ChatGPT is a chatbot that uses artificial intelligence to communicate with users through text-based messages and images. It uses modern machine learning technology to generate responses that sound natural and are meant to be relevant to the conversation.
The chatbot was trained in several phases.
- It is based on the GPT-3.5 language model (GPT stands for Generative Pre-trained Transformer), an improved version of GPT-3, which also comes from OpenAI. GPT is based on Transformers, a machine learning model presented by Google Brain, and has been trained through self-supervised learning. The quantity and variety of texts serve to enable the system to recognize and distinguish between the linguistic patterns of the texts: for example, that poems consist of lines or that technical texts often contain special terms. This step is also referred to as pre-training, as a model was created that could not yet perform the desired task, but the model parameters obtained provided favorable starting conditions for further fine-tuning.
- Afterwards, the language model was fine-tuned for the actual task through supervised learning: the generation of answers to previously asked questions. Ready-made answers have been provided for this. From the beginning of the training, the generated answers were evaluated by test subjects. If they were judged to be wrong or outlandish, the system tried out new settings or parameters. In the chatbot’s initial training, testers preferred longer responses, regardless of actual understanding or content, resulting in long responses from the chatbot.
- In the final step, the model was further optimized through reinforcement learning through human-influenced feedback (RLHF): For this purpose, another model (reward model) was first trained with supervised learning, which was taught to qualitatively evaluate ChatGPT’s answers in the form of a ranking by people. This additional enhancement via RLHF is used for the first time in version 4. Finally, the Proximal Policy Optimization algorithm was used, which used the reward model as the reward function to be maximized.
Based on their experience developing GPT and Codex, ChatGPT has been provided with protections designed to avoid incorrect or harmful answers. Nevertheless, the website points out that the current version is a publicly available research preview and that the chatbot may occasionally generate incorrect information.
The large number of parameters in the system’s calculations no longer allows the developers to fully understand the processes. “That’s why an AI calculation is often referred to as a black box – only input and output are recognizable from the outside.”
---
Role of ChatGPT in Chatbots
ChatGPT offers the opportunity for dialogical exchange. The quality of the answers fluctuates. According to the AI experts, it can express itself “brilliantly one moment and breathtakingly stupid the next.” ChatGPT can write texts in the style of business plans or school homework, among other things. A wide range of options and challenges arise in science communication when ChatGPT is used to summarize and explain knowledge.
The advanced communication capabilities (multimodal) with combinations of voice, image and/or text input and responses has been tested by many developers. Image input can also recognize text in the image, among other things. For example, if a tester photographs the front page of a newspaper and ask the chatbot for summaries of the articles, there will be satisfactory result. The results will not be satisfactory when entering a crossword puzzle and ordering it to be solved or writing step-by-step instructions from a graphic representation for assembling a piece of furniture. Examples of applications were that you could show a defective bicycle part and ask how it could be repaired, or that you could present a photo of the food in the refrigerator and ask what to cook with it.
Open AI has deliberately largely blocked people’s image input to prevent misuse. Compared to voice assistants such as Siri or Alexa, the voice responses of five female and male speaker variants sound more natural and allow longer dialogues. However, longer voice outputs are still time-consuming and interrupted by pauses.
ChatGPT can analyze and write program code in various programming languages. It could also be used to improve, annotate and detect errors in software code. For the time being, the use of chatbots will facilitate programming, especially as an aid. The fact that these tools completely replace programmers is not yet up-to-date. Hints are given on how programmers can appropriately contribute to the new environment.
Since ChatGPT was initially only trained with data up to September 2021, the software sometimes responds to inquiries about current events with the note, “As an AI model, I do not have direct access to real-time information.”
ChatGPT allows plug-ins to extend the functionality of this software, for example by interacting with application programming interfaces (APIs) from other software and service providers to retrieve real-time information, integrate company databases, perform certain calculations, or act on behalf of the user. For example, there are also plugins for travel planning, including personalized recommendations.
