The departure of Yann LeCun highlights the crisis of vision which opposes long-term science to short-term economic logic in AI.
Yann LeCun, Turing Prize winner and scientific director of artificial intelligence at Meta for more than a decade, announced his departure to create his own research structure.
The event, seemingly managerial, in reality reveals a major philosophical disagreement: that of the direction that artificial intelligence should take.
Behind this choice, two visions of progress oppose each other. On the one hand, the technology industry, led by Meta, OpenAI and Google, is multiplying giant language models (LLM) to respond to market logic and rapid innovation.
On the other hand, a more patient scientific approach, embodied by Yann LeCun, calls for an overhaul of AI architectures, capable of going beyond the fundamental limits of these models.
Meta, short-term logic
Under the leadership of Mark Zuckerberg, Meta recently reorganized its AI division by launching Superintelligence Labs, led by Alexandr Wang.
This new entity aims to accelerate the deployment of models like LLaMA 4, in order to catch up with the dynamics imposed by OpenAI and Google.
The context is tense. Meta AI chatbot adoption remains low, LLaMA 4 has received mixed reception, and financial markets are demanding tangible results. Despite investing more than $100 billion, Meta announced 600 job cuts in its AI division last fall.
This strategy reflects a priority given to rapid valorization rather than fundamental research.
It is this orientation that Yann LeCun now refuses to endorse.
The limits of language models
For several years, Yann LeCun has defended a clear position: current Large Language Models are not a credible path towards intelligence comparable to that of humans.
LLMs excel at generating plausible text, but fail to understand what they are producing.
They lack lasting memory, hierarchical reasoning, causal understanding, and grounding in the physical world.
For LeCun, these systems only imitate the statistical regularities of language. They are good at predicting the most likely outcome of a sentence, but incapable of constructing a representation of the world that allows them to reason about causes, effects or intentions.
This distinction is crucial: an AI capable of “speaking” does not mean an AI capable of “thinking”.
The alternative: World Models
Faced with this impasse, Yann LeCun proposes another path: that of the World Models.
These models seek to reproduce the way living beings learn to understand the world: not through language, but through perception and experience.
The idea is simple, but revolutionary: to provide machines with memory, spatial understanding and the ability to mentally simulate the world to predict its effects.
This is the basis of self-supervised learning, an approach that LeCun believes is essential to crossing the threshold into true artificial intelligence.
The I-JEPA (Image-Joint Embedding Predictive Architecture) project, developed under his direction at Meta, is a first illustration of this: it learns to predict missing representations of images from others, no longer by generating text, but by constructing abstract representations of the visual world.
The conflict between science and business
The disagreement between Yann LeCun and Meta symbolizes a broader conflict within the technology sector: that between the search for quarterly results and long-term scientific ambition.
The big AI players are banking on computing power, data volume and model scaling, believing that intelligence will emerge mechanically from size.
LeCun considers this belief to be a “conceptual bubble”. According to him, increasing parameters and GPUs are not a substitute for understanding.
Progress in AI requires architectural breakthroughs, not just engineering refinements.
A crisis of vision shared by other pioneers
LeCun is not isolated. Geoffrey Hinton, another Turing Prize winner and “father of deep learning”, left Google for different reasons, but revealing the same malaise: the loss of meaning and control in the race for superintelligence.
One warns of scientific limits, the other of ethical risks.
Their double departure highlights the same concern: the AI industry is progressing quickly, but without knowing where it is going.
A departure that gives new meaning to research
By leaving Meta to launch his own structure, Yann LeCun sends a strong message: science cannot be reduced to the speed of execution of a product.
His choice reaffirms the need to give time back to fundamental research and to reinvest the question of understanding, beyond performance.
The departure of a pioneer is not a withdrawal, but a signal: that of a return to scientific rigor in a field now dominated by market logic.
What if real progress in artificial intelligence came, paradoxically, through a slowdown?
