Edit: yep, found it.
SmoothGrad: removing noise by adding noise, https://arxiv.org/abs/1706.03825
Web page with explanations and examples
https://tensorflow.github.io/saliency/
I couldn't find the HN thread, but there was no discussion as far as I remember.
The effect is same one that occurs when you get a group of people together to estimate the number of jelly beans in a jar. All the estimators are biased, but if that bias is drawn from a zero mean distribution, deviation of the average bias goes down as the number of estimators increases.
I can certainly observe what's being selected once the state machine is generated, but I have no clue how it was constructed to make the features. Do determine that, I have to watch the state of the machine as it "grows" to the final result.
Beyond that, the convolution/max pool repeated steps could be understood to be applying something akin to a multi-level wavelet decomposition, which is pretty well understood. It's how classical matched filtering, Haar cascading, and a wide variety of proceeding image classification methods operated at their first steps too.
CNNs/Deep learning really doesn't seem like a black box at all when examined in sequence. But to me at least, randomized ensemble methods (random forest, etc.) are actually a bit more mysterious to me in their performance out of the box, with little tuning.