PARIS, FRANCE — The chief scientist for Facebook owner Meta on Tuesday said that generative AI, the technology behind ChatGPT, was already at a dead end, instead promising new artificial intelligence resembling human rationality.
“Today AI and machine learning really sucks. Humans have common sense, machines don’t,” Yann LeCun told reporters at a Meta launch event in Paris.
LeCun spoke as Meta announced its latest AI project — called image-based Joint Embedding Predictive Architecture, or JEPA.
The project seeks to move beyond ChatGPT-like generative AI and give machines the ability to conceptualize abstract ideas and not just regurgitate what exists online.
“Generative models are past, we will abandon them in favour of joint embedding predictive architecture,” LeCun said, touting the Meta project he will lead.
“My prediction is that in a few years, generative large language models will not be used any more, we will have a better thing to replace them,” he added.
LeCun is considered a major thinker on AI and has been a critic of the hype around the generative AI models that power ChatGPT or the image-based Dall-E since they launched last year.
LeCun believes that the fears and excitement surrounding generative AI grossly inflate its actual capabilities.
In a Facebook post, Meta CEO Mark Zuckerberg said the JEPA tool was open source, meaning it would be available to researchers to tinker with.
He said the aim was to develop AI that “more closely reflects how people understand the world.”
“We need models that perceive the world and make predictions. This research is another step in that direction,” Zuckerberg added.
Compared to its rivals, Meta has taken a more discrete approach to ChatGPT-style AI for its social media platforms Facebook and Instagram.
Meta infused generative AI in its products, but without the same publicity as Microsoft or Google.
In parallel, it has also released open-source AI models that require less computing power than the technology that powers ChatGPT.