What is ChatGPT?

Recently, the non-profit artificial intelligence research group OpenAI has launched an AI chatbot called ChatGPT, which can chat with users and answer a range of questions. In less than a week, it has already gained over one million users. As the model is still in the research preview stage, it is offering free trials to users, but it will not be permanently free.

So what exactly is this ChatGPT that has taken the internet by storm? How does it work? As a model capable of intelligent conversation, can it replace Google?

The launch of ChatGPT has caused a sensation and it has already gained over 100 million users.

On November 30, 2022, the non-profit artificial intelligence research group OpenAI announced the launch of a new type of chatbot model called ChatGPT, which has been welcomed by many users. As the model was initially in the research preview stage, it was offered to users for free trial. Currently, users can register and start interacting with ChatGPT for free, and it even supports multiple languages.

According to Bloomberg, ChatGPT has caused a sensation on the internet after its open trial. Users have shared detailed answers provided by ChatGPT on Twitter, from resolving historical controversies to composing poetry for cryptocurrencies, and were surprised by its capabilities. One user even shared their testing experience and claimed that ChatGPT could easily complete a 1000-word university paper.

In addition, Sam Altman, the CEO of OpenAI, revealed on December 5, 2022, that “since ChatGPT was launched last Wednesday, it has already gained over 1 million users.”

A report by UBS Group showed that just two months after the release of ChatGPT, its monthly active users had surpassed 100 million by the end of January 2023, making it the fastest-growing consumer application in history. It’s worth noting that TikTok took 9 months to reach 100 million users, while Instagram took 2 and a half years.

What is ChatGPT and what makes it special?

So, what exactly is the ChatGPT and why has it become an overnight sensation on the internet?

Before understanding what ChatGPT is, let’s first briefly introduce its company.

Introduction to OpenAI

Headquartered in San Francisco, OpenAI was founded seven years ago by Tesla CEO Elon Musk, Sam Altman, and other investors with the goal of developing artificial intelligence technology for the betterment of humanity. However, Musk left the company in 2018 due to differences in the direction of its development.

Previously, OpenAI gained fame for its release of the GPT-3 natural language processing model, which allows AI to intelligently respond to text. The currently released ChatGPT may be a prelude to the official release of OpenAI’s GPT-4.

Introduction to ChatGPT Chatbot

ChatGPT is a language model focused on conversation generation. It can generate corresponding responses based on user input. By learning a large amount of existing text, ChatGPT can simulate human conversation and has made significant progress in natural language understanding and generation.

After registering with ChatGPT, users can ask it to answer questions such as “Who was the US president in 1955?”, “What is the meaning of life?”, “What should I wear if it’s 4.4 degrees Celsius outside today?” Regarding clothing, ChatGPT responds that it depends on your activity. If you’re going out, you can wear a light jacket or sweater, pants, and shoes that don’t expose your toes. If you’re staying indoors, wear a T-shirt, jeans, or other comfortable clothes.

Compared to traditional AI text generation based solely on regular data from databases, ChatGPT’s weaknesses such as simple identification and prediction of given text content (parroting), and outdated databases are addressed. ChatGPT is adapted from OpenAI’s GPT-3.5 model, which, in addition to being fluent in conversation, can also remember the user’s previous conversation content and is more human-like. It has a certain understanding of events that have occurred since 2021 and can be said to be the most communicative AI to date, handling user natural language in a more humanized way.

However, ChatGPT refuses to answer many potentially harmful questions (such as those related to hate speech, violent content, or how to make bombs), as these answers would go against its “programming and purpose.” OpenAI achieves this by pre-setting all input with special prompts and using a technique called “reinforcement learning from human feedback” (RLHF) to fine-tune the AI model based on human scoring of its generated responses.

In addition to conversation, netizens around the world are trying to find potential uses for ChatGPT, such as writing poetry, correcting coding errors with detailed examples, writing new code, and expounding on the philosophy of hot dogs as sandwiches.

Official ChatGPT App Launch

After the release of ChatGPT, many people were disappointed that it did not have a mobile version and could only be used on the web or through third-party apps that connect to GPT-3/GPT3.5.

