I am a little bit bored with these GAN style attacks. They show that the networks haven't yet generalized well enough. They don't show that the technology, or e

I am a little bit bored with these GAN style attacks. They show that the networks haven't yet generalized well enough. They don't show that the technology, or even the approach, is broken. Personally, I think the main problem is some combination of a little bit of over claiming in research, and a metric ton of over hyping and generalizing in both media and business circles. Robust multi model ensembles with strong generalization ability will show up within the foreseeable future, and will be no more susceptible to optical illusions than human beings.