Gary Marcus

as much as I generally agree w/you, if you look closely you can see what might be a person reflection where it says person (but it’s vague) & TV isn’t a horrible fail… there’s enough DLs to critique with really bad errors- maybe go easy on the “silly, but plausible"

with "superhuman" abilities come superhuman mistakes?

It is possible that this is not a mistake. Suppose there is a sculpture to the right of the image frame that is being reflected. This could give us this image. [A bit implausible, I admit, because the sculpture would need to be upside down.]

That’s the kind of reasoning deep learning can’t do, but AI needs to do

This is a very good example showing the possible errors in a ML system based on function approximation only. The common sense logic based AI needs to be integrated with such ML systems in high risk scenarios to make the precision near perfect.