ChatGPT’s name is a portmanteau. It aims to explain how this conversational agent works, capable of imitating a natural discussion.
It’s a combination of letters that makes you smile in French, but is not meaningless in English. The name chosen to baptize ChatGPT, the conversational agent developed by the American company OpenAI, has a precise meaning, which serves to describe the work carried out by this computer program when one interacts with it.
ChatGPT is a portmanteau of “chat” and “GPT”. The first term is easy to understand: it refers to the instant discussions that we have with other Internet users (except that here, we are addressing a system based on advanced algorithms in artificial intelligence). Sometimes, we write chat in French, to make the distinction with the feline.
As for GPT, it is an acronym meaning “Generative Pre-trained Transformer”, which could roughly be translated as “Transformer generative pre-trained”. The wording is complex, but it basically designates ChatGPT’s ability to give the illusion of a natural conversation, as if you were talking to another person.
A system trained on large amounts of text
ChatGPT was trained from a lot of textual data — hence the notion of pre-training found in its name. Its strength is to produce structured and articulated answers, thanks to the common methods found in the field of artificial intelligence, in particular those of machine learning.
Specifically, ChatGPT is based on a reinforcement learning approach, but also supervised — that is, with a human adjustment in the assimilation process. Reinforcement learning is the equivalent of a kind of theoretical course: we show the system examples to which we associate a notion, to learn how to make the connection.
As for supervision, it manifested itself when ” human AI trainers provided conversations in which they played both sides: the user and an AI assistant “says ChatGPT. The answers they provided were integrated with data from another model, InstructGPT. Everything has been adapted to simulate a dialogue between the AI and the Internet user.
ChatGPT is the latest iteration of the language models designed by OpenAI. The conversational agent is a specialized version of GPT-3.5, the development of which was completed at the beginning of 2022. Before it, other language models were invented, including GPT-3, which had made people talk about it at the turn of the 2020s.
Subscribe to Numerama on Google News to not miss any news!