Yann LeCun at AI Summit 2025: New Horizons of AI

I firmly believe that to achieve true human-level intelligence, we need to fundamentally rethink how machines learn and interact with the world. This is not just a matter of computational capacity but of understanding how the human brain processes and reacts to complex information.

In this article:

At the AI Action Summit 2025, Yann LeCun, Meta’s chief AI scientist and NYU professor, shared groundbreaking insights on the future of Artificial Intelligence. LeCun, a pioneer in the field, emphasized the need to move beyond today’s language models like GPT and explore new architectures that truly understand and interact with the world in a human-like way.

The Current AI Paradigm: Limitations and Challenges

LeCun argues that despite significant advancements in AI, we are far from replicating human intelligence. "We can't even replicate the intelligence of a cat or a rat, let alone a human," he affirms. He points out that although we have systems capable of passing complex exams, we still do not have domestic robots that can clear the dinner table or level-five autonomous cars.

One of the reasons for this, according to LeCun, is that current AI systems are trained on large volumes of text but lack an understanding of the physical world that even a young human possesses. He highlights that a typical language model is trained on trillions of tokens, yet it still does not come close to the sensory experience a four-year-old child accumulates in their first interactions with the world.

Rethinking AI Architecture

To overcome these limitations, LeCun proposes a shift to joint prediction architectures, where AI not only predicts the next word in a sequence but understands the underlying structure of the world. He suggests using energy-based models that measure the compatibility between observations and proposed outputs, allowing deeper reasoning and contextual understanding.

These models, called JEA (Joint Embedding Architectures), eliminate the need to predict every detail, focusing on abstract representations that capture the essence of what is being observed. This is crucial for developing systems that can plan and execute actions based on mental models of the world, much like a human would.

The Importance of Memory and Hierarchical Planning

LeCun emphasizes that to achieve human-level intelligence, AI systems must have persistent memory and the ability to plan actions hierarchically. This means AI should be able to break down complex tasks into subtasks, just as a human plans a trip or performs daily tasks.

He points out that currently, hierarchical planning in robotics is done with manual representations, but the real innovation will be training architectures that learn these representations autonomously. "Animals do this, humans do this very well, but we are completely unable to do this with machines today," LeCun states.

The Future of AI: Universal Virtual Assistants

The future, according to LeCun, belongs to universal virtual assistants that will mediate all our interactions with the digital world. These systems cannot be monopolized by a few companies in Silicon Valley or China. Instead, he advocates that AI platforms should be open and collaborative, allowing everyone to contribute and benefit.

He warns that keeping science secret will only lead to delays and that global collaboration is essential for AI progress. "Open-source models are slowly but surely outperforming proprietary models," he states.

Conclusion: A New Era for Artificial Intelligence

Concluding his talk, LeCun highlights that true human-level intelligence in AI is still distant, but the path is clear. We need architectures that not only mimic human language but understand and interact with the world meaningfully and safely.

For those wishing to explore more about how AI is shaping the future, check out our article on the next wave of AI and the ethical challenges it brings.