Post
It's so weird to me that this is used as an example of Al working well. It's meant to by a guy getting distracted by a passing woman and his girlfriend getting annoyed about it, but the model has drawn a woman randomly getting mad at a passerby and her boyfriend being annoyed about it. We've built an "artificial intelligence" that can't understand one of the best-known memes on the internet. Any slightly online person could draw this meme better than the AI did — the drawing might be less technically neat, but the idea would come across better. We only recognise this image because it resembles the original. Nobody would understand the meme from the the AI version if they saw that one first.
https://media.mathstodon.xyz/media_attachments/files/114/285/808/379/498/321/original/547aa8fe52ccaa14.jpeg
Brief Dragon That's interesting because it mirrors issues with machine translation. Like an idiom translated literally, this is a case of data being insufficient to cover the lack of understanding. "Leering at a girl's ass" is not in your vocabulary if all you know about people's faces is Ghibli cartoons. And since a machine has no understanding of what's going on, it can't swap it with the appropriate cartoon equivalent. Exactly like MT fails at any creative solution at cultural issues...