The automatic or self-development of artificial intelligence (AI) has been speculated since the term started to be commonly used within the tech industry. A new generation of AI named DALL-E 2 has been identified by tech experts to have developed its own secret language or better-said vocabulary to create digital intellectual property.
This new AI model has been developed by some of the brightest minds around the world to create realistic images and art from a description in natural language. The AI takes text description as input and searches the internet for resources to create original and realistic images based on the text description given.
A Ph.D computer science student from the University of Texas in Austin has managed to decrypt the language used by DALL-E to process information and produce creative content. Giannis Daras with the help of Alexandros Dimakis has managed to identify the secret vocabulary used by this AI model.
Although the AI has been coded to use English, it seems to have created its own language as it works more efficiently when searching the internet for resources to use in order to create original content.
Besides some of the information presented on Twitter by the student, he has also written a research paper where he explains how he identified the hidden vocabulary of DALL-E-2. The students have used something called a "black-box method" which is used to discover the meaning behind random words that have some correspondence to visual concepts. Although they understand the basics of this weird technical language, it is hard to grasp how the AI developed this language.
Something similar occurred with an older AI engine used by Facebook which was shut down in 2017 as it created its own language which wasn't understood by Facebook engineers and therefore it was considered a threat to the platform.
In fact, a lot of AI models may be running under their own secret vocabulary which in most cases is self-developed. The reason why AI creates its own language is that it is easier for them to process information. An AI does not think or read a text like a human mind, they break input text descriptions up into what computer scientists name "Tokens".
However, there are times when these tokens created by AI are confused by words with two meanings such as match. AI self-developed languages can't comprehend words with two or more meanings.
This is not necessarily an "AI taking over the internet scenario" but due to such discrepancies between the AI's own language and English, sometimes the AI may not be able to process information accurately, therefore wasting processing power and also have people lose faith in AI.
The world is looking at a not very far future where AI can be implemented across all sectors due to its high efficiency and little need for maintenance. However, if we are talking about the financial or medical sector, these sorts of discrepancies between languages and the way AI understands our inputs could be fatal.