Artificial intelligence is capable of doing several tasks. However, have you heard of it developing its language? Yes – An artificial intelligence program has created its own language and nobody can understand it.
OpenAI is an AI systems developer whose programs are the best examples of super-computing. But even the best has their own twists.
OpenAI ‘s latest AI system is DALLE-E2 capable of generating realistic or artistic images based on text descriptions entered by users.
DALLE-E2 illustrates a breakthrough in machine learning. According to the OpenAI website, the program learned the relationship between images and the text adopted to narrate them.
A demonstration of the DALLE-E2 comprises interactive keywords which are provided to the users for playing and generating images – switching different keywords results in different images, styles, and subjects.
However, the system exhibits one peculiar behavior: it creates its own language out of random letter arrangements, for which researchers have no explanation.
Giannis Daras, a Ph.D. student in computer science at the University of Texas, detailed DALLE- E2’s mysterious new language on Twitter.
Daras instructed DALLE-E2 to generate an image of “farmers talking about vegetables,” which the program did, however, the farmers’ speech read “vicootes” – an unknown AI word.
Daras entered “vicootes” back into the DALLE-E2 system and received images of vegetables in return.
Daras wrote on Twitter that they fed the words ‘Apoploe vesrreaitars’ and got the word birds.
It appears that the farmers are discussing birds messing with their vegetables!
Daras and a co-author have published a paper on the “hidden vocabulary” of DALLE-E2.
They admit that telling DALLE-E2 to create images of words – An example given by Daras is when the command “an image of the word airplane” is given – usually results in DALLE-E2 spewing out “gibberish text.”
When that gibberish text is plugged back into DALLE-E2, it produces images of airplanes, which reveals something about how DALLE-E2 communicates with and thinks of itself.
Some AI researchers disputed that the gibberish text of DALLE-E2 is random noise.
Hopefully, we won’t find out after it’s too late that the DALLE-second E2’s language was a security flaw that needed to be patched.
