I'm also going from the other post which said that this is all simply 90's era algorithms scaled up. But using that form of neural net stuff, wouldn't we expect minor mistakes like this from time to time? Neural net does strange unexplained thing suddenly is an ancient tale.

it doesn't even have to do the 'are you paying attention' thing (which shows so many levels of awareness it is weird (but I guess they are just saying it is copying the test idea back at us (which is parroting, not cognition but whatever))) because it is aware, it could just be an error.