The neural network accidentally came up with a new language. Here’s what happened when they wrote her a request on it

The recently introduced DALL-E 2 neural network has already managed to impress programmers and ordinary people around the world: it generates images of a very high quality for any text query, even if things that do not exist in reality are indicated. However, it has a limitation: the neural network does not know how to work with text, therefore, when a request is entered to generate a dialogue or replica, it draws gibberish instead of words. However, the researchers found that this is not at all a meaningless set of letters, as it seemed before – it is a full-fledged language that DALL-E 2 understands.

For example, for the request “Two farmers talk about vegetables, with subtitles”, the neural network generates the following picture:

If you take the text from this image (“Vicootes”) and enter it as a request to generate a new image, DALL-E 2 will generate vegetable images:

This is not an accident: for example, when generating images for the request “Two whales talk about food, with subtitles”, the neural network generates a completely different phrase – if you enter them as a request, images with seafood will be created.

Moreover, DALL-E 2 understands combined queries from English and her own language. For example, the researchers found that “Apoploe vesrreaitais” are birds, and “Contarra ccetnxniams luryca tanniounons” are insects. On request “Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons” the neural network created pictures of birds that eat insects.

Source: Trash Box

You may also like

South Korea’s billing bills
World
Flora

South Korea’s billing bills

A bill was approved yesterday, Wednesday, by South Korea, which prohibits the use of mobile phones in the classrooms, as