Also, the main showcase of the 'zero' models was that they learnt with zero training data: the only input was interacting with the rules of the game (as opposed to learning to mimic human games), which seems to be the kind of approach the article is asking for.