At NVIDIA’s GTC 2025, Meta’s Chief AI Scientist Yann LeCun delivered a provocative message: the era of Large Language Models (LLMs) is nearing its end. In a candid discussion, LeCun criticized LLMs as “token generators” and advocated for a new generation of AI architectures grounded in reasoning, planning, and real-world understanding.
“I’m Not So Interested in LLMs Anymore”
LeCun, a Turing Award laureate and one of the pioneers of deep learning, expressed his waning interest in LLMs, stating, “I’m not so interested in LLMs anymore.” He elaborated that LLMs are primarily product-driven technologies that are reaching their limits, as they rely heavily on predicting the next token in a sequence without genuine understanding.
He further emphasized that this approach is inherently flawed: “When you train a system to predict tokens… you can never train it to predict the exact token that will follow.” This critique underscores his belief that LLMs, despite their impressive capabilities, are fundamentally limited in achieving true machine intelligence.
A Shift Toward World Models and Reasoning
LeCun is advocating for a paradigm shift in AI development, focusing on four key areas:
- World Models: Internal representations that allow AI systems to perceive, plan, and predict, akin to how humans understand the world.
- Persistent Memory: The ability for AI to retain and utilize information over time, moving beyond the transient nature of current models.
- Reasoning: Developing systems capable of logical inference and decision-making, rather than mere pattern recognition.
- Planning: Enabling AI to formulate and execute complex sequences of actions to achieve specific goals.
He believes that these components are essential for building AI systems that can truly understand and interact with the world, moving beyond the superficial capabilities of current LLMs.
Critique of Token Prediction
LeCun highlighted the limitations of training AI systems solely on token prediction, likening it to “writing a program without knowing how to write it… hopeless.” This analogy underscores his view that LLMs lack the structural understanding necessary for genuine intelligence and are confined by their reliance on vast amounts of data without true comprehension.
The Road Ahead: Joint Embedding Predictive Architectures
Looking forward, LeCun is exploring joint embedding predictive architectures—systems designed to forecast future states rather than just the next word. This approach aims to create AI that can model the world more accurately, enabling it to reason and plan effectively. He envisions AI systems that can understand the physical world, possess persistent memory, and exhibit advanced reasoning capabilities, moving closer to human-like intelligence.
A Reality Check on AGI Hype
As discussions around Artificial General Intelligence (AGI) gain momentum, LeCun offers a grounded perspective. He argues that the current trajectory, heavily reliant on scaling up LLMs, is unlikely to lead to AGI. Instead, he emphasizes the need for fundamentally new architectures that incorporate world models, reasoning, and planning. This science-first approach, he believes, is the true path toward achieving intelligent systems that can operate autonomously and effectively in complex environments.
Conclusion: Embracing a New Paradigm
Yann LeCun’s insights at GTC 2025 mark a pivotal moment in AI research. By challenging the dominance of LLMs and advocating for architectures grounded in real-world understanding and reasoning, he is steering the AI community toward a future where machines can think, plan, and interact with the world in a more human-like manner. This shift promises to redefine the boundaries of artificial intelligence and its role in society.
Source: LinkedIn