This page focuses on the technical challenges of our project. Click here for details about the ethical challenges.
Most chatbots are created around a specialized and very limited topic. They usually serve as a first contact for customers who have questions concering a certain service provided by the company offering the chatbot.
DEUS aims to have much longer philosophical conversations on complex topics and even in different schools of thoughts. So at the beginning we checked if any other teams have already started to train any AI on religious or philosophical topics but we found none.
On top of that training individual chat-bots for every single religion would not be possible within the limited resources of the University. On the one hand there are issues with the original texts. Many scriptures like the bible or the qumran use a very poetic language with a lot of images. This opens them up to many different, sometimes even conflicting interpretations.
In spring 2020 Sana Nouzri, the scientific supervisor of the project at the University of Luxembourg, requested access to GPT3 which was granted.
Working with GPT3
GPT3 is the state of the art model when it comes to conversational AI. The project of the OpenAI research and deployment company is being trained on all English texts available on the internet, thus representing the globally accessible knowledge available in a digital form.
The big drawback is that users have hardly any control nor can they change its training in any way. Nevertheless we continued along this path.
First experiments and the “Nice” mode
One of the surprising findings of the students in this first phase was that some DEUS-identity could turn out to argue in a very aggressive manner. Since GPT3 is being trained on all the available texts on the web, it will also draw on radical and brutal ideas represented in the internet.
Therefore we had created the “nice” mode as an option. If activated it sets up the chatbot to be more peaceful and open-minded as well as forbids the use of certain banned words that have been encountered during testing and reliably created very provocative and derogative interactions during the discussions.
This website is continuously improved. The last update of this specific subpage has been in March 2022. If you are interested in a collaboration or you have spotted a typo or a blocked link, please do not hesitate and contact us at email@example.com