All they seem to be creating is new APIs to interact with the models they created years ago. It's pretty clear that new models are not viable to train and that the supposed "scaling laws" are bovine feculence.