Yann LeCun, one of AI's most influential researchers, is betting that the path to artificial general intelligence requires something fundamentally different from scaling up language models: AI systems that understand how the physical world actually works. After stepping down as Meta's Chief AI Scientist in late 2025, LeCun founded AMI Labs (Advanced Machine Intelligence), a startup focused on building world models, internal representations that let AI systems predict how environments will evolve and what consequences their actions will produce before taking them.

This move represents a significant departure from the prevailing AI industry consensus. While most major tech companies continue pouring resources into larger and larger language models, LeCun argues that language alone cannot produce genuine intelligence. His new venture challenges one of the field's dominant assumptions: that sufficiently scaled language models will eventually achieve human-level reasoning and understanding.

What Makes World Models Different From Language Models?

The core difference lies in what these systems learn. Traditional language models, also called LLMs (Large Language Models), learn statistical relationships between words by predicting the next word in a sequence. World models, by contrast, aim to learn the actual causal structures and behavioral patterns that govern reality itself.

Rather than generating every output detail, world models convert observations into abstract, compressed representations and perform prediction within that compressed space. Think of how a child learns physics: they don't memorize every visual detail of a falling object. Instead, their brain gradually learns abstract principles like gravity, momentum, and object permanence. This is what world models attempt to replicate.

The technical backbone of AMI Labs is the Joint Embedding Predictive Architecture, or JEPA, a framework LeCun originally proposed in 2022. The system captures stable, meaningful structures while ignoring irrelevant variability, allowing AI to focus on what actually matters for understanding and planning.

What Capabilities Would World Models Enable?

If successful, world models could unlock several capabilities that remain out of reach for today's language-only systems:

• Persistent Memory: Maintaining coherent understanding of environments and objects over time, not just processing isolated text inputs.
• Environmental Understanding: Grasping spatial relationships, physical interactions, and how objects behave in the real world.
• Planning and Decision-Making: Predicting consequences of actions before taking them, enabling better reasoning about complex scenarios.
• Causal Reasoning: Understanding cause-and-effect relationships rather than just correlations in training data.
• Autonomous Interaction: Enabling robots and other systems to interact directly with physical environments based on learned understanding.

LeCun frequently references Moravec's paradox, which highlights how tasks humans find intuitive, such as perception, navigation, and physical interaction, remain extremely challenging for machines. From this perspective, simply increasing model size doesn't solve the underlying problem.

What Do Recent Research Papers Reveal About Progress?

Two research papers released in May 2026 provide the first substantial public evidence of how AMI Labs' vision is progressing. The first paper, titled "When Does LeJEPA Learn a World Model?", focuses on a concept called linear identifiability. It demonstrates that under specific conditions, a LeJEPA system can recover meaningful hidden variables from observations, such as object position, velocity, and underlying environmental states.

The mathematical proofs in this paper were verified using Lean 4, an interactive theorem-proving system. This level of formal verification exceeds traditional peer-review standards by allowing the logical derivation process to be independently checked by software, representing a significant methodological advancement for theoretical AI research.

However, the theorem also reveals a major engineering challenge. Many real-world robotics datasets are generated through goal-directed behavior rather than broad exploration. Such data collection strategies can violate the theorem's assumptions, meaning the theoretical guarantees may no longer apply. This gap between theory and practice remains one of the largest unresolved challenges in world model research.

The second paper, introducing the Stable World Models benchmark, evaluated practical robustness. The results are sobering: today's world models remain highly sensitive to seemingly minor changes. Success rates reached roughly 50 percent under standard conditions, but changing the agent's color reduced success rates to approximately 12 percent. Altering background colors dropped performance to around 6 percent.

These findings suggest that many models continue to rely on superficial visual cues rather than learning truly robust environmental representations. One of the benchmark's most important findings is that prediction quality does not necessarily translate into successful planning. A model may accurately forecast future observations while still failing to complete a task because it focuses on irrelevant features rather than meaningful causal structure.

How Is AMI Labs Structured and Positioned?

LeCun serves as Executive Chairman rather than CEO. Daily operations are led by Alexandre LeBrun, a former Meta FAIR researcher and co-founder of healthcare AI company Nabla. Headquartered in Paris, AMI Labs plans to expand internationally with offices in New York, Montreal, and Singapore.

LeCun's criticism of LLMs is one of the most consistent counterarguments to the prevailing scaling paradigm in AI. His position is not that language models are useless. Rather, he argues that language alone cannot produce the capabilities required for genuine intelligence. Because LLMs operate primarily in the domain of text, they lack direct understanding of physical interactions, spatial reasoning, cause-and-effect relationships, real-world planning, and action consequences.

His prediction is that future robotics systems will rely less on pure language-model reasoning and more on architectures capable of learning environmental dynamics and planning under uncertainty. This view remains controversial and far from universally accepted, but it forms the foundation of AMI Labs' research direction.

Why Does This Matter for AI's Future?

The stakes of this debate are high. If LeCun is correct, the industry's current focus on scaling language models represents a dead end for achieving artificial general intelligence, or AGI (artificial general intelligence). If he's wrong, AMI Labs' approach may prove to be a costly distraction from the path that actually works.

What makes this moment significant is that one of AI's most respected researchers is willing to stake his reputation and leave a major tech company to pursue this alternative vision. The research papers from May 2026 show both promise and profound limitations in current world model approaches. The theoretical framework suggests the approach is sound in principle, but the practical benchmarks reveal that building robust, generalizable world models remains an extremely difficult engineering problem.

AMI Labs is far from alone in pursuing this direction, though LeCun's entry into the space with substantial resources and credibility may accelerate research and investment in world models across the industry. The next few years will reveal whether this alternative path to AGI can overcome its current limitations or whether the scaling paradigm will continue to dominate AI development.

" }