It's been said that you can tell that a photographic image was generated by artificial intelligence if the number of fingers or teeth are wrong. The industry is well aware of that, and recent updates have dramatically improved how many fingers humans have in such images. But they still have a way to go in understanding humans overall.
Dan Leveille has been experimenting with the March 16th update to the AI program Midjourney. While the people generated are more realistic than ever, Leveille noticed it has a real problem understanding spaghetti and how people in the real world eat it. It's as if spaghetti were specifically invented to make a mess. You can see more images at Geeks Are Sexy.
But let's consider how neural networks learn. They are given boatloads of publicly-available images from the internet. Considering that the 'net is a repository of the world's knowledge, that makes sense. But what kind of spaghetti-eating photographs do everyday people post on social media? The times it makes a mess. There is little entertainment value in posting a picture of someone eating spaghetti in a normal manner, and we don't really want to share pictures of ourselves eating. But we love to make other people laugh. Midjourney was most likely inundated with thousands of images of toddlers and drunk people trying to manage a plate of spaghetti with hilarious results. So this is what you get.
Newest 1 Comment
thats how i usually eat spaghetti
Abusive comment hidden.
(Show it anyway.)