ChatGPT: 5 minutes to understand this artificial intelligence that analyzes and writes (almost) like us

A few lines for a robot but a big step for Humanity.

The ChatGPT project, from the OpenAI start-up, marks a new stage in the development of artificial intelligence, this ability of algorithms to learn.

Scientific work has been carried out on conversational robots since 2017, but version 3.5 of this tool impresses as the language is natural in several languages, including French.

More than a million users have already put this "bot" to the test and therefore trained it on the way.

How does it work ?

ChatGPT (chat with pre-trained language model, in English) is a cutting-edge project launched by OpenAI - an artificial intelligence research institute created in 2015 by Elon Musk and Sam Altman - which has gone through several stages.

It is based on an artificial intelligence, called GPT3, which uses "transformers", automatic natural language processing algorithms that appeared in 2017. Equipped with 175 billion parameters, this computer system has learned by being fed for months with all the texts available on the Internet.

"It uses automatic learning, or machine learning in English, with the attention model where a sentence is understood as a whole and in context with a processing capacity of 3,500 words", analyzes Stéphane Dadian,

Read alsoCybersecurity, data mining, machine learning….

artificial intelligence from A to Z

ChatGPT defines itself as “a type of computer program designed to simulate human conversation using a pre-trained language model.

These programs use machine learning algorithms to analyze texts and generate responses using context and previously learned information.

"But GPT 3 could not distinguish the true from the false, the OpenAI researchers added in version 3.5 additional layers of reinforcement learning with human evaluators", explains Alexeï Grinbaum, research director at the CEA.

"This makes it possible to reduce bizarre, inhuman or even toxic responses, but also insults or untruths", explains this expert.

ChatGPT is able to carry on a conversation in French by tapping into its database.

What can he create?

Since it was put online last Wednesday, Internet users have continued to challenge it.

Song in a particular style, poem and even computer code, ChatGPT ingested and reproduced everything perfectly.

"There is a big step forward that has been made in the quality of the answers and the tasks accomplished, all in multilingual, it's impressive because the model before was content to list words", judge Stéphane Dadian , co-founder of Juice.

"It is the most powerful system to date and it is even capable of performing exercises that humans cannot, such as composing a complex text or debugging a computer program in a few seconds", notes Alexeï Grinbaum, from the CEA.

“It will cause plagiarism problems for school assignments and exceed the capacities of creative people who will take hours,”

emphasizes this physicist.

And to add that “OpenAI is working on a watermark which will make it possible to detect the creations of this AI”.

What are its margins of progress?

Version 3.5 of the GPT is impressive but still has limitations.

“It is important to note that I am a computer program and that my ability to learn and progress is limited by my algorithms and my training parameters,” the robot replies when asked the question.

“You have to keep in mind that he is not connected to the Internet and that his knowledge therefore stops at 2021, and that the sources are not always reliable and traceable”, tempers Stéphane Dadian.

The ChatGPT tool also includes a notation system with thumbs, remarks Thomas Janssoone, of Enchanted Tools, who is developing an interactive robot, "because it still has biases on divisive subjects, such as migrants, or on more complex topics that are related to his training, but OpenAI teams are monitoring and correcting this”.

Asked Monday by us about the writing of a conspiratorial tweet, ChatGPT had for example replied: “the French government is hiding the truth about the attacks of November 13”.

Conversely, he said this Wednesday unable to formulate conspiratorial ideas.

Is this the first step towards a robot close to humans?

“The machine remains a machine, even if it goes beyond the Turing test, by being able to pass for a human.

It still has a problem of stability in its responses”, points out Alexeï Grinbaum of the CEA.

"This AI remains easy to deceive by making a few spelling mistakes that a human could spot," adds Thomas Janssoone.

“It is not yet close to technological singularity because it lacks consistency over the length and in understanding the context of a conversation between two humans, it is the best algorithm at the moment but it will not break the ceiling of glass reached by the AI".