Chatting With Chat GPT: Getting to Know OpenAI’s Latest Revolutionary Language Model
Vienna Parnell | February 20, 2023
Whether you are a software developer, student, or simply someone looking for answers, you have probably used ChatGPT, the latest craze that has taken over the Internet. With its ability to generate informative and well-written responses in a matter of seconds, the AI chatbot has impressed computer scientists and the general public alike, including myself.
I decided I wanted to learn more about this feat of technology, and who else would know more about ChatGPT than ChatGPT itself? Featuring snippets of conversation from my “interview” with the language model, I discuss the rise of the AI chatbot, briefly how it works, and its ethical implications for everyday users and university students.
*****
Launched by Open AI, an artificial intelligence research laboratory based in San Francisco and the mind behind digital image generator DALL-E, ChatGPT (Generative Pre-trained Transformer) is an advanced natural language processing model designed to respond to prompts provided by the user in a conversational manner. ChatGPT is proficient at explaining complex concepts, generating essays and fictional stories, and writing software code, among many other uses.
Like other Large Language Models (LLMs), ChatGPT determines connections between words by analyzing vast amounts of text—570 GB of data, to be precise, containing 300 billion words. The larger the input datasets, the more knowledgeable the chatbot is, and the more accurately it responds to a wide range of inquiries. Most people are already familiar with a basic version of a LLM: next token prediction, like the predictive text function on your phone. This feature works by keeping records of the user’s vocabulary and typing patterns, yet it is a far cry from the capabilities of OpenAI technologies. While chatbots have been around for decades, OpenAI stands out in its ability to emulate human conversation through machine learning techniques—but what makes ChatGPT so special?
Enter Reinforcement Learning From Human Feedback (RLHF), the novel technique that allows the chatbot to be uncannily creative, eloquent, and more “human” than its predecessors. I asked ChatGPT to provide a brief overview of this approach.
Essentially, by receiving feedback from the creator, the artificial agent is able to improve its decision-making and recognize what behavior is considered human and what is not. RLHF has been proved highly effective in conveying information in a smooth, conversational manner. This feat would not go unnoticed, as ChatGPT reached a staggering 100 million users two months after its initial release.
Even before becoming a mainstream sensation among the public, ChatGPT had been drawing attention from major investors for years behind-the-scenes. Back in 2019, Microsoft invested $1 billion into OpenAI and is currently confirmed to be continuing a multi-year partnership. With Bing only accounting for 10% of the world’s Internet searches in 2022, ChatGPT may be Microsoft’s opportunity to challenge Google’s dominance among search engines by helping them better understand user’s queries and preferences.
ChatGPT has also recently demonstrated other impressive capabilities, scoring passing grades on notoriously difficult exams: the US Medical Licensing evaluation, tests administered by the University of Minnesota Law School, and a Wharton Business School exam. These tests were composed of a combination of short-answer, multiple choice, and essay questions, and while ChatGPT did not ace them, it did perform better than many of its human counterparts.
ChatGPT also passed a Google technical interview, consisting of coding prompts suited for a Level 3 Engineer candidate with a $183K salary. While it struggled on behavioral questions aimed at evaluating leadership qualities, it excelled at generating answers to the questions (which would be more complicated than the one below).
While a remarkable achievement, the AI chatbot’s success draws attention to the increasing automation of the workforce and raises the question, “Will ChatGPT take my job?” Most sources say for now, no. In fact, it is more likely to supplement professionals rather than replace them entirely, and we might see artificial intelligence relieve humans of laborious, time-intensive computational tasks.
Job security has not been the only area of concern as of late. Students constitute a large proportion of ChatGPT’s user base, which has elicited a divided response among educators about regulating the language model. On one hand, it can help explain intricate terminology in simpler terms, but on the other hand, students may become dependent on its answers and adopt its responses as their own. At what point does using ChatGPT cross the line between being a useful learning aid to being outright plagiarism?
The New York Times recently published a series of articles addressing ChatGPT in academia and different universities’ responses against its use in coursework. Alarmed at its ability to write full-length, well-structured essays for a variety of subjects, some teachers are modifying their curriculum to remove take-home typed essays in favor of oral exams, handwritten papers, and group projects. Other instructors are trying to outsmart the chatbot by creating complex prompts that only humans could fully understand and respond to in the context of the class.
In the case that students circumvent these measures, some teachers have found hope in Princeton student Edward Tian’s GPTZero, an app that detects whether text has been generated by an AI chatbot. It works by analyzing inputted text for two measures: perplexity and burstiness. Perplexity represents how complex the sentence is, meaning the more “perplexed” the detector is, the more likely it was written by a human. Burstiness describes sentence structure, with more variation suggesting a human writer as opposed to an artificial one, which tends to write more uniformly. While GPTZero is not entirely foolproof, it is a noteworthy resource for teachers who are opposed to the use of ChatGPT.
Conversely, some teachers view ChatGPT positively, and instead of censoring its use, are integrating it into their curriculum. One such assignment involves having students study the AI-generated responses for their flaws and merits, and using it as a guide for their own work. ChatGPT might even lessen the workload of teachers by generating quizzes and developing personalized lesson plans.
Vanderbilt has yet to officially weigh in on the use of ChatGPT among its student body, leaving the monitoring of the chatbot to each individual professor. Several instructors, particularly in the engineering department, have explicitly warned against using the app for coding assignments. Other professors in more writing-intensive classes have similarly stated that consulting the chatbot for essays is not permitted. The general consensus is that most professors perceive ChatGPT as a resource that is exploited by students seeking easy shortcuts rather than proactively asking TA’s for help.
