ChatGPT

Figure 1: ChatGPT user interface. Source: OpenAI.

Developed by the AI research lab OpenAI, ChatGPT is a large language model (LLM) that generates text based on an input (prompt) from a user (figure 1). It interacts in a conversational way, answering follow-up questions, admitting to its mistakes, refusing inappropriate requests, and challenging incorrect premises. The LLM was trained on a massive amount of information and it's a derivative of InstructGPT, a program trained to follow natural language instructions in a prompt and give an elaborate response.

OpenAI's ChatGPT was introduced at the end of November 2022 during a public demo. In the following days, it was already considered the best artificial intelligence (AI) chatbot ever released to the general public. Its initial introduction was, according to the AI research lab, "to get users’ feedback and learn about its strengths and weaknesses." They are encouraged to report on problematic model outputs through the UI and false positives and negatives. During this phase, ChatGPT could be used for free at chat.openai.com.

The technology behind OpenAI's chatbot is based on an upgraded version of the GPT-3, GPT-3.5 (Generative Pre-Trained Transformer) . A new model, text-davinci-003, was introduced that is said to be able to handle more complex instructions and better long-form content production. GPT-3's successor is expected to arrive in 2023. The current model of iteration was created with ease of use in mind.

Reception

A few days after the release of ChatGPT, more than a million people had already signed up to test it, according to Greg Brockman, OpenAI's president. Screenshots of users' conversations with the AI model went viral on Twitter with the majority being astonished by the software's text production capabilities. It also captured the attention of tech leaders like Box CEO, Aaron Levie.

The immediate impact of the chatbot could have resulted from it being the first time that a powerful tool was made available to the general public for free and with a simple user interface.

Users had a diverse range of requests to ChatGPT, with some asking the AI to rewrite the 90s song, "Baby Got Back" in the style of "The Canterbury Tales", asking the answer to an AP English exam question, and asking for fairy-tale inspired home decor tips.

Features

Figure 2: General overview of the training process using reinforcement learning from human feedback. Source: OpenAI.

The ChatGPT model was trained using Reinforcement Learning from Human Feedback (RLHF) following the same methods as InstructGPT (figure 2); only with small differences in the data collection setup. Training for an initial model was made using supervised fine-tuning where human AI trainers had "conversations in which they played both sides—the user and an AI assistant."

Reinforcement learning uses a reward model for AI training. This was done by collecting comparison data consisting of two or more model responses ranked by quality. According to the official blog of OpenAI, "to collect this data, we took conversations that AI trainers had with the chatbot. We randomly selected a model-written message, sampled several alternative completions, and had AI trainers rank them. Using these reward models, we can fine-tune the model using Proximal Policy Optimization. We performed several iterations of this process."

GPT-3.5 was trained on a mix of text and code published before Q4, 2021. Both ChatGPT (a fine-tuned version of a model in the GPT-3.5 series) and GPT-3.5 were trained on an Azure AI supercomputing infrastructure. The model on which the chatbot is based, text-davinci-003, can handle more complex instructions with increased output quality and overall better in long-form writing (around 65% longer outputs than text-davinci-002). It also has fewer limitations (e.g. a reduction in "hallucinations") than previous versions and scores higher on human preference rating.

While its output is not perfect, ChatGPT has several distinctive features:

• An LLM capable of generating human-like text in response to a prompt.
• It answers and dialogs with users in a relevant way to the topic and situation at the time.
• Creates text in different formats and styles.
• Provides information on a wide range of topics.
• Can be embedded in a variety of applications (e.g. chatbots and virtual assistants).

Applications

OpenAI's ChatGPT has myriad applications, from generating computer code, to writing jokes, explaining scientific concepts in different levels of complexity, to writing college-level essays. Besides its regular text composition skills, it can also provide assistance, by helping programmers spot and fix errors in code or answer questions made to it, analogous to a Google search. Indeed, it was denominated as a possible "Google Killer". Due to ChatGPT remembering what a user has written before, some have suggested that it could be possible to create personalized therapy bots.

Overall, the potential of this application is enormous, possibly having an impact on different business sectors.

• AI prompt generation, for art creation in models like DALL-E 2 or Midjourney.
• Act like a virtual cloud, stringing together cloud services to achieve complex tasks.
• Content generation, changing how content is created for thousands of companies.
• Possible impact on education tech by answering to doubts.
• Sort, manage, and organize unstructured data.
• Generate SQL queries.
• Code writing.

Limitations

ChatGPT has several limitations. These range from writing incorrect or nonsensical answers to being excessively verbose. Such issues arise from the training data and also biases in it.

Dev.to compiled a list of the tool's limitations:

• May not provide accurate or relevant responses.
• May not have enough knowledge or information on particular topics, affecting the quality of responses.
• May have difficulty in understanding and interpreting complex or abstract concepts.
• It can have difficulties in understanding and responding to natural language.
• Inconsistancy in responses due to not being able to adapt to changing contexts or new information.
• May not generate original or creative answers.
• May not engage in a dynamic conversation.
• Its performance may decrease over long periods of time, resulting in lesser quality responses.
• Large volumes of input or multiple conversations at the same time may lead to delays or errors.
• Misunderstandings or bias due to the language model not being able to understand users with different backgrounds, experiences, or language abilities.

Implications

The impact of such AI models is potentially enormous. Like other content-generation applications (e.g. DALL-E and Midjourney) it affects the human element that, until now, has been responsible for the production of text or art. It could mean the end of white-collar knowledge work. Also, copyright issues are another problem that will arise. In education, homework and essays will not be a reliable way to evaluate students. AI-assisted tools keep evolving and will provide an easier way to plagiarize. In the future, tools like OpenAI's ChatGPT could be used as propaganda machines, flooding the public with authoritative-sounding information.