#oh yeah nigthwish and glaze are very heavy on your computer so do take that into account
Explore tagged Tumblr posts
Note
Hey ik this is not tournament related, but in case you didn't know and want to spread the word, Tumblr is selling everybody's data to AI companies.
Here is the staff post about it https://www.tumblr.com/loki-valeska/743539907313778688
And a post with more information and how to opt out https://www.tumblr.com/khaleesi/743504350780014592/
Hi thanks for the information and sorry for my late reply. I was a bit low on spoons this week and I wanted to form thoughts about this.
Because the thing is, I am doing a PhD at an AI department in real-life. Not in generative AI, in fact Iām partly doing this because I distrust how organisations are currently using AI. But so this is my field of expertise and I wanted to share some insights.
First of all yes do try to opt out. We have no guarantee how useful thatās going to be, but they donāt need to be given your data that easily.
Secondly, I am just so confused as to why? Why would you want to use tumblr posts to train your model? Everyone in the field surely knows about the garbage in, garbage out rule? AI models that need to be trained on data are doing nothing more than making statistical predictions based on the data theyāve seen. Garbage in, garbage out therefore refers to the fact that if your data is shit, your results will also be shit. And like not to be mean but a LOT of tumblr posts are not something I would want to see from a large language model.
Thirdly Iāve seen multiple posts encouraging people to use nightshade and glaze on their art but also posts wondering what exactly it is these programs do to your art. The thing is, generative ai models are kinda stupid, they just learn to associate certain patterns in pictures with certain words. However these patterns are typically not patterns weād want them to pick up on. An example would be a model that you want to differentiate between pictures of birds and dogs, but instead of learning to look for say wings, it learns that pictures of birds usually have a blue sky as background and so a picture of a bird in the grass will be labelled as ādogā.
So what glaze and nightshade are more or less doing is exploiting this stupidness by changing a few pixels in your art that will give it a very different label when an AI looks at it. I can look up papers for people who want to know the details, but this is the essence of it.
To see how much influence this might have on your art, see this meme I made a few years ago based on the paper āIntriguing properties of neural networksā, Figure 5 by Szegedy et al. (2013)
Finally, staff said in that post that they gave us the option to opt out because of the maybe upcoming AI act in Europe. I was under the impression that they should give us this opportunity because of the GDPR and that the AI act is supposed to be more about the use of AI and less about the creation and data aspect but nevertheless this shows that the EU has a real ability to influence these kinds of things and the European Parliament elections are coming up this year, so please go vote and also read up on what the parties are saying about AI and other technologies beforehand (next to everything else you care about) (also relevant for other elections of course but the EU has a good track record on this topic).
Anyway sorry for the long talk, but as I said this is my area and so I felt the need to clarify some things. Feel free to send me more asks if you want to know something specific!
#misc#ai#tumblr#oh yeah nigthwish and glaze are very heavy on your computer so do take that into account#nightshade*
95 notes
Ā·
View notes