> the generator wasn't trained explicitly to fool a detector but to fool humans.
Yes, but “next week”’s one will be trained to fool humans AND this week’s Grover - which is still likely to be easy.
Next week’s Grover will Be harder to get right against this model. But will likely be built.
The faker model of two weeks from now will be harder to build (against the updated week Grover), and so on.
There is no fallacy here. An attacker usually has easier time than a defender, because the defender must respond to every attack whereas an attacker can vary and pick every detail of their attack.
Yes, but “next week”’s one will be trained to fool humans AND this week’s Grover - which is still likely to be easy.
Next week’s Grover will Be harder to get right against this model. But will likely be built.
The faker model of two weeks from now will be harder to build (against the updated week Grover), and so on.
There is no fallacy here. An attacker usually has easier time than a defender, because the defender must respond to every attack whereas an attacker can vary and pick every detail of their attack.