Zero of their flagship models have been released AFAIK, e.g. the pre-DALLE model, CLIP, has not had weights for its ViT-H/16 model released from a paper 19 months ago.
This was before DALL-E and way before DALL-E 2.
Not that I think they need to. It’s just weird people defend them as being open just because they release crippled versions of their model.
They've said that they intend to release the model soon-ish for anyone to run on their own machine, with no censorship*. I don't think the other major players, like OpenAI, have committed to doing the same.
* They're actually working on a filter right now, but IIRC it's an optional one, for when you don't want to accidentally generate NSFW output.
I'm so sorry I thought the discussion was about OpenAI, I'm glad you clarified it was about StableDiffusion. Hope I didn't muddy the waters too much for others.
You may be interested to know that the code for SD has already been released on Github, and they've given the weights to a bunch of researchers in preparation for full release. I've also heard that one of the researchers leaked the weights earlier today, and 4chan has been using them for, uh, stuff.
You're right that they could always change their minds and that would suck, but so far they seem to be being up front.