Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
“We don’t have robots that are as good at understanding the physical world as mice,” said Yan Lekun, one of the world’s leading figures in artificial intelligence.
He worked at Facebook owner Meta for a decade, where he was chief AI scientist, but left in 2015. He left in 2025 and founded Advanced Machine Intelligence Labs (AMI Labs).
Its aim is to move AI beyond systems like ChatGPT, Claude and Gemini. They have their uses, he says, but they can never solve complex situations in the real world, like getting a robot to do household chores.
On the sidelines of the French technology conference Vivatech, “they are not a path to human level or human-like intelligence, or animal-like intelligence, because they cannot interact with real-world information, and they are not built for that.”
So, Paris-based AMI Labs is busy developing new artificial intelligence that is not based on the technology behind ChatGPT and its competitors.
Investors think it has potential. Earlier this year, AMI Labs announced it had raised more than $1 billion (£760m), with investors including US computer chip giant Nvidia and Amazon founder Jeff Bezos’ private wealth management fund.
That so-called seed funding round – the first ever startup fundraising – was one of the largest of its kind in Europe.
Large-scale language models (LLMs) like ChatGPT are excellent for things like coding, mathematical problems, and text generation, LeCun says.
But he argues that these are well-defined and predictable problems.
“They[LLMs]basically just accumulate knowledge… they can refresh something, you train them to refresh it, but they’re not particularly smart. They don’t have the underlying understanding,” he says.
In the real world, there is a bewildering array of consequences for any action, requiring a more flexible artificial intelligence.
Lekun holds a pen upright on its tip. What happens when you leave, he asks? Even a toddler knows that the pen will fall. But no one can guess which way the pen might fall, there’s no way of knowing.
But LLM can try to make a prediction about the pen’s next move based on statistical patterns from the training data.
The prediction will definitely be wrong, because the system is not because of the actual reality of the situation – it is generating what seems to be statistically reasonable.
LeCun says the system his company is building, called Joint Embedding Predictive Architecture (JEPA), is designed to solve such problems.
It creates real-world abstractions that allow for evaluating the results of actions.
Creating these summaries involves difficult math, but they filter out essentially useless data, leaving the AI with only useful pictures of the world.
In the case of the pen, the AI knows that there is no point in trying to predict which way the pen will fall.
