AI Pioneer Says Today's Artificial Intelligence Is 'Not Smart' Enough for the Real World

One of the world's leading artificial intelligence researchers has argued that today's most advanced AI systems are still far from possessing genuine intelligence, saying they are unable to understand or reason about the real world.

Yann LeCun, former chief AI scientist at Meta and founder of Paris-based Advanced Machine Intelligence Labs (AMI Labs), believes current large language models (LLMs), including ChatGPT, Claude and Gemini, have reached the limits of what their underlying technology can achieve. Although these systems perform exceptionally well in tasks such as writing, programming, mathematics and text generation, he says they are fundamentally incapable of understanding the physical environment in the same way that humans or even animals do.

Speaking during the VivaTech technology conference in France, LeCun said current AI models simply memorize patterns found in enormous datasets rather than developing a true understanding of reality. According to him, they can reproduce information that appears convincing but cannot reason through unpredictable situations encountered in everyday life.

To illustrate the problem, LeCun used the example of balancing a pen on its tip. While even a young child understands that the pen will eventually fall, nobody can accurately predict the exact direction it will take. Current language models, however, attempt to generate a single statistically likely answer rather than recognizing that the outcome is inherently uncertain.

LeCun believes the future lies in a different architecture known as Joint Embedding Predictive Architecture (JEPA), which his company is developing. Instead of predicting every possible detail, JEPA creates simplified internal representations of the world, allowing AI to understand which information is important while ignoring unnecessary details. This approach could enable machines to make decisions more like humans.

The research has attracted major financial backing. Earlier this year, AMI Labs secured more than one billion dollars in seed funding from investors including Nvidia and a fund managing the private wealth of Amazon founder Jeff Bezos. The investment ranks among the largest early-stage funding rounds for a European AI company.

The limitations of current AI are especially evident in robotics. While humanoid robots have become increasingly capable, teaching them to safely perform household tasks such as ironing clothes or loading a dishwasher remains extremely difficult. LeCun argues that simply making language models larger will never solve these challenges.

Many researchers share this assessment. Ingmar Posner, Professor of Applied Artificial Intelligence at the University of Oxford, says the next generation of AI must be able to explain why events happen, understand cause-and-effect relationships and evaluate alternative actions before making decisions.

Posner's research team has spent several years developing what are known as "World Models," AI systems designed to build internal simulations of reality. Rather than relying only on statistical patterns, these systems organize knowledge in a way that allows them to predict consequences, adapt to new situations and improve decision-making.

Several leading AI companies are pursuing similar approaches. Google DeepMind is developing the Genie model, autonomous driving company Wayve is working on Gaia, while AI pioneer Fei-Fei Li founded World Labs to advance world-model technology.

LeCun expects AMI Labs to begin deploying its new AI system in industrial environments next year before expanding to broader applications. Ultimately, he believes future AI systems will become highly capable assistants rather than replacements for humans.

According to LeCun, even if artificial intelligence eventually surpasses human intelligence in many tasks, people will continue to play the central role in deciding what problems should be solved, what technologies should be created and what goals society should pursue.