It doesn’t have to be insanely specific, but I guess when I prompt it with “pizza with pineapple”, why can’t it go through all of the connections that helped it understand what I mean by “pizza” and “pineapple”, and then site what images helped it and how much each image did
In: 0
Because the way the models work, they draw on ALL the samples in its corpus.
That would be like you being asked to write an original poem and cite all the things you’ve ever learned that helped you create the poem.
Are you referring to AI image generation models?
They are trained on a large number of images, but they no longer contain the images. The images have taught them all sorts of things about what things look like, but the models no longer have access to the images they were trained on. If the models still contained all the training images, they would be 10,000 times their current size, and nobody would be reasonably able to download them.
Assuming you are talking about an image generation AI, because the source is every image it has in its database.
It uses the images of what you want to know what to draw and everything else to know what is similar to what you want and the rest to know what not to draw.
Though, that answer isn’t even entirely accurate because it doesn’t really have a source because it’s a giant pile of neural nets, math, patterns, grammar and random numbers that has mostly learned how to draw things that you tell it to so what it’s drawing isn’t entirely connected to any sources more than any others.
As a analogy, asking an AI for its source is kind of like cutting down a forest, mulching them all down into a giant vat of pulp, mixing it up and turning it into paper, then asking which tree a particular sheet came from. It might have come from a few more than others, but it’s impossible to say for sure and kind of meaningless at this point anyway.
Latest Answers