I see a conflict between the argument that training LLMs on publicly visible comments (or books or articles) is stealing, and open weight LLM models. If intellectual property is interpreted like that, it will make free LLMs illegal to use, since the original creators of the training data have not licensed this use (even though this data is publicly readable on websites).

I would consider it the worst possible outcome if only the AI corporations would be able to profit from the global treasure of our accumulated knowledge. And I suspect that is what is going to happen because they can lobby for some kind of broad licensing deal and pay them off, but for open source it will not work. I believe that is how they will monopolize AI. Then they will truly have stolen it, because they have taken it away from everybody else.