Please explain how a human brain does it.
Earlier in this thread we got into a discussion of we don't know exactly how the brain does it, but we know how it does not. So being able to explain how the brain does it isn't a requirement to being able to discuss how AI does it in a way the brain does not.
But all of that said, let me give a try of the
process using human vs. AI art.
A human art capable of producing a realistic image similar to what AI arts does envisions the various objects which exist in a 3D world, establishes a point of view from which to render it. They translate from 3D to 2D, including what can been seen and what blocks view, where light sources are, perspective and foreshortening, etc. They start with the real world and then from that move to how does it look in an image. Heck, we've had had 3D render programs for a long time that are made to emulate that process, and it's what 3D games do.
AI art is models generating it via statistical analysis. It does not involve that process at all, since there was never a 3D model to translate. A human artist could redo the exact same scene but drawn a few degrees to the left and a foot forward. An AI model can't.
Edit: because here's the thing: in my profession (teaching) we are really struggling with what to do about AI, since in many ways, it writes better than most humans. But also since it suggests the a lot of the things we thought were exceptional about humans...maybe not so much.
"Writes better."
One common issue with LLMs (Large Language Models - AI writing) is what they are now calling "hallucinations". I'm not fond of that as a descriptor but it's in common usage. If they have information, they can use the information. If they don't have the information, they often will make up information. Not so different from human - except that they can't tell they made up the information. They can sprinkle falsehoods and incorrect information in, and don't know.
An example of this was with ChatGPT-3.5, we were playing a new boardgame, Dice Theme Park, and asked it for strategies. There were whole sections about Mascots and such that just didn't exist in the game, but were presented with the confidence of everything else.
A human writer would know when they are bulling around. But there is no "they" to understand this with LLMs. We anthropomorphize them because it seems like someone talking to us, and because we as humans anthropomorphize lots of things. Pets. Cars. Computers. What have you.
Instead it's taking the current and previous prompts and statistically generating words. It's spicy autocorrect. Yes, it's the Porsche of conversation compared to the Horse-and-Buggy of conversation of autocorrect, but being more advanced just means it's better at it's job, that it picks the right words, not that it's actually thinking about the concepts.
Generating output from input that looks human - yes. Is generated by the same process - not at all.
Frankly, it's the anthropomorphism that's a big part of the perception issue. Because people treat it like a human, they mistakenly compare it to how a human would learn.
