No you are right. You can compare this to almost any AI that uses a similar strategy, to predict the future and maximize some goal. The cool thing about this is that it defines it's goal as something general purpose enough to create interesting behaviors in a lot of different situations.

But the hard part of AI is actually predicting the future in the first place. Exploring a search tree with billions and billions of possibilities. Or worse, trying to figure out how the world works in the first place from a limited set of observations, then doing that. Figuring out the goal of the AI was never really the hard part.

> The cool thing about this is that it defines it's goal as something general purpose enough to create interesting behaviors in a lot of different situations.

Yep, that is the first thing that stood out to me about both the NES AI and this new paper. Both are defined in ways that are so far removed from the tasks they've been put to that they seem to seek out their own goals relevant to the context.