Comment by A_D_E_P_T
3 years ago
There's a mistaken assumption here. The author seems to believe that these generators operate on the basis of random chance, hence a 200x200 pixel image in RGB has a possibility space of 10^300000 images.
...But an overwhelming majority of those images are random noise, and are effectively homogeneous and interchangeable. They're perfectly identical.
Like human artists, the image generators are (very strongly) biased to create "structured" images, which are derivative of existing artistic works or natural representations in the form of photographs. In principle, there's no a priori qualitative difference between the AIs and humans in this respect -- no human art is created de novo, but is always a continuation of, or a reaction to, existing forms.
Further, there's no reason to believe that (a) the size of the possibility space matters in a quantitative sense, and (b) that AIs will never be as capable of qualitative "originality" -- in conception, composition, or subject matter representation -- as a human artist would be.
I put my caveats on indistinguishable images. And three is the challange - to make a better estimation of an effective number of distinguishable images.
> (b) that AIs will never be as capable of qualitative "originality" -- in conception, composition, or subject matter representation -- as a human artist would be.
Sure, AI already creates original things. The point is that questions like "create a drawing that presents friendship" can be approached from many, many different, and what is crucial - subjective, ways.