On May 19, 2023, the official ChatGPT mobile version was officially launched, offering a choice between the GPT-3.5 and the latest and most powerful GPT-4 models, allowing users to chat with AI to their heart’s content.

ChatGPT APP will be free to use, initially only available in the United States and certain markets. Now it has been launched in 152 countries. According to OpenAI, the launch of the iOS app is the first step in transforming their most advanced research into a practical tool, and they will continue to lower the barriers to use in the future.

What is the technology background behind ChatGPT and what is the GPT-3 model?

The technology behind ChatGPT is the GPT-3 model. Essentially, GPT-3 is a simple statistical language model. From a machine learning perspective, a language model is a modeling of the probability distribution of word sequences, using already spoken segments as conditions to predict the probability distribution of different words appearing at the next moment. On the one hand, a language model can measure the degree to which a sentence conforms to language grammar (e.g., whether the response generated by a human-machine dialogue system is natural and fluent), and on the other hand, it can also be used to predict the generation of new sentences. For example, for a segment “It’s 12 o’clock noon, let’s go to the restaurant together,” a language model can predict the possible words that may appear after “restaurant.” A common language model would predict that the next word is “eating,” while a powerful one can capture temporal information and predict contextually relevant words like “having lunch.”

Typically, the strength of a language model depends mainly on two factors: first, whether the model can use all the historical contextual information. In the above example, if the semantic information “12 noon” cannot be captured, the language model can barely predict the next word “having lunch.” Second, it depends on whether there is enough rich historical context available for the model to learn, i.e., whether the training corpus is rich enough. Since the language model belongs to self-supervised learning, the optimization goal is to maximize the language model probability of the observed text, so any text can be used as training data without annotation.

Due to its stronger performance and more obvious parameters, GPT-3 contains more topical text and is obviously better than its predecessor GPT-2. As the largest dense neural network, GPT-3 can convert web descriptions into corresponding code, imitate human narratives, create customized poetry, generate game scripts, and even imitate deceased philosophers – predicting the true meaning of life. Moreover, GPT-3 does not require fine-tuning. In handling grammatical problems, it only needs some samples of output types (a small amount of learning). It can be said that GPT-3 seems to have fulfilled all our imagination of language experts.

What are the limitations of GPT-3?

However, GPT-3 is not perfect. One of the main concerns people have with artificial intelligence today is that chatbots and text-generating tools, among others, may indiscriminately learn from all text on the internet, regardless of quality, and produce erroneous, offensive, or even aggressive language output, which will significantly affect their next applications.

OpenAI has also announced that it will release a more powerful GPT-4 in the near future. It is said that GPT-4 will be released in 2023 and will be able to pass the Turing test and be advanced enough to be indistinguishable from humans. In addition, the cost of introducing GPT-4 to businesses will also be significantly reduced.

Although the advancement of GPT-4 is promising, it is important to consider the potential risks and challenges associated with it. As developers and researchers, we must continue to work towards creating ethical and responsible AI that prioritizes the safety and well-being of society.

A more powerful GPT-4 is now online! GPT-5 is also in the works.

On March 14, 2023, OpenAI, the developer of ChatGPT, announced that the latest version of their large language model, ‘GPT-4’, is now available. According to OpenAI, the accuracy of GPT-4 is 40% higher than version 3.5. The model has already achieved ‘human-level performance’ on many professional tests, beating 90% of test-takers in the US college entrance SAT exam and the Bar examination for lawyers. Currently, GPT-4 is available for trial use.

OpenAI states that GPT-4 will be available for ChatGPT’s paying subscribers and can also be used as part of an application development interface (API) so that programmers can integrate this AI into their apps. As for GPT-4’s API, OpenAI is currently accepting waitlist forms, and users will be charged based on the number of times they call the API.

In addition, on March 26, OpenAI announced details about GPT-4.5. According to the official information, GPT-4.5 is an important intermediate version between GPT-4 and the upcoming GPT-5. GPT-5 is expected to be released in Q4 2023 and may achieve the goal of AGI (artificial general intelligence).