Various artificial intelligence systems are already capable of creating things that seem completely alien to humans. Moreover, there have been cases in history where artificial intelligence systems invented their own language, alarmed researchers so much that they even preferred to reset the system, breaking the chain of intelligence that went in a direction they could not imagine. This time, the DALL-E 2 system, created by Google specialists, was originally intended to create images based on a textual verbal description.

Note that the neural networks of the DALL-E 2 system are pre-trained on a large number of images accompanied by a detailed text description. Built-in neural connections create a kind of database containing relationships between symbols, words, and elements of images.

“Miracles” with the DALL-E 2 system were noticed when independently given the task of creating a text description and the corresponding image. For the vast majority of names, it turned out to be in the bowels of the system, which, as he puts it, works in its own incomprehensible language.

The researchers believe that the main reason for what happened was that the descriptions of the images the system was trained on were presented in different languages. This forced the system to bring these descriptions to some “mean denominator” and compare the result with the elements of the images. The principles the system uses to create the “middle denominator” are still a mystery to scientists, as is the mystery of a kind of new language made up of averaged data.

Researchers see a potential danger in the artificial intelligence system creating its own language. Imagine this language being deciphered by someone who doesn’t have very good goals, then that person will be able to make this language a task for artificial intelligence and in doing so attack the Internet or other critical infrastructure. And other people won’t even be able to understand what happened and for what reason.

Therefore, the researchers isolate the DALL-E 2 system from the rest of the world, observing the processes taking place in it and what will subsequently lead to its further development. However, it is worth noting that there is another version of such a system in the world, the DALL-E Mini, available to the public, but with much more limited capabilities compared to the DALL-E 2.