One of the people most responsible for making modern artificial intelligence possible now believes the industry has taken a wrong turn. Not a small wrong turn. A fundamental one.
Yann LeCun, the Turing Award-winning researcher who left Meta in 2025, told the BBC at VivaTech in Paris that ChatGPT, Claude, and Gemini are not a path towards human level or human-like intelligence, and that AI still does not understand the physical world as well as a rat.
That is a striking statement from a man who helped build the foundations that those systems stand on. And he is not just criticising from the sidelines. He has put over a billion dollars behind a different approach.
The Man Who Built Modern AI
To understand why LeCun's position matters, you need to know who he is. He is one of the three researchers, alongside Geoffrey Hinton and Yoshua Bengio, whose work on neural networks and deep learning made the current generation of AI possible. They shared the Turing Award in 2018, the highest honour in computer science. He spent years as Meta's chief AI scientist, running the company's Fundamental AI Research lab.
He is not a fringe voice with an alternative theory. He is one of the architects of the very technology he is now criticising.
He believes that the industry's current obsession with large language models is wrong-headed and will ultimately fail to solve many pressing problems. He is also a staunch advocate for open-source AI and criticises the closed approach of frontier labs like OpenAI and Anthropic.
What He Left Meta to Build
In late 2025, LeCun co-founded Advanced Machine Intelligence Labs, AMI, pronounced like the French word for friend, alongside Alexandre LeBrun. The company is headquartered in Paris, with offices in New York, Montreal, and Singapore. LeBrun serves as CEO. LeCun is Executive Chairman. The leadership team is drawn almost entirely from Meta's FAIR research organisation. [medium]
AMI Labs raised $1.03 billion in March 2026, the largest seed round in European history, at a $3.5 billion pre-money valuation. Backers include Nvidia and Bezos Expeditions, the fund managing Jeff Bezos's private wealth. Mark Cuban and former Google CEO Eric Schmidt also backed the round.
The scale of that funding reflects how seriously the research community and investment world are taking LeCun's challenge to the current AI consensus.
The Problem With Language Models
LeCun's critique of large language models is specific and technical. It is not that they are useless, clearly they are not. It is that they are built on a foundation that has a hard ceiling, and that ceiling sits well below genuine intelligence.
LeCun's research centres on world models, systems that understand cause and effect rather than just predicting text tokens. He argues current approaches hit a fundamental ceiling.
The way a large language model works is by predicting what word or token comes next in a sequence, based on patterns learned from enormous amounts of text. It is extraordinarily good at this. But it has never watched a ball fall to the ground. It has never felt resistance when pushing against something heavy. It has no internal model of how the physical world actually behaves, because everything it knows came from descriptions of the world rather than experience of it.
LeCun frequently argues that intelligence is not a magic dust you sprinkle on data. It is the ability to predict the consequences of your actions in the real world.
By that definition, current AI systems are not intelligent. They are extraordinarily sophisticated pattern matchers. But pattern matching and understanding are not the same thing, and the gap between them becomes very visible the moment you ask an AI system to reason about a genuinely novel physical situation it has never seen described before.
The Architecture He Believes Will Work
One of the key concepts in LeCun's framework is Joint Embedding Predictive Architecture, known as JEPA, which he and his collaborators have been developing as an alternative to generative models like diffusion networks or autoregressive transformers. Rather than learning to reconstruct raw data pixel-by-pixel or token-by-token, JEPA learns in an abstract representation space.
The distinction matters. A system that learns to predict the world at an abstract level, understanding that objects have weight, that actions have consequences, that space has structure, can transfer that understanding to new situations. It is not memorising descriptions. It is building a model.
AMI Labs is building world models using JEPA to give AI systems the physical and causal reasoning that pure language models lack.
Think of it as the difference between a child who has read every book ever written about swimming, and a child who has actually been in the water. The second child understands something the first one does not, regardless of how many books have been read.
The Broader Split in AI Research
LeCun is not alone in his scepticism about the long-term ceiling of large language models. But he is one of the most prominent voices making the argument, and one of the few putting serious institutional and financial weight behind an alternative.
The push for more adaptable machine intelligence reflects a genuine split emerging among leading researchers about where AI goes next.
On one side are the labs that believe scaling language models, making them bigger, feeding them more data, giving them more compute, will continue to produce meaningful gains in capability and will eventually lead to systems approaching human-level intelligence. OpenAI, Anthropic, and Google DeepMind are broadly in this camp.
On the other side are researchers like LeCun who believe language modelling is a fundamentally limited approach, and that the next breakthrough will require a different architecture built on different principles, one that engages with the physical world, reasons about cause and effect, and builds internal representations rather than surface-level statistical patterns.
LeCun's open-source strategy poses a direct challenge to proprietary AI giants. If the entire startup ecosystem builds on JEPA principles, the dominance of proprietary models from Microsoft and Google faces a systemic threat.
What This Means for the Future of AI
The stakes of this argument are not academic. The direction the industry takes over the next five to ten years will determine what kind of AI exists in the world, how it is controlled, who owns it, and what it can and cannot do.
If LeCun succeeds, the era of large language models will be remembered as a transitional, wasteful phase before the advent of true reasoned machine intelligence that understands the world much like we do.
If the scaling advocates are right, the systems that already exist will simply keep improving and the world models approach will remain an interesting research direction that never quite delivered on its promise.
Nobody knows yet which side is correct. But the fact that one of the people most responsible for modern AI has staked over a billion dollars on the bet that the current path is wrong is worth paying close attention to.
The godfather of deep learning thinks the industry he helped build has gone wrong. He is building what he believes should come next. The next few years will tell us whether he is right.