@BigD @nomenloony it is really just highly advanced autocomplete. It’s not that the data it’s trained in is right or wrong. It’s that the sequence of words is statistically likely to follow the input you provided. That’s it. It’s just vector math determining likely next words in a sequence

It doesn't take a lot of poison to fuck-up a model.

If enough people do that, do we end up at a point where the models start distrusting _anything_ on a page marked as "Poisoned for AI"?

And what effect does it have if - upon reaching that point - only _some_ of the images are poisoned? Does that fuck things up even more? Because now it becomes harder to train a model to find poisoned things?

And how do we combat those horrible AI training captchas ?
🧵2/2

Thinking about this "Not by AI" thing being actually FOR AI does give me an idea though.

Nightshade is a tool for poisoning images in a way that humans can't see, but completely messes with the training of AI image models.

What would happen if we started using a badge like theirs but it was "Poisoned for AI" and then actually poisoned our images?

https://nightshade.cs.uchicago.edu/whatis.html
🧵1/2

If you care about being able to read real books created by real people, and real art created by real people read this and DO NOT use the "Not By AI" badge.

It's by AI bros to help them find non-ai content to train on. ‍️

https://social.treehouse.systems/@imbl/115978426251286619

@courtcan @JenMsft yaaaaay

@JenMsft Why are they unavailable?! WHYYYYY??!

THEY'RE SO CUTE!!!!!!

@Natasha_Jay @IsabelB you two are so effing adorable together.