I think all models should be at least open weights when trained on public data or better the training data itself has to be open. Right now it looks like those companies can just put IP on public knowledge which is the inverse of what the internet was supposed to do.