“We don’t have robots that are as good as rats at understanding the physical world,” says Yann LeCun, one of the leading figures in the world of artificial intelligence.
He worked for a decade at Meta, the owner of Facebook, where he was chief AI scientist, but left in 2025 and founded Advanced Machine Intelligence Labs (AMI Labs).
Their goal is to take AI beyond existing systems like ChatGPT, Cloud, and Gemini. He says they have their uses, but they will never be able to deal with complex situations in the real world, such as getting a robot to do household chores.
“They are not the path to human level or human-like intelligence, or even animal-like intelligence, because they can’t deal with real-world data, they’re simply not built for that,” he told me on the sidelines of VivaTech, France’s leading technology conference.
So, Paris-based AMI Labs is busy developing a new type of artificial intelligence not based on the technology behind ChatGPT and its rivals.
Investors think 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.
That so-called seed funding round – the initial round of start-up fundraising – was the largest of its kind in Europe.
LeCun says that large language models (LLMs) like ChatGPT are extremely good at certain things like coding, mathematical problems, and generating text.
But he argues that these are well-defined and predictable problems.
“They [LLMs] Basically just store knowledge… They can regurgitate something, you train them to regurgitate it, but they’re not particularly smart. They have no inherent understanding,” he says.
In the real world any action has a surprising range of consequences, requiring a more flexible type of artificial intelligence.
LeCun holds a pen upright on its tip. He asks, what happens when you let go? Even a small child would know that the pen would tip over. But no human being would bother to guess, there being no way of telling in which direction the pen might fall.
But an LLM might try to generate a single prediction about the pen’s next move 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 appears to be statistically plausible.
LeCun says the system his company is developing, called Joint Embedding Predictive Architecture (JEPA), is set up to tackle such problems.
It creates abstractions of the real world that allow it to assess the consequences of actions.
Creating these abstractions involves difficult mathematics, but essentially they filter out useless information, leaving the AI with useful pictures of the world.
In the case of the pen, the AI would know that there is no point in trying to guess which direction the pen will fall.
