In November 2025, Yann LeCun — Turing Award winner, co-founder of modern deep learning, Meta's Chief AI Scientist for twelve years — resigned. In March 2026, his new laboratory, AMI Labs, closed a $1.03 billion funding round at a $3.5 billion valuation, the largest seed round in European startup history. The reason: LeCun believes large language models will never achieve general intelligence — and that the path forward runs through a completely different architecture.
The Thesis: LLMs Are a Dead End for AGI
LeCun's position is direct and contradicts the dominant field consensus: models trained to predict the next word in text sequences cannot, in principle, develop the type of physical world understanding necessary for general intelligence. "Today's AI is dumber than a cat," LeCun said on multiple public occasions in 2026. "A cat understands intuitive physics. It knows solid objects don't pass through each other. No LLM truly knows that."
The alternative he proposes is world models — systems that learn an internal representation of how the environment works, including causal relationships, physics, and consequences of actions — rather than statistical patterns in text. The central architecture he develops is JEPA (Joint Embedding Predictive Architecture), which learns by predicting abstract representations of image regions or scenes — not the pixels themselves, but the concepts they represent.
JEPA: Predicting Concepts, Not Pixels
The technical distinction between JEPA and generative models is fundamental. A generative video model (like OpenAI's Sora) proves it "understood" a scene by attempting to recreate it pixel by pixel. LeCun argues this approach is fundamentally misguided: most of what happens in a video — leaves trembling, light flickering, individual air molecules — is intrinsically unpredictable. Forcing a network to predict this wastes capacity and corrupts representations.
JEPA avoids the problem by predicting in the space of abstract representations. Instead of "what will this pixel be," it asks "what will this concept be." In May 2026, two preprints from LeCun's group formally proved when the JEPA architecture can recover real-world structure — and identified how far current implementations still fall from that standard.
Who Is on the Other Side
LeCun's bet is not alone. Fei-Fei Li founded World Labs focused on "spatial intelligence" — raising $1 billion to develop models that understand three-dimensional space. DeepMind launched Genie 3, capable of simulating 3D worlds in real time. NVIDIA entered with Cosmos, a suite of foundation models for robotics-focused world models — with 2 million downloads by mid-2026.
However, the field is genuinely divided. In March 2026, at the Spring School AI For Impact, LeCun and a rival researcher staged what was described as "a forensic disagreement about what intelligence is for." The central tension: are current transformers and LLMs sufficient for AGI with more scale and multimodal data, or is a fundamental architectural shift necessary?
What Is at Stake for Robotics and Agents
The most immediate use case for world models is not language — it is robotics and autonomous agents. A robot that needs to manipulate physical objects must understand that a glass breaks if it falls, that liquids flow downward, that surfaces have textures that affect grip. No current LLM has this understanding robustly — it has textual approximations of physical concepts, not causal world models.
If LeCun is right, the next generation of autonomous agents — the systems that labs like OpenAI and Anthropic consider the next major market — will require an architecture that current transformers simply do not provide.
