It's impressive, but I think it's still in the same category as even the best LLMs: the demos look good and they can be quite useful, but you can never quite trust them. You really can't just have an LLM write a whole report for you - who knows what facts it'll make up, what it'll miss? You really can't use this to generate video for work, who knows where the little artifacts are (it's easier to tell with video).

The future of these high-fidelity (but not perfect) generative AI systems is in realizing we're going to need "humans in the loop". This means designing to output human-manipulable data - perhaps models/skeletons/textures instead of whole output. Pixels are hard to manipulate directly!

As for entertainment, already we see people sick of CGI - will people really want to pay for AI-generated video?