Reading the paper I'm frustrated they aren't releasing the full model. Their reasons seem really weak to me. They think it can be used to generate offensive pictures or deep fakes and it encodes hegemonic language (e.g. "girl toys" gets more pink and fluffy toys than "boy toys").
These arguments don't make sense for a few reasons - only academics and people capable of reproducing their results should have access to convincing deepfakes? Why is that "safer"? To me it seems less safe, if everyone had this power images would quickly become untrusted, if only a few have this power they could exploit trust in images longer.
Second, this argument would hold for things like Photoshop. It makes creating fake images easier. Should Photoshop be banned or weakened because people might create offensive pictures with it?
Third, the moment some company implements a good version of this and integrates it with their tools they will make it widely available to users. What use is there in delaying this capability for a couple years so some other company can profit from it?