[lanternfish]

The problem is that it's pretty easy to fine tune an adversarial model to fool classifiers. In some sense, building that classifier is a very productive way to improve the model, so it's something the generative side is already doing. That makes it hard to win the arms race from the classification side.

[randomdata]

But also the solution. With enough classify -> beat the classifier iteration, the works will be indistinguishable from human production, at which point there is no longer a need for classification.

[lanternfish]

This assumes that the limit is full generative parity, and not just a our SOTA classification techniques maxing out.