Contrary to text and the big piles of "liberated" data hanging around for anyone looking hard enough to grab, the training data for video seems to be harder to access for opensource / research / individuals. Google has Youtube, OpenAI can pay whatever fee any proprietary data bank requires. There's a moat right there that I can't see how to overcome.
Weird to say I guess, but meta might release an open source model too. And they do have plenty of data to feed their models. Arguably more data than openAI should have as they don't really own any social media.
Thing is, anyway, as soon as one model is open there will be copies of it, fine-tune implementations. People don't care that much about ownership of data I would say if they actually have access to the models that are produced by gathering this data.
Ultimately, to me, an open source model for this tool makes a lot of sense. They use publicly available data and the models become publicly available.
I for one am quite excited for this tooling to become better and better so I can make the adaptation of a book I love into a movie I imagine it can be. At least I can have a lot of fun trying.
