According to a computer science PhD student at the University of Texas at Austin, DALL-E2 appears to have created its own language.
DALL-E2 made a big splash earlier this year with its ability to turn text into pictures, and now it sounds like it may be even more powerful and mysterious than we thought.
In order to generate an image of "farmers talking about vegetables", Daras asked the algorithm to use the word "vicotes." It spits out pictures of vegetables if you put the word "vicotes" into the program.
The image of the farmers contained a text that looked like a joke. You get a bunch of pictures of birds when you feed that into the system.
Farmers are talking about birds and vegetables. The writer was named Daras.
This advanced image-generating artificial intelligence appears to have generated a "hidden vocabulary" underlying or working in parallel with its primary function, according to a new yet-to-be-peer-reviewed paper written with a colleague.
The discovery of the DALLE-2 language creates many interesting security and interpretability challenges, according to the author.
The results aren't 100 percent consistent as he notes in the paper. Sometimes the prompt "Contarra ccetnxniams luryca tanniounons" will generate pictures of bugs while other times it will generate pictures of mostly animals.
A lot of people are not convinced by the theory.
BenjaminHilton asked, "My best guess?" Random chance is what it is. It looks like a name for some birds or bugs.
This is starting to look more like random noise than a secret DALL-E language to me.
Vice points out that there is a chance that we're readying too far into it.
It's going to be interesting to see more scrutiny of the claims from the research community.
If it passes muster, we may be looking at a fascinating new development in the field, or we may be giving it too much credit and reading between the lines.
We don't understand how image-Generating Artificial Intelligence keeps doing weird things.