ChatGPT: configuration, errors, limits… how does this amazing artificial intelligence work?

Since its launch in November 2022, the chatbot using artificial intelligence (AI) ChatGPT has been making a lot of noise. But what does he really have in his stomach?

Summarize a text, write a poem or an article, provide an email template, solve a calculation, generate code… In seconds, ChatGPT provides answers, often stunning, to (almost) anything you ask of it. But how does he do it?

ChatGPT, what is it?

Interviewed by Le Parisien, Olivier Martinez, consultant in product management and Artificial Intelligence (AI), defines ChatGPT as “a machine that mixes generative AI with conversational AI”.

“Conversational AI, we already know, is the chatbot that answers next door every time you ask it a question. Like ‘call customer service’ when you can never reach it, details- Generative AI is an AI that can generate text, sound, images and video from automatic language processing. When spoken to, it identifies ideas and from that she can answer.”

Concretely, ChatGPT is trained “to give, from a textual input, the following sequence of the most probable words”, all from “175 billion parameters that uses pre-trained linguistic representations from huge databases (Wikipedia, web pages) to produce texts that mimic human language in multiple languages,” says Laurence Devillers, specialist of digital ethics, in a forum in Le Monde.

“By construction, the model only recorded information about the words and parts of words and their context of appearance”, she adds.

What are his limits ?

If it arouses the enthusiasm of many users, ChatGPT is also targeted by a certain number of critics because of its limits. By its nature, ChatGPT is only trained to predict the probability of the next word in a sentence.

Thereby, he does not understand your question or even his own answer and does not “verify at any time the veracity of the information” he provides, emphasizes Laure Soulier, IT lecturer at ISIR, for Sciences et avenir.

It is therefore better to know a minimum of the subject on which you are questioning him. Many erroneous answers provided by ChatGPT have already been relayed on the web.

To maximize your chances of obtaining an answer that meets your expectations, it is advised to brief ChatGPT by providing it with context which will allow him to better adjust his text. “For example, if I need legal advice, I tell him ‘Act like a lawyer and give me legal answers'”, illustrates Olivier Martinez.

As Laure Soulier points out, ChatGPT does not guarantee respect “for copyright either, because it is not immune to copying bits of sentences or passages in full that it has already seen during his apprenticeship”. It is probably risky to use it to write a dissertation or a press article, for example.

As perfectible as ChatGPT is, its exceptional launch has not gone unnoticed in the IT world. Thus, the red alert was quickly triggered at Google, which sees this AI as a “threat”. Long cautious about this type of tool, the net giant could well start dancing in turn.

Leave a Comment

Your email address will not be published. Required fields are marked *