Several basic facts about ChatGPT :
- The official website for ChatGPT is: https://chat.openai.com/chat . This interface is a very simple dialogue tool where you can input any question you want to chat about or inquire about;
Currently, there is no official ChatGPT application. Any app, computer client, or other website claiming to be related to ChatGPT may be developed based on OpenAI's API interface or may be fake. Please be discerning;
-
ChatGPT was developed by OpenAI, and to use it, you need to have an account on the OpenAI website. To register an account, you need to meet two conditions: First, you need to access their website
https://openai.com/ through Science Internet 🪜. Second, during the registration process, you need to use a non-mainland China phone number to receive a verification code to complete the confirmation;
-
ChatGPT belongs to OpenAI, a company founded in 2015. Recently, Microsoft invested billions of dollars to become a major shareholder. Therefore, you may see a lot of information about Microsoft and this product on the internet, such as their office intelligent assistant CoPilot released on March 16th.
Now, let's get to the main topic of what ChatGPT really is and how it was created.
Ⅰ. Introduction to ChatGPT
First, ChatGPT is a natural language processing (NLP) model developed by OpenAI, based on the GPT-3 and subsequent architectures, and specifically optimized for interactive dialogue with human users. We can break it down as follows:
Chat: Chat, meaning the model is mainly used for interactive dialogue with human users.
GPT: Stands for "Generative Pre-trained Transformer", which is the core architecture of the model.
-
G (Generative): Generative, indicating that the model can generate new content, such as text, answers, etc.
-
P (Pre-trained): Pre-trained, meaning the model has been pre-trained on a large amount of text data, enabling it to understand and generate natural language.
-
T (Transformer): Transformer, a neural network architecture used for natural language processing tasks. It was proposed in a 2017 paper and applied by OpenAI to the GPT model. This architecture has powerful parallel processing and attention mechanisms.
In summary, ChatGPT is a generative pre-trained transformer model optimized for conversing with humans. Currently, Google's BERT model also uses the Transformer architecture, which is the most mainstream model architecture in the NLP field and the basis for large-scale model training.
Ⅱ. Development History of GPT Models
-
GPT: Released in June 2018, it was OpenAI's first model based on the Transformer architecture, adopting generative and pre-training methods. It achieved significant success in natural language understanding and generation tasks.
-
GPT-2: Released in February 2019, the model parameters increased from 120 million in the previous generation to 1.5 billion. This enabled GPT-2 to achieve significant performance improvements in various NLP tasks, such as reading comprehension, machine translation, and summary generation.
-
GPT-3: Released in June 2020, the third version has over 175 billion parameters, a 1000-fold increase from the first generation. At this point, the GPT model performed excellently on various natural language processing tasks and could generate highly persuasive text. An important feature of GPT-3 is its ability to perform zero-shot learning and few-shot learning by adjusting input and output formats without explicit fine-tuning. Because of this, GPT gained the ability to start conversations with humans from scratch, and the first version of ChatGPT was based on GPT-3.
-
GPT-3.5: Released in November 2022, it is an improved version of GPT-3, with model parameters reaching 200 billion. This version of GPT used reinforcement learning from human feedback (RLHF) and improved on various natural language processing tasks, especially in conversation, where it can generate more natural, fluent, and interesting text.
-
GPT-4: Released in February 2023, it is currently the most advanced natural language generation model with over 500 billion parameters. The core new capability of this version of GPT is support for multimodal input and output (such as images, audio, video, etc.), which is a revolutionary change, equivalent to ChatGPT being able to "hear" and "see" the real world after modification.
Ⅲ. Brief Description of ChatGPT's Application Scenarios
In fact, anything related to "semantic understanding" and "language generation" can be largely solved by ChatGPT:
"Semantic Understanding": Such as the simplest translation (requiring understanding of word meanings), intelligent customer service (requiring understanding of customer requests), article summarization, meeting minutes, etc.
"Language Generation": This refers to when your result requires it to help you generate a piece of text or code, which it can handle very well, such as answering questions, writing articles, writing code for a webpage, etc.
However, as we mentioned earlier, GPT-4 has multimodal capabilities, it can also understand and generate video, images, audio, etc. This is essentially because when these contents are stored in computers, they are just 0s and 1s, so it can deconstruct these contents into something similar to language to achieve corresponding capabilities.
The OpenAI official website also provides some references for what they can do, which you can check out.