“We do not have robots which might be almost nearly as good at understanding the bodily world as a rat,” says Yann LeCun, one of many main figures on the earth of synthetic intelligence.
He labored at Fb-owner, Meta, for a decade, the place he was chief AI scientist, however left in 2025 and based Superior Machine Intelligence Labs (AMI Labs).
His aim is to maneuver AI past present methods like ChatGPT, Claude and Gemini. They’ve their makes use of, he says, however won’t ever be capable of deal with sophisticated conditions in the true world, like getting a robotic to do family chores.
“They are not a path in the direction of human degree or human-like intelligence, and even animal-like intelligence, as a result of they can’t cope with actual world knowledge, they only aren’t constructed for that,” he tells me on the sidelines of VivaTech, France’s main expertise convention.
So, Paris-based AMI Labs is busy creating a brand new sort of synthetic intelligence not based mostly on the tech behind ChatGPT and its rivals.
Buyers assume it has potential. Earlier this yr AMI Labs introduced that it had raised greater than $1bn (£760m), with traders together with US pc chip large Nvidia and the fund that manages the personal wealth of Amazon-founder Jeff Bezos.
That so-called seed funding spherical – the earliest spherical of start-up fundraising – was one of many largest of its variety in Europe.
Giant Language Fashions (LLMs) like ChatGPT are extraordinarily good at some issues like coding, mathematical issues and producing textual content, LeCun says.
However he argues that these are properly outlined and predictable issues.
“They [LLMs] mainly simply accumulate data… They will regurgitate one thing, you prepare them to regurgitate, however they are not significantly sensible. They do not have an underlying understanding,” he says.
In the true world there’s a bewildering array of outcomes to any motion, which requires a extra versatile sort of synthetic intelligence.
LeCun holds a pen upright on its tip. What occurs whenever you let go, he asks? Even a toddler would know that the pen would topple over. However no human would trouble to guess by which route the pen would possibly fall, there is no option to inform.
However an LLM would possibly attempt to generate a single prediction concerning the pen’s subsequent transfer based mostly on statistical patterns from its coaching knowledge.
The prediction would virtually definitely be improper, as a result of the system isn’t reasoning concerning the bodily actuality of the state of affairs – it’s producing what seems to be statistically believable.
LeCun says the system his firm is creating, referred to as Joint Embedding Predictive Structure (JEPA), is ready as much as cope with issues like that.
It creates abstractions of the true world that permit it to evaluate the outcomes of actions.
Creating these abstractions entails troublesome maths, however basically they filter out ineffective data, simply leaving the AI with helpful photos of the world.
Within the case of the pen, the AI would know that there is no level in making an attempt to foretell which approach the pen would fall.
