Bugs in the mutual interaction between ChatGPT and DALL-E 3 show that they communicate with each other in human language.
So writes Ars Technica. Bugs in images generated with ChatGPT and DALL-E 3 show that both AI models communicate with each other in a special, even human, way in the generation process.
Bugs found in the image results show, among other things, the prompts and responses that both AI models ask each other. The communication takes place in plain language.
This involved instructions that both systems gave each other on how to act when OpenAI’s servers for processing end-user prompts had reached their maximum capacity.
Communication in ‘normal’ English
Notable in this was, for example, the use of capital letters to make it clear what it was about. In normal virtual text usage, capital letters represent saying something in a loud tone. This therefore gives the AI models real human traits, writes Ars Technica.
According to the tech website, in the past, the APIs of both AI models mainly communicated with specialized data formats that were not readable by humans. Nowadays, according to the discovered bugs, this mutual interaction can thus clearly take place in plain English. This functionality was previously implemented by OpenAI with ChatGPT plugins.
Dependence on prompt engineering
Experts say it is therefore fascinating to what extent the AI giant still relies on prompt engineering to develop its generative models. In addition, they marvel at the amount of polite language put into this normally hidden communication.
This would come from training GPT-4, for example, on millions of documents that also contain a lot of polite language.