Raw LLM Responses

currently pursuing a bachelor's (and then master's) in computer science and the comment at 14:53 made me BURST out laughing "the programmers are just gonna make the ai recognize (the) poison itself and try to cure (it)" is just... tell me you don't know how artificial intelligence works without telling me you don't know. artificial intelligence models cannot actually "see" images like we do since they are MACHINES and *CODE*, they literally don't have EYES. What they "see" is really a series of analyses of the pixels (formatted in arrays, vectors, lists, etc) aka pure data. give it enough pictures of dogs and the model will realize "oh! these lighter-colored pixels are always sticking together in this certain portion of this dataset! I will make my output look like that!" it doesn't realize (or see) what it's actually reading and outputting, and this is why we always get those really weird "hair melts into clothes" and "her hand becomes her neck" anomolies: the models cannot see they're making those errors, because they cannot see in general. trying to program an ai model that can predict if an image is poisoned is, to make a long story short, an impossible and very, very stupid idea. because these algorithms can only "see" data (individual pixels), there is no precise and clear way to do what that commenter claims, ESPECIALLY with the "curing" bit. that part almost made me throw up laughing. Theoretically, the best someone could do is train a model on a large enough dataset of poisoned artpieces to try and determine the patterns of poisoning; I say "theortetically" because 1. that would be a very hard job to begin with and 2. poisoning softwares like Nightshade and Glaze can be evolved into using different poisoning patterns or evolve to have a mutating pattern that cannot be easily identified by an algorithm (like how some modern-day viruses use metamorphic and polymorphic code to evade anti-malware systems); in either case it's a losing battle for ai tech bros, which is why we haven't seen any "anti-glaze" or "anti-nightshade" brags pop up from these people: they simply can't do it!