Generative artificial intelligence

Artificial intelligence (AI) was born in the 1950s, under the impetus of pioneers such as John McCarthy, Marvin Minsky and Claude Shannon. Its aim is to implement systems on computers that simulate functions deemed to be intelligent: speech and image recognition, reasoning, decision-making, etc.

Such systems are based on different types of models, in particular neural networks, or neuromimetic networks, which draw their inspiration from the human or animal cortical model: a set of very simple units ("neurons") in very large numbers and highly interconnected. A major advantage is their ability to learn from examples. Around 2010, spectacular results in many fields (go game, image interpretation, speech recognition, natural written language processing, diagnostics) highlighted a particular type of these models: Deep Neural Networks.

Such models are characterized by a large number of layers of neurons, up to several hundred. Deep learning of these models requires three conditions:

• high-performance algorithms (improved backpropagation of the error gradient);

• sometimes considerable computing resources (specialized processors such as those from Nvidia);

• the availability of large quantities of learning data, in particular Big Data, the digital data we all produce on a massive scale every day (voice and text messages, GPS signals, climate information, purchases, bank transactions, scientific publications, newspapers and magazines, etc.).

Among these deep neural networks, one model has proved particularly successful. These are convolutional networks, initially designed for image processing and later extended to a wide range of applications.

Generative AI uses deep neural networks to produce a text, image, video, music, etc., on demand. This production results from a short textual description called a prompt. This article presents the different types of generative AI models and describes how they work. The fields of text (such as ChatGPT) and image (such as MidJourney) are particularly considered.