I'm looking forward to people finding ways to throw their sabots into the gears. Stickers on stop signs. AI only works because we give them easily interpretabl

I'm looking forward to people finding ways to throw their sabots into the gears. Stickers on stop signs. AI only works because we give them easily interpretable data - the neck (bottleneck or metaphorical neck) is the data. And the data is never complete, the entirety of deep learning only works because the objective function gains more local extrema with higher dimensions, so the probability of finding one of those increases as dimensionality increases. And someone correct me if I'm wrong but to my knowledge, deep learning models, by their nonlinearity are the opposite of local, you change your training data somewhere locally and your entire model changes everywhere. So incorporating edge-cases is a fiddly and very costly affair (Edit: and only possible to a degree).