[deadbeeves]

Tool improvements only apply a constant factor to the effectiveness of learning. Creating a generative model applies an unbounded factor to the effectiveness of learning because, as I said, the only limit is how much computing resources are available to humanity. If a single person was able to copy themselves at practically no cost and the copy retained all the knowledge of the original then the two situations would be equivalent, but that's impossible. Having n people with the same skill multiplies the cost of learning by n. Having n instances of an AI with the same skill multiplies the cost of learning by 1.

[MichaelZuo]

Right, but the 'unbounded factor' is irrelevant because the output will quickly trend into random noise.

And only the most interesting top few million art pieces will actually attract the attention of any concrete individual.

For a current example, there's already billions of man-hours worth of AI spam writing, indexed by Google, that is likely not actually read by even a single person on Earth.

[deadbeeves]

Whether it's irrelevant is a matter of opinion. The fact remains that a machine being able to copy the artistic style of a human makes it so that anyone can produce output in the style of that human by just feeding the machine electricity. That inherently devalues the style the artist has painstakingly developed. If someone wants a piece of art in that artist's style they don't have to go to that artist, they just need to request the machine for what they want. Is the machine's output of low quality? Maybe. Will there be people for whom that low quality still makes them want to seek out the human? No doubt. It doesn't change the fact that the style is still devalued, nor that there exist artists who would want to prevent that.