Other than class policies, there are other reasons to be wary of ChatGPT. For example, it struggles with basic conversions between number systems and math-based riddles. It also sometimes confidently asserts nonsensical answers due to its inability to distinguish between fact and fiction, as erroneous information that it has processed during training can resurface in its responses. Another main caveat is that ChatGPT is limited to 2021 data, so it is not up-to-date on recent developments. So, maybe think again before asking it to write your essay on the modern political landscape.
So, what is next for ChatGPT? In recent news, OpenAI has announced ChatGPT Plus, a $20/month plan that will be available for users to purchase in the coming weeks. This option would provide subscribers with more reliable response times, particularly during peak, high-traffic hours. Rest assured, OpenAI will still offer the current free-to-use version of ChatGPT. If this higher tier seems pricey, it is worth noting that keeping the service running is extremely expensive; some estimates report that OpenAI spends $100,000 per day, with each query costing the company nearly 1 cent. While OpenAI continues to enjoy multi-million dollar investments, a subscription-based service catering toward commercial users of ChatGPT would contribute toward reducing the financial burden of keeping the chatbot running.
Whether you are willing to spend $20 to use the premium version of ChatGPT, the service is probably here for the long run. Currently averaging 13 million unique visitors per day, the AI chatbot continues to captivate its global audience. However you decide to use ChatGPT, hopefully you are now a little more informed about its background and ethical implications.
Bibliography
Bowman E. 2023 Jan 9. A college student created an app that can tell whether AI wrote an essay. NPR. https://www.npr.org/2023/01/09/1147549845/gptzero-ai-chatgpt-edward-tian-plagiarism.
Cerullo M. 2023 Feb 1. ChatGPT user base is growing faster than TikTok. wwwcbsnewscom. https://www.cbsnews.com/news/chatgpt-chatbot-tiktok-ai-artificial-intelligence/.
ChatGPT. chatopenaicom. https://chat.openai.com/.
Clark M. 2023 Feb 1. OpenAI announces ChatGPT Plus at $20 a month. The Verge. [accessed 2023 Feb 10]. https://www.theverge.com/2023/2/1/23581561/chatgpt-plus-paid-option-20-openai-waitlist.
D’Andrea A. 2023 Jan 30. Will ChatGPT take your job? New program shows AI could be “competing” for work: experts – National | Globalnews.ca. Global News. https://globalnews.ca/news/9445651/chatgpt-job-market-artificial-intelligence/.
Dreibelbis E. 2023 Feb 1. ChatGPT Passes Google Coding Interview for Level 3 Engineer With $183K Salary. PCMAG. https://www.pcmag.com/news/chatgpt-passes-google-coding-interview-for-level-3-engineer-with-183k-salary.
From Chatter to ChatGPT: A Once Simple Tool Has Become an Industry Mainstay. 2023. JD Supra. [accessed 2023 Feb 10]. https://www.jdsupra.com/legalnews/from-chatter-to-chatgpt-a-once-simple-4524550/?.
GPTZero. gptzerome. https://gptzero.me/.
Huang K. 2023 Jan 16. Alarmed by A.I. Chatbots, Universities Start Revamping How They Teach. The New York Times. https://www.nytimes.com/2023/01/16/technology/chatgpt-artificial-intelligence-universities.html.
Hughes A. 2023 Jan 4. ChatGPT: Everything you need to know about OpenAI’s GPT-3 tool. BBC Science Focus Magazine. https://www.sciencefocus.com/future-technology/gpt-3/.
Koenders L, Könning D. 2023. What is ChatGPT capable of, and what are its limitations? – Educational Development & Training – Utrecht University. wwwuunl. [accessed 2023 Feb 10]. https://www.uu.nl/en/education/educational-development-training/knowledge-dossier/what-is-chatgpt-capable-of-and-what-are-its-limitations#:~:text=ChatGPT%20is%20a%20chatbot%20based.
Movement Q ai-Powering a PW. 2023 Jan 27. Microsoft Confirms Its $10 Billion Investment Into ChatGPT, Changing How Microsoft Competes With Google, Apple And Other Tech Giants. Forbes. [accessed 2023 Feb 10]. https://www.forbes.com/sites/qai/2023/01/27/microsoft-confirms-its-10-billion-investment-into-chatgpt-changing-how-microsoft-competes-with-google-apple-and-other-tech-giants/?sh=4515b7643624.
Network TL. 2023 Feb 2. What Students Are Saying About ChatGPT. The New York Times. https://www.nytimes.com/2023/02/02/learning/students-chatgpt.html.
OpenAI. 2019 Apr 25. OpenAI. https://openai.com/.
OpenAI’s ChatGPT Reportedly Costs $100,000 a Day to Run – CIOCoverage- Driven for Technology Leaders. CIO Coverage. [accessed 2023 Feb 10]. https://www.ciocoverage.com/openais-chatgpt-reportedly-costs-100000-a-day-to-run/.
Roose K. 2023 Jan 12. Don’t Ban ChatGPT in Schools. Teach With It. The New York Times. https://www.nytimes.com/2023/01/12/technology/chatgpt-schools-teachers.html.
Rosenberg S. 2023 Jan 24. What ChatGPT can’t do. Axios. https://www.axios.com/2023/01/24/chatgpt-errors-ai-limitations.
Ruby M. 2023 Jan 31. How ChatGPT Works: The Models Behind The Bot. Medium. https://towardsdatascience.com/how-chatgpt-works-the-models-behind-the-bot-1ce5fca96286.
Southern MG. 2023 Jan 11. OpenAI’s ChatGPT Update Brings Improved Accuracy. Search Engine Journal. https://www.searchenginejournal.com/openai-chatgpt-update/476116/.
Wilde J. 2023 Jan 26. ChatGPT passes medical, law, and business exams. Morning Brew. https://www.morningbrew.com/daily/stories/2023/01/26/chatgpt-passes-medical-law-business-exams.