(3) is fascinating and begs the question of how we discover these representations. I imagine - like science - it's a bit of theory, a bit of empiricism. Constan

(3) is fascinating and begs the question of how we discover these representations. I imagine - like science - it's a bit of theory, a bit of empiricism. Constantly tinkering with our theories and our data to evaluate if new models improve and reliably predict performance. (7) is also interesting. Ambiguous success criteria is one problem. I'd also include delayed feedback and multivariate feedback loops. The longer the feedback loop, the more likely confounds can enter into the training regimen, rendering it harder to discern whether the new training ("treatment") actually has a significant effect. This is a problem in all fields: when we are trying to measure long-term results, it gets very difficult to decide whether any proposed intervention (new diets, new exercise regimens, new marketing campaigns) have any effect.