ChatGPT: Everything you need to know about the chatbot

Everyone has an opinion about ChatGPT. Media outlets tout its ability to generate well-written, relevant content. Engineers see it as a technological disruption, and entrepreneurs see it as a new world to explore, full of promise. All in all, the chatbot powered by artificial intelligence has been creating a buzz since its public launch in November 2022: it took only two months to reach the 100 million users mark, compared to 9 months for TikTok and 2 years for Google Maps, thus setting a new record.

This wave of curiosity seems more than legitimate, as the step forward made by the chatGPT technology is impressive. But what is really behind ChatGPT? How does this technology work? What can we do with it? And what are its limits? Here is a summary of everything you need to know on the subject.

What is ChatGPT?

ChatGPT is the name of the interface that provides access to an artificial intelligence-based language learning model, GPT-3, capable of generating text in a way that mimics human speech. In other words, ChatGPT is the name given to the application – which is nothing but an enhanced chatbot – to which users can ask questions, and from which they can get answers.

On the face of it, this is not so far from what virtual assistants already do in smartphones and smart speakers, like Siri, Google Assistant, and Alexa. Except that ChatGPT goes further thanks to its learning model, since it allows, in a way, to “chat” with the machine, as you would with a friend or a colleague. This is precisely what is amazing about this program: it becomes difficult to distinguish machine-generated text from human-generated writing.

Its main function is to generate text in response to questions asked by Internet users. In this sense, ChatGPT can produce content adapted to the user’s request, identify errors, and reject requests that it deems inappropriate, but also develop its knowledge over the course of exchanges. Below is an example of a response to a “simple” question.

This technology has brought the American artificial intelligence company OpenIA – one of the very many companies around the world working on AI-based tools – to the forefront. Based in San Francisco, OpenIA was founded in 2015 by Elon Musk and Sam Altman with the status of an association, before becoming a for-profit company in 2019, at the time of the development of GPT-2, the ancestor of the model presented in 2022.

How does ChatGPT work?

To work, ChatGPT relies on the GPT-3.5 (for Generative Pre-Training Transformer) language model. The program has been “trained” on vast amounts of data that allow it to make predictions, to “glue” words together while producing meaning. This is called a Large Language Model (LLM) because of the very large number of parameters used – 175 billion for this program.

Thanks to a deep learning system that gives it the keys to understanding the use of words in their context, it manages to give the illusion of human behavior by proposing answers that are both coherent in content and correct in form, by adapting its interventions as it is asked for clarifications, and by learning from the exchanges with users.

In essence, ChatGPT combines encyclopedic knowledge with an uncanny ability to mimic language.

However, this technology is not new, as it is based on the Transformer neural network developed by Google, which also powers the Mountain View firm’s own natural language learning model, LaMDA (found in Google Bard). In addition, other companies are working on equivalent models.

But the revolution here is more in the use than in the technique: by integrating the program with an intuitive interface, OpenIA has allowed the general public to interact directly with ChatGPT, to experience it concretely, and to push it to its limits. This an excellent way for the company to collect feedback (you can “rate” the answer given by the application) and to test its limits in real-time, by making Internet users around the world participate.

How to use Open Ai’s chatbot?

Since its public launch in November 2022, ChatGPT is free to use. Any Internet user can try it by going to this page and creating an account. It is a web application that does not require any installation.

Once on the interface, you just have to type your request in the dedicated field and click on the send button. Here is an example:

As a conversational AI, ChatGPT is not only capable of answering questions but also of remembering what you asked and bouncing off of it to offer another answer. So you can clarify your question, ask for a rephrasing, or request additional information without having to repeat the query – as in the example below.

In the language of chatbots, a request corresponds to a “prompt”, i.e. a sentence that the program interprets to be able to answer it. The artificial intelligence cuts the sentences into shorter sections that it compares to the learning model it has been trained on, and then automatically generates its response. The form of the prompt plays a significant role in the program’s usability: in short, you can get a lot out of ChatGPT, but you still have to know how to ask it!

The service is free at the time of writing, which allows OpenIA to get a maximum of feedback. At the same time, the company launched a paid plan at the beginning of February 2023 at a price of about 20 euros per month – ChatGPT Plus. Subscribers are given priority when the tool is overloaded (as frequently happens with the free version) and responses are generated more quickly.

