What AI can and cannot do
Hey guys!With this article I want to clarify some common misconceptions about AI in general and image creation in particular and thus help for a better understanding of what it is that we are working with here.First things first: The term AI is misleadingWe're not dealing with real AI here, not at all. What we are dealing with here are scripts that are able to make decisions on their own, but only when prompted to do so.Many people confuse our actual concept of AI with that of science fiction movies, and the reality could not be further away from that. In SciFi AIs are thinking, conscious machines with an insane amount of control over their surroundings, which makes for a good movie villain, but is not how it works at all.Our AIs are not AIs, they are comparison tools that can create derivative work out of their databases. So when you type in "red sports car at nighttime on a vivid street with neon lights," the script will look up its database for every word in the prompt, collect fitting images and image patterns and then create a mix of all the elements given in the prompt. For every element there is a certain probability of being depicted and it depends on the model how high this probability is (given you don't use emphasis).To make this totally clear: The way our AIs work means that they cannot create anything they haven't seen before!So if one would create a model or a checkpoint with no car pictures in the training data, you could not create a single car with that model because there are no pictures it could derive something from.So what does this mean?It means several things. First: Our AIs do not think. This fact alone disqualifies them from being called AI to begin with.It also means they do not understand, grasp or get an idea of what you want it to do.So whenever you try to create an atmospheric image, it's not the AI that creates the atmosphere, it's you. Let's take a horror image for example. You can already tell the AI to create an atmospheric horror image, but I want you to be aware what happens in the background when you type in these words and hit "create"The script runs through its database and looks for images that contain "horror" and "atmospheric" as keywords and will most probably find images with dark backgrounds, stone walls, chains, skeletons and fog.So depending on the training data both approaches will lead to pretty similar images - be it "skeleton laying on stone floor, chains on the wall, foggy, dark background" or be it "horror image with skeleton that has a chilling atmosphere"Which of both prompts is better suited for the image you want to create solely depends on the model and the way its training data was created and implemented.So no romance in AI?I know this sounds a bit sad to many, since they often imagine that if they use emotional words and descriptions the AI will better understand what they mean... Well you can use AI this way ("mesmerizing beauty and breathtaking atmosphere") and get really good results. But, and that's my point, you can use totally technical descriptions ("beautiful, posing, sunset") and get the exact same results.Because our AIs neither understand anything, nor can they grasp the meaning of something. They only compare pictures and picture patterns that fit the prompt you give it, and that's all they do. I know this sounds disillusioning to some people, but it is what it is.Language models as counter argumentYes, language models like ChatGPT are different. They can understand things on a technical level. They are like a child that never in his life has seen a real dog but only read about it. So it can perfectly describe the properties of a dog, but at the end of the day it still does not know what a dog is. You can believe me here, I talked to several AIs about this very topic, and all told me the same.So are AIs great tools that will bring technology and society forward? Most probably yes.Can they think? No.Can they feel? No.Are they conscious? Hell no. They still are scripts. Computer programs that stop doing anything when you do not input anything, and that is not how a conscious being works.But are they cool? Hell yeah, they are, and I'm eager to see where all this is going.
