Yann LeCun Shocks: LLMs are “nonsense”, AI must pivot to achieve human-level intelligence

• Yann LeCun, one of the godfathers of modern AI, declared that Large Language Models (LLMs) cannot reach human-level intelligence and called the current approach “nonsense.”
• Just two weeks after leaving Meta, he announced a new vision at the AI-PULSE conference at Station F (Paris).
• The reason for leaving Meta stemmed from strategic disagreements: Meta continues to pour massive resources into LLMs, while LeCun argues LLMs only mimic data and do not truly “understand” the world.
• His new direction focuses on “world models”—systems that build models of how the world functions and predict the consequences of actions.
• LeCun believes that only when AI possesses the ability to reason and plan based on an understanding of reality will human-level intelligence be possible.
• He rejects the view that “scaling current technology is enough,” arguing for a complete shift in the AI thinking paradigm.
• LeCun’s new startup is based in Paris, reflecting his ambition to build a European AI hub independent of the American-style hyper-data center model.
• According to LeCun, training the V-Jepa-2 model requires only a few thousand GPUs.
• This number is significantly smaller than Elon Musk’s Colossus supercomputer, where the Grok chatbot operates with over 200,000 GPUs.
• He argues that Europe does not need to race in building massive data centers to achieve an AI breakthrough.
• LeCun has not revealed project details yet but promised an official announcement next January.
• The project is expected to challenge the dominant thinking of LLMs in the global AI industry.

📌 Yann LeCun, one of the godfathers of modern AI, declared that Large Language Models (LLMs) cannot reach human-level intelligence and called the current approach “nonsense.” His new direction focuses on “world models”—systems that build models of how the world functions and predict the consequences of actions. Leaving Meta to found a startup in Paris, training models with only a few thousand GPUs, and comparing them to 200,000-GPU systems demonstrates a completely different path.