SevenSpheres are the Swiss Army Knife I use for getting to the essence of things. For this particular one, it was triggered by the acrosstic for pain, which came from a post on X this morning. It shows how learning is a sort of a balancing act between pain and pleasure, which are both inputs, love and hate which can be outputs or internal states, and good and bad, which are internal states.
It does confirm that a singular reward function is not enough for an artificial intelligence to learn properly. But good and bad cannot be added as an afterthought outside the model itself. Because like children, computers will essentially learn from the example of their "parents", and not the teachings they receive.
Of course nothing is as simple as the above diagram. Love and hate are not opposites, but more like a spectrum where one is the extreme, and corresponds to the total absence of the other. Same for good and bad, and pain and pleasure. Nowadays it feels like the world in general and social media in particular are trying to pull us into the extremes, but I think we are better off trying to remain in the center...
