ChatGPT: AI Genie for You
(Before reading this article, I urge you to read the prerequisite articles on Artificial Intelligence part 1, part 2 and the Machine learning article)
The story of Aladdin and his genie who gets out of a magic lamp and creates miracles in the life of his master, was one of the most intriguing stories of Arabian Nights. I’m sure you must have wished for a similar magical creature to serve you, right? Even I wanted the blue spirit as a playmate, a friend and most of all – as a loyal servant for whom my wish would be the command! But alas, that desire remained unfulfilled.
But not anymore. ChatGPT, a new text-based interactive computer program can act as an artificial intelligence (AI) – powered genie for you by answering your questions, writing poems and generating ideas for you. With the help of ChatGPT, you can achieve a phenomenal rise in your productivity. It acts as your own personal assistant in writing, researching, coding, brainstorming, etc. Sounds like a Genie, right?!!
The WH questions – what is it, how does it work and why is it important?
ChatGPT explains itself as – “a large-scale language model that uses deep learning techniques to generate human-like text. The input to the model is a sequence of words, and the output is a predicted next word. This prediction is based on the probability distribution over the vocabulary, which is computed by the model using the input sequence and its internal parameters.”
Thus, ChatGPT is an artificial intelligence-based computer program produced by OpenAI, a California based AI company, which is currently in the prototyping stage. GPT expands to Generative Pre-trained Transformer. In order to understand this disruptive technology in depth, we shall proceed in a step-by-step manner.
In the earlier paragraph, we have seen a glimpse of all that ChatGPT can do. Before going further let us first understand some of the terms that we require for understanding the mechanisms behind the workings of ChatGPT.
1. Generative AI – As opposed to a predictive or analytical artificial intelligence system, Generative AI writes articulate text, generates ideas, paints, creates poems and song lyrics as well as manipulates existing information to create newer knowledge.
2. Pre-trained model – As you might have read in the previous article on Machine Learning, a machine is taught using models. These models are trained using a vast amount of data, which enables the model to make a prediction when encountered with similar situations. Larger is the dataset used for training, better is the accuracy of the model. Hence ChatGPT has been trained on a humongous amount of data ranging from information on the internet to advanced scientific concepts such as quantum physics from books, articles, etc. This pool of knowledge gives the platform an ability to answer any question.
3. Transformer – It is a type of deep-learning model used in Natural Language Processing (NLP) and computer vision (CV). NLP is a specialized branch of machine learning that focuses on interpretation and generation of human language by the machines.
ChatGPT uses OpenAI’s GPT-3.5 software engine, hence differs from its predecessors in the fact that it can converse like a human being and not throw an illogical collection of words at you. The structure, format and vocabulary of the chatbot’s responses highly match with those of a human. For the uninitiated in the world of natural language processing, this is a historically significant milestone in the field of AI and its subsets of Machine Learning and NLP. This complexity in making a machine understand human speech, stems from the theoretically infinite languages, their multiple accents and pronunciations, homonyms as well as the nuanced approaches to speech such as context and sarcasm. Hence, the success of ChatGPT is being hailed as a stepping stone to many more advancements in the field of NLP and AI.
How was ChatGPT trained?
To understand in depth how the AI – enabled chatbot works, it is imperative to understand how it was trained. The process of training the model is ongoing, with the feedback of users and the responses generated to their queries, being used as a training dataset to improve accuracy and eliminate inherent bias. ChatGPT was trained using supervised, transfer and reinforcement learning techniques. Let us briefly understand these approaches to machine learning (ML) –
1) Supervised Learning –
This approach to creating AI, focuses on training the machine to recognize the relationship between the input and the label of the dataset. This label provides more information about the input data, such as its category. There exists an underlying pattern between the input data and the label, which then is used to generate output for a new set of similar data. For example, to train a model to recognize whether a food item is sweet or savoury, the list of food items is the input given to the ML model and savoury or sweet is the label. The machine, through an analysis of the ingredients and the labels, determines that the food items having higher percentage of sugar are sweet while those containing salt as the chief seasoning element are savoury.
