Generative AI has transformed how we interact with technology. Large language models can write code, answer complex questions, translate languages, and generate human-like conversations. But are they the path to human-level intelligence?
Yann LeCun doesn't think so.
During his VivaTech 2026 session, the Turing Award winner, former Chief AI Scientist at Meta, and co-founder of AMI Labs challenged one of the biggest assumptions driving today's AI race. Large language models, he argued, are powerful, useful, and often superhuman in specific tasks, but they're not the foundation on which truly intelligent machines will be built.
Why Language Models Cannot Understand the World
At the heart of LeCun's argument is a simple question: how do humans learn? Before children can read or write, they learn by exploring and interacting with the world around them. They develop an understanding of objects, movement, cause and effect, and how actions lead to consequences. That mental model becomes the foundation for reasoning, planning, and decision-making.
Language, LeCun argued, captures only a small fraction of that experience. A language model trained on text learns statistical relationships between words. It can generate convincing responses because it has seen enormous amounts of written information. But predicting language is fundamentally different from understanding the physical world that language describes.
That distinction, LeCun argued, explains why today's AI systems can appear extraordinarily capable while still making mistakes that seem obvious to humans.
This is shaping the work at AMI Labs. Rather than building increasingly large systems designed to predict the next word, LeCun is focused on creating AI systems that learn world models: internal representations of reality that allow machines to understand how the world works, reason about cause and effect, and plan actions to achieve goals under real-world constraints.
It is a fundamentally different approach to artificial intelligence.
Instead of relying primarily on text, these systems would learn through observation, interaction, and experience, much more like humans do.
LeCun acknowledged that this is the harder path. Progress may take years rather than months, particularly at a time when much of the AI industry is focused on rapidly scaling today's language models. But for LeCun, increasing the size of existing models alone will not produce human-level intelligence.
Why an AI Monoculture Concerns Yann LeCun
The discussion extended beyond technology to the way AI itself is being developed. As frontier models become larger and more expensive to train, development has become concentrated among a relatively small number of companies with access to enormous computing resources. According to the Stanford AI Index, the cost of training leading foundation models has increased dramatically in recent years, making frontier AI development accessible to only a handful of organizations. LeCun believes that concentration carries significant risks. "That creates this monoculture," he said.
In his view, a small number of companies should not determine how billions of people interact with AI or access information.
Instead, he argued for a more open ecosystem, one where many organizations contribute to research, many models coexist, and innovation happens across a broad scientific community rather than within a few private platforms. For LeCun, openness is not simply a philosophical preference. It is a practical way to accelerate scientific progress while encouraging competition and innovation.
The Case for Open-Source AI
That philosophy also explains LeCun's position on one of AI's most closely watched debates: open-source models. As governments and companies discuss new safeguards for increasingly capable AI systems, some have argued that the most powerful models should remain tightly controlled. LeCun strongly disagrees. Comparing efforts to restrict open AI research with earlier attempts to limit access to scientific knowledge, he argued that openness has historically accelerated innovation rather than hindered it. "If you're going to block this because you think it's dangerous, you are being a medieval obscurantist." His comments reflected a broader debate taking place across the AI community. How should society balance openness with safety? And who should control the technologies likely to shape the next generation of innovation?
What Comes After Generative AI?
Although LeCun challenged many of today's assumptions, he was equally careful to recognize what current language models already do exceptionally well. From coding and translation to content generation, they have demonstrated remarkable capabilities across a wide range of specialized tasks.
But throughout the session, he returned to three ideas: language is only one way humans learn, AI systems need to understand the physical world rather than simply describe it, and future breakthroughs will require new architectures beyond today's large language models.
That debate matters because companies, governments, and investors are collectively investing hundreds of billions of dollars into artificial intelligence. The direction researchers choose today will shape the next generation of AI systems for years to come.
For LeCun, the future of AI will depend less on building bigger models and more on building smarter ones.
His vision is a long-term one, measured not by the next product launch, but by the scientific breakthroughs still to come. Whether or not the industry ultimately follows that path, his ideas continue to shape one of the most important debates in artificial intelligence today.
Watch the replay of Yann LeCun's VivaTech 2026 session, Beyond Language Models, to explore his vision for the future of AI, world models, and open-source artificial intelligence.
