Ok,
So I thought to show myself random configurations of circle/rectangle (for now) on a flexible canvas.
Since all those things/colors/shapes/size/location along with canvas aspect ratio can work together to impress me or not impress me.

The point is I want to have enough data points to train a neural net so it learns what I think looks good so that I can then iterate over millions and millions of random configurations and have the neural net vote/rate for me (in place of me).

And then if I select the top scoring (highest ratings) it should be something decent.

Now I don't have much experience in Neural nets but ChatGPT was able to give me code that I can work with, whenever I want to extend functionality I just ask it how do I do this and boom sample code given (it's really great).

Anyways some of these are really impressive to me way better than what I can come up with on my own.

Check them out here:
First Go at it
Second Go at it

I am currently trying something new apparently called continuous learning where I train the network then grab 10 or (11 because of a counting bug) samples from 11 different ranges of scoring then I rate them what I think they're worth and then refeed that into the network let it train a little then spit out another 11 in different ranges of scoring and repeat again and again so that I continually shape the rating system over more and more sample data that's valuable for training.

Personally I like the 2nd Go better than the 1st.

Once I am bored of this combination, or find the piece I absolutely love the most or whatever the condition is that makes me want to jump to the next: I'll then try to add maybe a random triangle or something as it'll change the dynamics so i'll train another network to try to get the most pleasing combination out of that .... there's so much room to play and there's so many surprises along the way.

When I am shown the random pieces for me to rate I might not see the piece i see that is best. because I can't physically go through millions of random ones, but with neural net, it can fly through them really fast and then if trained well, the ones it returns are usually really good.