In my opinion, the real breakthrough described in this article is not bigger models to read the web, but models that can experiment on their own and learn from these experiments to generate new ideas.

If this happens, then we indeed enter a non-linear regime.