(Paris) A senior official at Meta (Facebook, Instagram) believes that generative artificial intelligence – Chat’s technology – is already at an impasse and promises a new AI closer to human intelligence.
“Today’s AI and machine learning really suck. Humans have common sense, but machines don’t,” Yann LeCun, scientific director for artificial intelligence at Meta, told reporters on Tuesday.
He was speaking at an event in Paris when the Californian group has just presented a project for a new artificial intelligence architecture, called “Image Joint Embedding Predictive Architecture” (I-JEPA).
This computer vision technology “learns […] by creating an internal model of the outside world, comparing abstract representations of images [rather than comparing the pixels themselves],” Meta explained.
Currently, generative AI programs like ChatGPT or Bard (Google) are based on language models trained on huge databases to be able to predict which word comes after another, to the point of producing all kinds of texts (dissertations, speeches, poems, etc.).
Software like DALL-E or Midjourney works according to the same principle, to produce images.
“It’s already a revolution,” admitted Yann LeCun. “If you train [a model] with 1000 or 2000 billion tokens, it seems to be able to understand. But he makes stupid mistakes, factual or logical.”
“Generative methods focus too much on details, instead of capturing big, predictable concepts,” Meta said in a statement. This makes it difficult for them to “generate [images] of human hands accurately.”
The I-JEPA, which is based on the vision of the French scientist, is supposed to allow the machine to reason in an abstract way, much like humans.
It will be open source, that is to say open to researchers who wish to test it.
The goal is to develop AI that “better reflects how people understand the world,” Meta boss Mark Zuckerberg said on his Facebook profile.
“Generative models are a thing of the past, we are going to abandon them for predictive architectures”, like the one presented by the American group, assured Yann LeCun.
He also promises a “new rebirth of humanity” through AI, when “we all talk to intelligent [AI] agents”.
Yann LeCun is considered one of the major thinkers of this technology which has generated a lot of enthusiasm, but also concern over the past six months, since the success of ChatGPT (OpenAI).
He was very critical of all these reactions, judging that they lead to overestimating the real capabilities of the AI.
