Yes, because we’re no longer at the point of “programming” by writing in instructions (in the context of AI models). We are legitimately moving towards training AI the way a child learns. These AI image generators are given millions upon millions of images with descriptions, then they learn the general features. So once it’s done and you give it a command, it does its best to create something from what it learned. And right now these AI’s don’t differentiate between the number of hands. Mainly because a good majority of images are showing clear views of hands. Most hands will be at an angle or even sideways. It also has no concept of how many fingers a person has. It’s actually an incredibly good image.
This is totes an actually super real picture, but I do feel bad for this kid. He’s lost a finger on one hand and regrown it on the other.
The wrong number of limbs seems like a very common AI mistake. Is that really hard to program/teach an AI system?
Yes, because we’re no longer at the point of “programming” by writing in instructions (in the context of AI models). We are legitimately moving towards training AI the way a child learns. These AI image generators are given millions upon millions of images with descriptions, then they learn the general features. So once it’s done and you give it a command, it does its best to create something from what it learned. And right now these AI’s don’t differentiate between the number of hands. Mainly because a good majority of images are showing clear views of hands. Most hands will be at an angle or even sideways. It also has no concept of how many fingers a person has. It’s actually an incredibly good image.
What goes next to a finger? Another finger. So they offer fingers next to fingers and don’t always stop at the right one.
https://youtu.be/24yjRbBah3w?si=ZSc3WDWyK6gM2aLJ
Best comment in thread