It even seems to me that plateau-ing is an inevitability. Unless there is some way to reach AGI which you could deduce from a combination of ideas which already exist in the training dataset, but that seems rather unlikely.