Assist. Prof. Dr. Ruaa Abdulridha Saeed Alsabah
Assist. Prof. Dr. Ashwan A. Abdulmunem
Computer Science and Information Technology
A new revolution in the search engine industry is looming with the emergence of “ChatGPT,” a modern technology that utilizes artificial intelligence to provide accurate answers to people’s queries while also admitting errors and dismissing inappropriate questions. OpenAI, a research startup co-founded by Sam Altman, Elon Musk, Ilya Sutskever, John Schulman, and Greg Brockman created an innovative paradigm of natural language processing which is called ChatGPT [1]. The idea of ChatGPT is to generate human-like text answers to prompts using a Natural Language Processing (NLP) model that trains massive quantities of data and comprises billions of documents and web pages [3]. OpenAI’s ChatGPT is a strong language model that generates human-like prose using deep learning techniques. In 2018, OpenAI presented ChatGPT, based on the GPT (Generative Pre-trained Transformer) architecture, which is a deep learning model trained on vast amounts of text data over 570GB. In the beginning, GPT-1 trained and dealt with 117 million parameters using a deep learning method. This method was called transformers on a sizable quantity of text data downloaded from the internet. In February 2019, GPT-2 significantly enhanced and started with 1.5 billion parameters.
Due to concerns over the model’s potential for improper usage, the full version of GPT-2 did not release. It was accessible to the general public by OpenAI (only 8% of the original model’s size)[4]. In June 2020, GPT-3 was made available with 175 billion parameters, and the queue was later eliminated in November 2021. In November 2022, by the time ChatGPT was live, it evolved to version 3.5. Users who joined the ChatGPT Plus subscribers were given limited text-only access to GPT-4 in March 2023. It may, however, react to both graphics and text. Although GPT-4 has a restricted availability, it has drawn notice for performing better than its forerunner. GPT-4 has better performance than GPT-3.5 Because of the more parameters adjusted and larger model during training in a neural network. OpenAI, until this time as of April 2023, has not yet disclosed any details on training methodologies utilized, the data, or the computer power used to create the language model [5]. GPT-5 is expected to be In November 2023 by OpenAI.
Although the concept of artificial intelligence is not new, it has been making headlines more frequently lately. The technology has been around since the 1940s when computing methods were used to intercept German radio messages during World War II. It wasn’t until the late 1990s that the technology gained widespread attention when a computer defeated a world chess champion. Since then, smart computing has become more readily available and integrated into our daily lives. 2023 will go down in history as the year of the artificial intelligence revolution, with the widespread adoption of the ChatGPT platform and, later, its upgraded version, GPT4, causing significant upheaval in most industries, education, and the media, leading to endless debates about the credibility of the content generated by the technology. ChatGPT has been the subject of controversy in the technology and media communities, as it generates copies of articles instead of images and videos. However, it has faced legal challenges regarding intellectual property rights for the content it copies and regenerates from the internet.
Figure 1 shows ChatGPT methods. To train this model, reinforcement learning from human feedback (RLHF) was used to train this model, which is the same as Instruct GPT except the data collection order is slightly different. By having human AI trainers, supervised fine-tuning is used to train a prototype. These trainers act as assistant AI and the user in chats. Access to sample written recommendations is provided to the Trainers to assist them in formulating their responses. The InstructGPT dataset is converted to a conversational format, and then merged with the new dialogue dataset. To construct a reward model for reinforcement learning comparative data are needed, which include at least two sample responses ordered by quality. To obtain this data, Chabot interactions with AI trainers are used. A typically written statement is then randomly selected, sampled from several possible conclusions, and asked to rate the AI trainers. In addition, to adjust the model with other reward models, proximal policy optimization may be used. This method has been repeated many times [6].
ChatGPT become one of the fastest-growing consumer apps in history, with an estimated 100 million active users monthly. The high language processing skills and the massive knowledge base are the main reason to make machine communication easier and more natural and to transform the way humans engage with new technology. Although ChatGPT is an exciting technology with a wide range of important applications in different variety of fields and impressive language processing capabilities, it still has some drawbacks and difficulties, such as bias and the infrequent generation of absurd output known as “hallucination”[7].
Sometimes ChatGPT provides responses that are accurate but are illogical or erroneous. Additionally, ChatGPT is sensitive to repeated attempts at the same question, and the input phrase can be changed. The model regularly makes use of superfluous words and phrases, such as saying it is a language model created by OpenAI repeatedly. Over-optimization issues and biases in the training data (trainers prefer longer responses that seem more detailed) are the root causes of these issues. The ideal model should ask clarifying questions when the user submits an ambiguous query. Instead, the majority of the time, our current models assume what the user means.
Though ChatGPT has the potential to save time and money through the automation of tedious tasks and by using a revolutionized natural language processing, it is important to be aware of the risks and take precautions to reduce them because it has the potential to produce biased results, raise ethical questions, and be misused. Because of this, several nations, including Italy, have outlawed the use of this AI [8]. To guarantee that ChatGPT is used responsibly and that everyone benefits from it, this must be done. The professions that are most likely to be affected are those that need routine and repetitive work, although these models have demonstrated amazing skills [9]. It is doubtful that these models will be able to replace persons or individuals in all tasks and circumstances completely. Even though emotion human, intuition, intelligence, and creativity are crucial for many tasks, machines may find it difficult to replicate these traits. However, with human input and direction, machines may be able to develop entirely original concepts [10].
References
[1] Anand, S. (2022). Success story of openai- the makers of the CHATGPT. StartupTalky. Retrieved from https: //startuptalky.com/openai-success-story/.
[2] Kirmani, A. R. (2022). Artificial intelligence-enabled science poetry. ACS Energy Letters, 8, 574-576. [3] Abrusini. (2023). ChatGPT: A Brief Introduction and Considerations for Academic Integrity. The Innovative Instructor. Retrieved from https://ii.library.jhu.edu/2023/01/30/chatgpt-a-brief-introduction-and-considerati ons-for-academic-integrity/
[4] Marr B. (2023). CHATGPT: What are hallucinations and why are they a problem for AI systems. Retrieved from https://bernardmarr.com/chatgpt-what-are-hallucinations-and-why-are-they-a-problem-for-ai-systems/
[5] Hao, K. (2020). OpenAI has released the largest version yet of its fake-news-spewing AI. MIT Technology Review. Retrieved from https://www.technologyreview.com/2019/08/29/133218
[6] OpenAI November 30, 2022 Introducing ChatGPT. Retrieved from https://openai.com/blog/chatgpt
[7] Heaven, W. D. (2023). GPT-4 is bigger and better than chatgpt-but Openai won’t say why. MIT Technology Review. Retrieved from https://www.technologyreview.com/2023/03/14/1069823
[8] Rodriguez, C. (2023). Which countries will follow aer nation’s shock chatgpt ban? Forbes Australia. Retrieved from https://www.forbes.com.au/news/innovation/chatgpt-ban-which-countries-will-follow-italy-in-block ing-ai-giant/
[9] McCallum, S. (2023). CHATGPT banned in Italy over privacy concerns. BBC News. Retrieved from https://www. bbc.com/news/technology-65139406
[10] Bahrini, Aram, et al (2023). “ChatGPT: Applications, Opportunities, and Threats.” arXiv preprint arXiv:2304.09103
