Also, "fan ports" is doing a lot of the heavy lifting. They're assets extracted from commercial games, uploaded with whatever license the person who extracted and packaged them chooses.

From their argument, if they truly think copyright should be abolished, an LLM capable of remixing that work into whatever you want would be a good thing, right? Ultimately this would empower the end-users of the models.

There's an argument to be made that these models trained on the world's knowledge shouldn't be locked behind closed models controlled only by big corps, but in this case it's good that anyone can download data and train such a model, right?

I can't speak for roblox intentions, and I am well aware of their history. But so far they seem to have published at least some of their model work to HuggingFace: https://huggingface.co/Roblox/

What nuance am I missing?