No, ChatGPT isn't smart!2 reading minutes

ChatGPT software doesn't analyze or synthesize, nor is it creative, but generates «a credible follow-up to a given text». The result can then be adorned with the trappings of human intelligence. This is enough to arouse fears and anxieties.

So as not to give in superficially to the general excitement, miracle sellers or nightmares, it's worth taking a look at the technical workings of a tool like ChatGPT. Doctor of Philosophy Thibault Giraud, host of the Monsieur Phi YouTube channel, takes a closer look at the subject in a video posted on January 8 entitled «What is ChatGPT really capable of?.

Accompanied by Lê Nguyên Hoang, a researcher at EPFL, he explains the key to ChatGPT's operation: «Most AIs (editor's note: artificial intelligence) that generate text are based on so-called large language models (“big language models”)». These models are designed to perform a single, very precise task: «predicting the next word in a text». What sets them apart at the moment," explains Thibault Giraud, "is their ability to do this extremely efficiently.

Read also | Discover our ChatGPT dan operations Le Regard Libre N°94

On the one hand, it is based on an immense amount of training data («a gigantic sum of texts found on the Internet [such as the entire] Wikipedia, but also millions of other pages»). On the other hand, it relies on the sophistication of the algorithms, since in order to ’improve its predictions during the training phase, the language model can adjust a dizzying number of parameters«. For the language model behind ChatGPT, we're talking about »at least 175 billion parameters«.

In short, it's important to understand that such a tool doesn't «answer» the questions put to it, as a human being would. It doesn't analyze or synthesize, nor does it demonstrate creativity, but it does generate «a credible follow-up to a given text». The result can then be adorned with the trappings of human intelligence - just trappings! Enough to motivate many fears and fantasies.

Write to the author: antoine.bernhard@leregardlibre.com

You have just read a commentary from our print edition (Le Regard Libre N°94).