What can you do with ChatGPT?

With 175 billion parameters, ChatGPT can be used for a wide variety of applications. Note that this learning model is limited to written language only, and cannot generate images, videos, or sounds. The requests are formulated in writing and the answers are generated in the same way.

In concrete terms, what can we do with it? Here are some examples of how ChatGPT can be used.

• Write a basic blog post that can be used as a draft for more in-depth content.

• Produce any type of textual content: email, post for social networks, list… An example with a checklist dedicated to moving.

Write a text imitating the writing style of a famous author (a feature often highlighted to emphasize the imitation capabilities of the program, capable of fetching texts from the designated author and reproducing their tone).

Summarize a document. For this example, we asked ChatGPT to summarize the article “SEO Optimization” posted on this blog and to propose a 5 key points answer.

Find topics for future content (articles, lectures, presentations, etc.). In the example below, ChatGPT has provided a list of practical tips for each topic.

Ask for a detailed plan to write a school assignment or a professional document.

Solve a complex calculation. We give an accessible example here, but it also works with more advanced equations.

Write a fictional story. No doubt that the story of the panda cyclist will be emulated!

Generate computer code in the desired language, or identify errors in an existing piece of code. You can also learn to code by asking ChatGPT-specific questions according to your level.

Overall, ChatGPT is a time-saving tool that can be used to complete tasks such as building a writing base or identifying solutions to specific problems. It is particularly relevant when it comes to code – so much so that Microsoft CEO Satya Nadella estimates that 80% of computer code will be automatically generated in the future (source). This does not prevent it from facing a number of limitations.

What are the limitations of ChatGPT?

Yes, the application is amazing. Yes, it generates structured and coherent answers. But you have to keep in mind that ChatGPT is only a computer program and that it does not “think” – in this sense, the notion of artificial intelligence is rather overused in this case. The program cannot develop reasoning, empathy, or a personal opinion. In fact, it doesn’t even really understand what it is being asked to do: it looks at the context of the terms to make sense of them and produce an answer.

It should also be noted that ChatGPT is not a “tool” in the strict sense of the word, in that it has no specific function: it does not “do” anything unique, but can, instead, serve several applications and be very versatile. It is a lever to achieve a certain goal, not an end in itself.

Moreover, its information is limited. The corpus of data it has trained on stops halfway through the year 2021, so ChatGPT is unable to answer questions about recent news – as can be seen in the example below. It can, however, learn from its conversations with users. And OpenIA plans to connect the next version of its program to the web.

Also, the responses generated by the program suffer from multiple flaws. They are generally of average quality, often sketchy or superficial, and always politically correct. ChatGPT doesn’t “wet” itself and is always careful not to upset anyone, even on questions that are not binding.

Its answers can also become far-fetched if the question is formulated ambiguously. Thus, artificial intelligence sometimes gives the impression of gluing together notions taken at random. The example below, with kangaroo eggs (spoiler: kangaroos don’t lay eggs!), is quite obvious. But on more complex topics, a misinformed user is likely to take at face value a biased, truncated, or even completely false answer. Because ChatGPT is unable to identify an error in its own response.

Finally, the massive use of ChatGPT raises ethical questions. Since the program has no concept of right or wrong, it can respond to any request without regard to its morality. That’s why OpenIA’s engineers have set up safeguards that prevent the chatbot from explaining how to make a bomb or how to poison a person. However, nothing prevents a user from having the program generate fake news or a misleading document, and spread it massively on the web – it is enough, for that, to find the right way to ask the question.

It is important to keep these limitations in mind to make ChatGPT a consistent and ethically viable use.

ChatGPT has more surprises in store. OpenIA has already scheduled the release of the fourth version of its language model (GPT-4), trained on more than 100 trillion parameters, for the first half of 2023. For its part, Microsoft has announced the program’s upcoming integration into Bing – an announcement that prompted Google to reveal its own AI-based chatbot, Bard, which could also make its way to the search engine.

Nevertheless, it’s important to keep in mind that ChatGPT, despite its impressive capabilities, is not about to surpass human skills. And that it will remain, probably for a long time, an improved chatbot, but a chatbot nonetheless.