/ic/ - Artwork/Critique

>>6830031
even for the base model, enough is known about the process that it can be redone from scratch given we have the compute and the data.
and i pretty much told you what i think about the copyright issue.

>No I know you're delusional about foregoing the legality of copyright for some reason, but it's the crux of the issue and why there's a huge problem with the whole thing. Hand waving it away cause you don't care doesn't solve anything.
i'm not. i'm just telling you that i care more about the ethics than the law. and i do think the copyright claim is shaky. your earlier argument was the only one that convinced me there was even an argument to begin with.

>Who changes the weights in the model? the model is incapable of doing anything consciously by itself. It can't passively "learn" or actively seek out and learn like a human.
it is "taught" by what the humans give it.
a convolutional NN (which is one part of SD) does literally that. you give it an image of a car and a mask telling it "this is the car", "these pixels belong to the car", "these don't".
do you, as a human, think you could know what a car is and that it was in fact called a car without ever being told or shown?
then, later on, everything it can do after training, it can do so because of the changes made to it during training. how is it not learning? how is it not literally applying the information it gathered from before and using it?
if you prompt "blue eyes" it will steer the noise towards showing blue eyes, based on the blue eyes it has seen.
do you at least understand why i make so many countless references to human learning?

1/2