ChatGPT is a language model chatbot developed by OpenAI based on GPT-3.5. ChatGPT, named after OpenAI's Generative Pre-trained Transformer (GPT) models, provides answers to user questions and allows users to interact in a conversational way with follow-up questions and responses. The language model produces an output for a user's request with the ability to write code, poems, songs, essays, stories (inspired by a specific author), and more, generating material to complete a task rather than only being a source of information. ChatGPT is capable of admitting its mistakes, challenging incorrect premises, and rejecting requests considered inappropriate. The chatbot is similar to another OpenAI model, InstructGPT, which is trained to follow instructions and provide a detailed response.
ChatGPT is based on OpenAI's GPT-3.5, a series of models trained on a blend of text and code from before the end of 2021. Large language models such as GPT-3.5 are trained on massive amounts of data from across the internet (including text and code) and are able to understand what a question is asking and provide human-like responses. OpenAI's GPT-3 model was trained on 570 gigabytes of text and has over 175 billion parameters, a significant increase compared to its predecessor GPT-2 with 1.5 billion parameters. ChatGPT adds an additional layer of training to GPT-3.5 models called Reinforcement Learning with Human Feedback (RLHF). This training uses human feedback to improve the model, allowing ChatGPT to better follow directions and generate superior responses.
OpenAI plans an iterative deployment of ChatGPT, with a research preview of the model released on November 30, 2022. During the research release, ChatGPT is freely available online to anyone with an OpenAI account. Within the first five days of ChatGPT's public release, over a million users registered to use the chatbot. OpenAI's iterative deployment integrates RLHF to increase the safe use of its AI systems while reducing harmful or untruthful outputs. During the research preview, OpenAI will acquire user feedback on the strengths and weaknesses of ChatGPT. The UI allows users to provide feedback on problematic outputs and false positives and negatives from the external content filter.
Reports in January 2023 suggest Microsoft is planning to launch a new version of its search engine, Bing, which incorporates ChatGPT to answer queries. This was followed by reports ChatGPT could be integrated into Microsoft Office applications, such as Word, PowerPoint, and Outlook, so users can automatically generate text from simple prompts. Microsoft has already incorporated OpenAI's GPT language model to improve Outlook search results.
A UBS Study published on February 1, 2023, estimated ChatGPT reached 100 million monthly active users in January. Reaching this figure only two months after its launch makes ChatGPT the fastest-growing consumer application in history. For comparison, it took TikTok nine months and Instagram 2.5 years to reach 100 million monthly active users. The study found an average of 13 million unique visitors used ChatGPT each day in January 2023, doubling its numbers from December 2022.
On January 11, 2023, OpenAI's official Discord shared a waitlist for an experimental paid version of the language model provisionally called ChatGPT Professional. The new service will offer improved availability and faster responses (without throttling) and give users at least twice as many the number of daily answers as the free version. The waiting list for the ChatGPT professional pilot is open to anyone and comes with a questionnaire about how the user currently uses the chatbot and what price they would consider fair.
On February 1, 2023, OpenAI launched a pilot subscription plan in the United States called ChatGPT Plus. Available for $20 per month, ChatGPT plus offers subscribers benefits that include the following:
- Access to ChatGPT, even during peak times
- Faster response times
- Priority access to new features and improvements
OpenAI will begin the process of inviting people from their waiting list with plans to expand access to additional countries. The company plans to refine and expand ChatGPT Plus based on subscriber feedback and explore options for lower-cost plans, business plans, and data packs for improved availability.
On May 18, 2023, OpenAI launched its ChatGPT app for iOS. The app is free, and it syncs history across devices and integrates OpenAI's open source speech recognition system Whisper.
ChatGPT has received criticism from some, arguing it could produce harmful outputs, lead to job losses, and be used for plagiarism, with particular concerns that students could use the model to produce their work. The International Conference on Machine Learning (ICML) announced it has banned authors from using AI tools, such as ChatGPT, to write scientific papers. In January 2023, research showed ChatGPT could pass the final exam for the University of Pennsylvania’s Wharton School MBA program. In response, OpenAI researcher Scott Aaronson stated the company is working on a system for “statistically watermarking the outputs” of ChatGPT to prevent cheating in academia and the production of propaganda. In a lecture at the University of Texas, Aaronson stated:
We want it to be much harder to take a GPT output and pass it off as if it came from a human... This could be helpful for preventing academic plagiarism, obviously, but also, for example, mass generation of propaganda – you know, spamming every blog with seemingly on-topic comments supporting Russia’s invasion of Ukraine without even a building full of trolls in Moscow. Or impersonating someone’s writing style in order to incriminate them.
In an interview given to StrictlyVC, OpenAI CEO Sam Altman stated the company will develop methods to help schools identify AI plagiarism. However, he warned that full detection could not be guaranteed. Altman added that society has integrated new technologies into the classroom before and that those technologies will only generate more positive impact for users down the line.
We're going to try and do some things in the short term. There may be ways we can help teachers be a little more likely to detect output of a GPT-like system. But honestly, a determined person will get around them... Generative text is something we all need to adapt to. We adapted to calculators and changed what we tested for in math class, I imagine. This is a more extreme version of that, no doubt, but also the benefits of it are more extreme, as well.
On January 31, 2023, OpenAI released a new AI classifier trained to distinguish between human-written and AI-written text. While the classifier works for text written by a variety of AI models, it is not 100% reliable. OpenAI made the classifier publicly available to get feedback. After criticism that ChatGPT has a left-wing bias, OpenAI shared information on how the model produces its responses in February 2023.
ChatGPT is a fine-tuned version of models in the GPT-3.5 series, trained using RLHF—the same methods as OpenAI's InstructGPT with only slight differences in data collection. An initial model utilized supervised fine-tuning with human AI trainers providing conversations in which they played both sides, the user and the AI assistant. Human trainers were given access to model-written suggestions in order to compose responses. This dataset was mixed with the InstructGPT dataset and transformed into a dialogue format.
To build a reward model for reinforcement learning, conversations between human trainers and the chatbot were collected as comparison data (the ability to rank two or more responses ranked by quality). Model-written messages were randomly selected, sampling several alternative completions, and trainers ranked their quality. This information was fed back into the model using proximal policy optimization. The process was repeated several iterations to improve performance. OpenAI researchers published a pair of papers on incorporating human feedback into large language models:
- Learning to summarize from human feedback, published in February 2022, describes a method of datasets for human comparisons to train a language model.
- Training language models to follow instructions with human feedback, published in March 2022, describes a method for incorporating human feedback into large language models to align their outputs with user intent.
ChatGPT continues to have limitations, producing incorrect or nonsensical answers. The model is sensitive to changes in input phrasing or repeating the same prompt multiple times. For example, ChatGPT may claim to not know the answer to a question or understand the input but answer correctly when the input is slightly rephrased. The model can overuse certain phrases, due to biases in the training data, such as trainers preferring specific responses. Additionally, the model can sometimes respond to harmful instructions or demonstrate biased behavior. OpenAI plans to regularly update the model to improve these areas.
Learning to summarize from human feedback
Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, Paul Christiano
February 15, 2020
Training language models to follow instructions with human feedback
Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, Ryan Lowe
March 4, 2022