“We don’t have robots that understand the physical world nearly as well as rats,” says Yann LeCun, one of the leading figures in the world of artificial intelligence.
He worked for ten years at Facebook owner Meta, where he was chief artificial intelligence scientist, but left in 2025 to found Advanced Machine Intelligence Labs (AMI Labs).
Its goal is to take artificial intelligence beyond current systems such as ChatGPT, Claude and Gemini. They have their uses, he says, but they will never be able to handle complex real-world situations, such as getting a robot to do housework.
“They don’t lead to human-level intelligence, human-like intelligence, or even animal-like intelligence, because they can’t deal with real-world data, they’re just not built for that,” he tells me on the sidelines of VivaTech, France’s leading technology conference.
So, the Parisian company AMI Labs is busy developing a new type of artificial intelligence, not based on the technology of ChatGPT and its competitors.
Investors believe it has potential. Earlier this year, AMI Labs announced it had raised more than $1 billion (£760 million) with investors including US computer chip giant Nvidia and the fund that manages Amazon founder Jeff Bezos’ private wealth.
This so-called seed funding round – the earliest fundraising round for startups – was one of the largest of its kind in Europe.
According to LeCun, large language models (LLMs) like ChatGPT are extremely good at some things, such as coding, math problems, and text generation.
But he argues that these are clearly defined and predictable problems.
“They (LLMs) basically just accumulate knowledge… They might regurgitate things, you train them to regurgitate, but they are not particularly smart. They don’t have a basic understanding,” he says.
In the real world, there is a staggering variety of outcomes for any action, requiring a more flexible type of artificial intelligence.
LeCun holds the pen vertically at the tip. What happens when you let go, he asks? Even the baby knew that the pen would fall. But no one would bother to guess in which direction the pen might fall, it is impossible to say.
But LLM can try to generate a single prediction about the next pen movement based on statistical patterns from its training data.
The prediction will almost certainly be wrong because the system is not reasoning about the physical reality of the situation—it is generating what seems statistically plausible.
LeCun says the system his company is developing, called Joint Embedding Predictive Architecture (JEPA), is designed to solve such problems.
It creates abstractions of the real world that allow it to evaluate the results of actions.
Creating these abstractions requires complex mathematical calculations, but essentially they filter out useless information, leaving only useful pictures of the world for the AI.
In the case of the pen, the AI will know that there is no point in trying to predict which way the pen will fall.