“AIs cannot yet reason and plan”

What will happen after ChatGPT and Midjourney? Behind the rise of generative artificial intelligence (AI) tools, capable of creating texts and images, what are the giants of the sector working on? At the Festival of World, Sunday September 17, Naila Murray discussed the research avenues of Meta, the parent company of Facebook and Instagram, where she directs the AI ​​laboratories in Paris, London, Zurich and Tel Aviv.

Read also: Article reserved for our subscribers AI: Meta puts pressure on OpenAI and Google with an open source model

The major language processing models – on which many tools, including ChatGPT, are based – “have a statistical logic” And “cannot yet reason and plan”, underlined Mme Murray. One way of explaining that this software is limited to creating realistic texts, resembling the billions of pages they have ingested. In its fundamental research, Meta is therefore working on AIs which would be endowed with these capacities and which could forge a “representation of the world”.

Behind this abstract expression lies the idea of ​​allowing software to learn things, without human supervision, by watching videos or even observing reality, via video sensors. “This is inspired by the ability that humans or animals have to understand how the world works, which allows them to make predictions: if they see a person walking, they expect that person to pose their foot on the ground because they understand what walking is. If they see me throw a ball up in the air, they know it’s going to come down, etc. “, explained M.me Murray.

Software that imitates human “common sense”

With this in mind, Yann LeCun, the director of artificial intelligence research at Meta, launched in June I-Jepa, an attempt to create a new AI model capable of learning, by ingesting images, abstract representations of things (the paws of a dog, the forehead of a dog, etc.), without making the effort to analyze each pixel, as the do current image generation software.

The goal is to imitate the ” common sense “ human. And to ultimately create machines that “learn much faster, plan how to accomplish complex tasks, and adapt to unexpected situations,” explains the company. Mr. LeCun, one of the French researchers recognized as the fathers of modern AI, has sometimes given the example of software capable, if asked to organize a journey from Paris to New York, of giving the route to leave the building, order a taxi, guide us to the airport, buy a plane ticket, etc.

You have 52.27% of this article left to read. The rest is reserved for subscribers.

source site-30