Comment by redgreenblueyel

I have to disagree. If there exists an underlying abstract, general pattern of coherence, you can optimize against it without the need of any current or next real world state. The next real world state is then resolved during inference by a system that produces high coherence. But the training can be independent. I am not claiming that there exists such learnable signal - i have no clue either, but wanted to note that the article somewhat suggests this in its plots.