We have recently seen two major announcements on the new class of Image generating AI models which can take natural language prompts / triggers and generate fascinating images.

The first one is DALLE2 from OpenAI and the second one from Google called as Imagen.

When the entry barrier to generate amazing images is lowered down to writing some prompts, it was about time that someone who do some creepy experiments and try to reveal more insights into the black box AI models.

Giannis Daras posted this thread on Twitter giving a title ´DALLE2 has a secret language´. I am still awaiting my access to DALLE2 to try similar experiments. Google on the other hand has decided not to open the access. So, we will not learn until researchers at Google try to do similar experiments at their end and publish more insights.

In the paper that Daras has published, he gives examples of a prompt ´Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons´it generates images of birds eating bugs. Similarly when they ask DALLE2 to generate the images with subtitles, the subtitles are gibberish text.

We do not fully understand what is happening here but it seems like that AI model has a representation of world model in its novel way. We probably need to have more deeper studies and experiments to understand if this representation is just another language / proto language (or just some incomplete vocabulary of concepts) and how superior / inferior it is in representing compared to other languages like English or French etc.