I wouldn’t be so quick to count disruption of Hollywood out. Hollywood won’t disappear but movie productions will require many fewer people in the future. The moviemaking business will change, significantly, and possibly for the better, as costs are wrung out of the system.
I'd be curious to get your opinion on the following ideas:
1. Synthetic data much?
Apart from Shutterstock videos (with whom OAI has signed a partnership for DALLE-3), what do you think about the training data including synthetically generated videos? To me, human face close-ups look similar to UE5 Metahumans, and on the generated videos which include landscapes, nature and backgrounds - do you think NERFs or UE5 Nanite/Unity generated content might be at play?
2. Compute implications
From the tech report, I've noticed quite a few referenced papers on temporal-diffusion. Since they specify in the report that the entire video is composed in latent space - at once, as in they don't rely on diffussing only key-frames and then complete in betweens like other TSR (temporal super resolution) methods propose. My logical reasoning is that applying diffusing-steps over an entire stack of frames is compute hungry - do you think this is the case?
I explore 7 non-obvious cultural and social implications & questions about what this might mean (digging deeper than just disinformation and democratisation, which are what everyone says!)
These include: imagination > skills; but skills will still matter, they'll just be more accessible; collaboration will be humanity's superpower; purpose will be everything... to drop a few teasers.
Sora should spin out as an entirely different company. It should specialize in video search and also offer a YouTube competitor.
Unfortunately you can't expect OpenAI to think very strategically. Sam Altman is too busy playing Venture Capital game of thrones.
I wouldn’t be so quick to count disruption of Hollywood out. Hollywood won’t disappear but movie productions will require many fewer people in the future. The moviemaking business will change, significantly, and possibly for the better, as costs are wrung out of the system.
Interesting ideas Bojan!
I'd be curious to get your opinion on the following ideas:
1. Synthetic data much?
Apart from Shutterstock videos (with whom OAI has signed a partnership for DALLE-3), what do you think about the training data including synthetically generated videos? To me, human face close-ups look similar to UE5 Metahumans, and on the generated videos which include landscapes, nature and backgrounds - do you think NERFs or UE5 Nanite/Unity generated content might be at play?
2. Compute implications
From the tech report, I've noticed quite a few referenced papers on temporal-diffusion. Since they specify in the report that the entire video is composed in latent space - at once, as in they don't rely on diffussing only key-frames and then complete in betweens like other TSR (temporal super resolution) methods propose. My logical reasoning is that applying diffusing-steps over an entire stack of frames is compute hungry - do you think this is the case?
Thank you
Completely agree.
I explore 7 non-obvious cultural and social implications & questions about what this might mean (digging deeper than just disinformation and democratisation, which are what everyone says!)
These include: imagination > skills; but skills will still matter, they'll just be more accessible; collaboration will be humanity's superpower; purpose will be everything... to drop a few teasers.
Hope you enjoy :)
https://thefuturenormal.substack.com/p/open-ais-sora-text-to-video-7-non