Raw LLM Responses

@ThatWeirdGuy43 no not really true if i look back. Dall-E 1 took the same type of scraped data and noone cared. Why did noone care back then? becasue it looked like crap. But now it looks very legit and in some/many cases even better than human output. This is why there is NOW outrage and not a year ago. I would bet an eye that if the models outputs would look like garbage noone would care. To give you another argument. Everything is remixed and sampled in some way nowadays. Do you like hiphop? well it is based on using breakbeatsections of old funk tracks from the 70s and Turntablists (DJs) using two of the same record and loop them via crossfading. This whole genre (the msot popular genre in music rn) has sprung into existence out of sampling others work in a transformative way. Sure training the Image models is in so far different as it is way way more efficient at iterating and blending between different learned cocnepts, styles and motifs. But in it's raw principle it is not really different than hiphop. Especially since ti is sparking new forms/media of art. For example AIanimation or using reallife footage as inputs to dream over and making animated dreams of that data. Or if you want to think a bit further creating full feature lwength movies out of a vast latent space being the result of all of humanities publically avaliable data. "but wait so it is the work of others that created this right?" no it is not. You need 5 thousand A100 GPU's to train these models. These GPUs cost about 20k - 30k dollar each alone. Running them takes a lot of money to spend. The data that has been consentually shared on different platforms such as instagram, deviant art and co falls under fair use as long as you use that data transformatively. Since thsoe models DO NOT CREATE ANY SORT OF STENCIL of the original imagery and do not use a single bunch of pixels from the originals but instead learn (yes jsut like a human brain) visual concepts and have them embedded together with their Text description ( collected via webpage metatext) into latentspace. Latentspace is a sort of visual library similar to a mindmap ot a plot of concepts where similar concepts or visual characteristics are closer to each other and form clusters so to say. These do not contain the input images as i said but instead learned cocnepts on what are the characteristics of those images. How is a sky shaded, how is an eye drawn, how is a cute cartoon dog stylized. That is pretty similar to the internal images you see in your mind when imagining something (unless you have aphantasia) If i tell you to imagine a picture of an anime boy fighting a dragon you have an fuzzy image in your mind right? You see things like how an eye is drawn or how the boy is smirking while grabing onto his sword to slay the dragon. This is pretty much how these models work. Just as you have seen a lot of anime that tought your brain how anime is looking those models do similar connections and form their visual library. Sure they are not making creative decissions other than interpolating between those concepts to match a given input prompt or input image by the user to generate the best fitting output corresponding to a position in that multidimensional latentspace(for example crude painting you made in paint and load into the model of a pose with a prompt like " knight holding a sword" I do not mean training tho just using the model.) So as an artist myself that worked traditionally, digitally, in 2d and 3d and with AI I can tell you this is extremely transformative. Maybe more transformative than anything i have ever seen tbh. Sure it uses the dataset to learn concepts right. But the output is still not the input. So banning this would mean to have consistency you would need to ban all of anime (same style) all of fanart and cosplay (design and style of private IP) all of reviews for movies and games (copyrighted material is being used without consent- sometimes to critique/discredit the reviewed product) and tbh the most scary part... research of other AI modfels such as large language models like gpt-4 which could ultimately solve many humanitarian problems and make humanity way more productive.. or even algorhythmy like AlphaFold which is currently being used to create a lotz of medical advancements and was just exactly like those image models trained on data of individuals that have not given SPECIFIC CONSENT to be trained into those proteinfolding systems. So if we would heavily restrict datascraping we would ultimately not save as many lifes as we would otherwise since we would slow down progress and 10 million die yearly of cancer alone (which might be curable thru those technologies) Do not forget that Art is one of the smallest parts of AI research. Who says that text data like the scraped data used to train language models is not as well fair to use for research after the artists-against-art mob has their say in a trial. I fear we are not doing the right thing in combatting this to save the old status quo but are in fact delaying necessary reformation of our whole system since those algos are coming FAST. and any month not spent on makin our society future proof will cause much more devastation and chaos later when the majority of people WILL BE AFFECTED by automation and us not having mechanisms to combat this systemic problem in place. Honestly a big disservice to humanity if you ask me. And I am an artist that studied contemporary arts at a acedemy (painting) and works since a decade in the entertainment industry as a creative. So I am well aware about the disruption and pain this causes to artists. but still this is a necessary evil thinking about climate change, world hunger, cancer and so on. We wont be alone as creative workers. This will hit MOST industries sooner than most people might think. Hell I am heavy into machine learning since a decade and it hit even me out of the blue how great diffusion models became all of a sudden. Hope this brings you a bit of my perspective on the subject and why i have this opinion. I want the best for all of us. And the current situation (pre AI) is not it. in fact its a living hell for many people on this planet.