the way people think. Two aspects I would like to see improved are instruction following and providing relevant feedback.