Thus, in case of ChatGPT, the programmers train the chatbot using ideal responses to a question. So, the creators of ChatGPT reprise both the roles – that of the user of ChatGPT and also that of an assistant to the AI bot, who is responsible for generating response to prompts.
2) Reinforcement Learning –
This is a rewards-based approach to training a machine. It borrows from the analogous concept in psychology, with the same name. The human trainers evaluate and rate the responses generated by ChatGPT to specific prompts. These rankings are then used to fine tune the model by encouraging the chatbot to generate responses similar to those that were rewarded with a high rank.
3) Transfer learning (TL) –
It is a method of applying the knowledge acquired while solving one problem to solve a similar problem. Through TL, ChatGPT replies to queries that are new but similar to the ones it is trained for. This approach is inspired by the similar approach taken by human brain to devise solutions to novel problems by dipping into its pool of knowledge acquired through previous experiences.
Limitations of chat GPT
1. Misleading Information – Despite drastic improvement in detection of fake news by GPT-3.5 engine over that of its predecessors, ChatGPT’s answers are prone to misleading information. StackOverflow, a website for computer code, banned ChatGPT generated code because it was found incorrect.
2. Hateful, misogynistic & racist content – Because of a learning bias, known as Inductive Bias, encountered by the model during training phase, some of the responses of ChatGPT have been red-flagged for their insensitivity towards gender, minorities and other marginalized communities.
3. Role in academic or professional cheating – According to an article by New York Times, universities have begun changing the nature of exams and the evaluation criteria, in order to counter students using ChatGPT to generate assignments and mark answers in multiple-choice questions. Cases involving cheating in a coding interview by making ChatGPT write code have also been seen.
4. Increased threat of cyber-attacks – According to an assessment by World Economic Forum (WEF), the threat to cybersecurity of systems is slated to rise, given the ease with which ChatGPT can generate malicious programs and files. Cyber-attacks shall no longer need expert hackers. Hence, WEF projects an increased investment and high demand for jobs in the cybersecurity domain.
After a lot of discussion in the media regarding limitations of ChatGPT and its possible use to create mischief, OpenAI has released a statement informing modifications to ChatGPT that shall curb generation of biased and hateful information, along with an increased awareness regarding the possible harmful end-use of the code generated by it. The company also aims to flag the responses of ChatGPT in some way so that they can be recognized in case of cheating. These measures, the company claims, will help make ChatGPT a benign and helpful tool, as intended by its creators.
The rapid rise of ChatGPT
While it took Netflix 3.5 years and Instagram 2.5 months to reach the mark of 1 million, ChatGPT created a record by reaching a million users in just 5 days. This was achieved because of the tremendous popularity enjoyed by ChatGPT amongst the masses.
The global market of AI was valued at 1.36 billion USD in 2022, with a compound annual growth rate of 38% from 2022-2030. With the disruption caused by ChatGPT, these numbers will only rise. As the cyber-threat goes up, so does the demand for cybersecurity professionals. In the coming years, jobs for ML & AI engineers and data scientists will see an upward trend. Hence, upskilling yourself in these fields will reward you handsomely in the long run. The humanity is at a critical juncture in the history of technology. Weak AI programs like ChatGPT have tremendous potential to simplify our lives, provided, we use them wisely, because there is an old adage that goes – technology is not born moral or corrupt; its users determine its nature.
Vishvali Deo is an E&TC (Electronics and Telecommunication) Engineer by education and Software Engineer by Profession. She believes that 'Technology is a Great Democratising and Equalising Force' and hence is on a mission to make the general public understand seemingly complex technologies in a simple manner.
She is convinced that the root of today's world problems lie in the past, hence she has also pursued post-graduation in History. She has a keen interest and a good grip over Economics, Political Science and Environmental Engineering. She has a penchant for working with Women and spreading Digital Literacy amongst them, with the aim of their empowerment. She also strives to provide Free Quality Education to children and counsels young adults. Besides, she is also skilled at Public Speaking, having won many awards in Elocution & Debate Competitions and Technical Paper Presentations.