Sora 1 was also lauded as being incredibly good at physics based on the early cherry-picked examples. The phrase "world simulator" was thrown around a lot. That didn't last long once people finally got their hands on it though.

Kind of wondersome if they will start to combine LLM generation with actual world models/GPU engines. Imagine that your model generates the wireframes, the Engine generates the physics and then another model fills in the actual visuals, and gaps... So you have realistic physics and gaps are filled in... Will also help with image retention more, if objects moved behind each other.

It was so much more hyped than that. They made it sound like Hollywood was in big trouble. It is going to have the same problems as Midjourney. You just don't have that much control of the scene. The process is to make thousands of random variations and cherry pick the good stuff because you can't do anything else.

The space dog and ice skater demo make it seem still very close to Sora 1