Algorithm Behind the Artificial Intelligence in ChatGPT

By Adi Permana

Editor Adi Permana

BANDUNG, itb.ac.id – In recent years, the world of artificial intelligence has shown impressive advancements. One of the recent achievements is a sophisticated natural language model development, ChatGPT. The model that is developed by OpenAI has gained massive popularity due to its ability to naturally interact with users.

To talk about this topic, ITB PR Reporter interviewed Dr. Eng. Ayu Purwarianti, S.T., M.T., who is a lecturer from the Informatics Research Group, School of Electrical Engineering and Informatics (SEEI) ITB. Dr. Ayu is also the Head of Artificial Intelligence Center at ITB.

“ChatGPT is one of the applications of Artificial Intelligence, more specifically in the branch of natural language processing which utilizes Pre-trained Generative Large Language Model. This model of Artificial Intelligence was built on unsupervised data,” said Dr. Ayu.

Dr. Ayu explained that one technique in Artificial Intelligence is machine learning, which when we build an AI model with machine learning, we need to provide it with data. The data then will be inputted to the machine learning algorithm to create an AI model from the studied patterns. The data may be supervised or unsupervised.

Dr. Ayu then continued, if supervised data is used then it needs to be given a label manually, an example of this is data from social media which need to attach a positive, neutral, or negative by a data annotator. We currently also have a semi-unsupervised data type which will be labeled automatically but still needs to be reviewed later. While unsupervised data do not need to be attached with labels, these two kinds of data are used for different purposes.


How ChatGPT Works
Source: Dr. Eng. Ayu Purwarianti’s Presentation

Furthermore, Dr. Ayu explained that ChatGPT uses generative techniques which are built from trillions of sentences processed by deep learning methods. When we ask a question, ChatGPT will provide a generated word-by-word answer. Because its data source is so massive, most of its generated answers are correct.

Dr. Ayu Purwarianti (Photo: ITB PR Documentation)

When ChatGPT processes this massive input of sentences into its deep learning algorithm, it does not store knowledge. What it does is adjust the weight of its deep learning, ChatGPT model implicitly stores the correlation between words.

ChatGPT has provided significant contributions in expanding the use of artificial intelligence in day-to-day conversations. It is expected that with this advancement of technology, we ought to act more wisely in addressing the use of ChatGPT, as well as the consequences that follow.

Reporter: Satria Octavianus Nababan (Informatics, 2021)
Photo: Freepic
Translator: Favian Aldilla R (Civil Engineering, 2019)