Meta unveils an AI model close to human intelligence

Will artificial intelligence ever surp that of humans or animals? The scenario has nothing to do with science fiction if we are to believe Meta, the parent company of Facebook. On Tuesday, Yann LeCun, the group’s “AI chief scientist”, unveiled in Paris on the eve of the VivaTech show a new model of artificial intelligence (AI) which will be able to “understand the underlying world”. And therefore to learn, reason and plan tasks, going much further than current models.

Called “JEPA” (for “Joint Embedding Predictive Architecture”), this model is based on a new architecture designed by star seeker Last year. “The goal is to have machines at least as intelligent as humans, if not more,” explains Yann LeCun. “But today, ‘machine learning’ is really rotten compared to what humans can do. […] Therefore, something huge eludes us. »

An intelligence close to that of the dog

This model marks a major break with generative AI models, on which ChatGPT, the OpenAI chatbot, is based. Instead of milling huge amounts of data or text inserted by humans, JEPA compares abstract representations of images or sounds. The idea is to replicate the functioning of the human brain which analyzes, every day, without realizing it, a lot of information simply by observing the world around it. But “this common sense information is key to intelligence”, explains Meta.

By using representations, the hope is therefore that these models become self-learning. In this way, they will no longer have to “guess” the missing information to generate the right text or the right image – which today, in ChatGPT, generates a lot of errors, the famous “hallucinations”.

The advantage of JEPA is also that it is faster and requires less computing power. “Broad language models are still useful today, but in five years we won’t be using them anymore,” predicts Yann LeCun. Our work will take years, if not decades. At first, we will achieve an AI close to the intelligence of dogs! »

A “flashy” product

One of the first use cases of this JEPA model is image recognition. Meta, which already uses AI to eliminate terrorist or ographic images on Facebook or Instagram, will be able to use it in this context. “An advertiser will be able to find out how many photos of Louis Vuitton bags there are in all of Instagram”, ures the scientist.

But in the longer term, this AI will be able to develop complete marketing or advertising strategies for major brands, write on a legal argument, etc. But today, generative AI can simply generate a first draft of text or the basic creative concept.

With JEPA, Meta also wants to show that it remains a powerful player in AI against OpenAI, the creator of the star robot ChatGPT. And that he was not in the wrong fight by pivoting towards the metaverse at the end of 2021… “We are far from being behind in AI, this idea is completely false. We’ve been using AI internally for years. Maybe our projects weren’t so ‘flashy’”, answers Naila Murray, the research director of FAIR – Meta’s network of AI research laboratories – for the EMEA zone.

Will JEPA impose itself until the language models disappear? Meta, which put it in open source, relies on the community of developers and researchers to improve it. But the group is cautious and realistic, given the speed at which this field is evolving. “In the end, perhaps an even different model will prevail”, recognizes the Canadian Joëlle Pineau, the vice-president of AI research at Meta. “There are still thirty years of research to come in AI,” smiles Yann LeCun.