Yann LeCun’s World Model Vision: Inside AMI Labs and JEPA
The AI industry remains heavily focused on scaling Large Language Models (LLMs), but one of the field’s most influential researchers is pursuing a fundamentally different path. After leaving Meta at the end of 2025 following more than a decade as Chief AI Scientist, Turing Award winner Yann LeCun launched a new venture centered on what he believes is the missing ingredient for artificial general intelligence: world models.
His startup, AMI Labs, is building AI systems designed to understand the underlying structure of reality rather than simply predicting the next token in a sequence. The company’s thesis challenges one of the dominant assumptions in modern AI—that sufficiently scaled language models will eventually achieve human-level intelligence.
Two research papers released in May 2026 provide the first substantial public evidence of how this vision is progressing. Together, they establish both the theoretical promise and the practical limitations of today’s world model research.
🌍 What AMI Labs Means by a World Model #
AMI Labs, short for Advanced Machine Intelligence, was founded to develop AI systems capable of understanding and interacting with the physical world.
According to LeCun, a world model is an internal representation that enables an AI system to predict how its environment will evolve and what consequences its actions will produce before those actions are taken.
Unlike language models that learn statistical relationships between words, world models aim to learn causal structures and behavioral patterns governing reality itself.
The long-term objectives include:
- Persistent memory
- Environmental understanding
- Planning and decision-making
- Causal reasoning
- Autonomous interaction with physical environments
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.
🧠 JEPA: The Foundation of LeCun’s AI Strategy #
The technical backbone of AMI Labs is the Joint Embedding Predictive Architecture (JEPA), a framework originally proposed by LeCun in 2022.
Predicting Representations Instead of Outputs #
The central idea behind JEPA differs significantly from traditional generative AI systems.
Most generative models attempt to predict every output detail, whether that means:
- The next word in a sentence
- The next video frame
- Every pixel in an image
This approach requires significant computational resources and forces models to learn many details that are inherently unpredictable.
JEPA takes a different route.
Rather than generating raw outputs, the system converts observations into abstract latent representations and performs prediction within that compressed space. The goal is to capture stable, meaningful structures while ignoring irrelevant variability.
Why This Matters #
Consider how humans learn physics.
A child does not memorize every visual detail of a falling object. Instead, the brain gradually learns abstract principles such as gravity, momentum, and object permanence.
JEPA attempts to replicate this process by learning higher-level representations that can be used for prediction and planning.
This distinction forms the basis of LeCun’s argument that future AI systems will require something fundamentally different from next-token prediction.
🚫 Why LeCun Believes LLMs Are Not Enough #
LeCun’s criticism of LLMs is one of the most consistent counterarguments to the prevailing scaling paradigm.
His position is not that language models are useless. Rather, he argues that language alone cannot produce the capabilities required for genuine intelligence.
The Limits of Text-Based Learning #
According to LeCun, LLMs remain constrained by their training environment.
Because they operate primarily in the domain of text, they lack direct understanding of:
- Physical interactions
- Spatial reasoning
- Cause-and-effect relationships
- Real-world planning
- Action consequences
He 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, increasing model size alone does not solve the underlying problem.
Implications for Robotics #
LeCun has repeatedly argued that robotics will eventually require architectures capable of modeling the physical world directly.
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.
📐 The LeJEPA Paper and the Mathematics of World Models #
One of the most significant developments from AMI Labs’ research ecosystem emerged in a paper titled When Does LeJEPA Learn a World Model?
Authored by researchers including David Klindt, Yann LeCun, and Randall Balestriero, the paper focuses on a concept known as linear identifiability.
What Is Linear Identifiability? #
In practical terms, the theorem shows that under specific conditions, a LeJEPA system can recover meaningful hidden variables from observations.
Examples include:
- Object position
- Velocity
- Underlying environmental states
Rather than exploiting superficial statistical shortcuts, the model can theoretically learn representations aligned with genuine latent factors.
The Critical Conditions #
The guarantee applies only when several assumptions hold simultaneously:
- Latent variables follow a Gaussian distribution.
- System dynamics evolve under stationary additive noise.
- Training data broadly explores the state space.
The most important conclusion is that the Gaussian assumption is not merely helpful—it is mathematically necessary for the proof to hold within the defined framework.
Formal Verification with Lean 4 #
A notable aspect of the work is that the mathematical proofs 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.
For researchers focused on theoretical AI foundations, this represents a significant methodological advancement.
The Practical Limitation #
While mathematically rigorous, 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.
📊 Stable World Models Benchmark Reveals Major Weaknesses #
If the first paper establishes theoretical feasibility, the second paper evaluates practical robustness.
The Stable World Models benchmark was introduced to measure how reliably existing world model systems perform under environmental variations.
Current Systems Are Fragile #
Results indicate that today’s world models remain highly sensitive to seemingly minor changes.
In one representative object-manipulation task:
- Success rates reached roughly 50% under standard conditions.
- Changing the agent’s color reduced success rates to approximately 12%.
- Altering background colors dropped performance to around 6%.
- Visual distractions further degraded results across all evaluated systems.
These findings suggest that many models continue to rely on superficial visual cues rather than learning truly robust environmental representations.
Prediction Accuracy Is Not Enough #
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.
This distinction highlights why planning and reasoning remain difficult problems despite rapid progress in predictive modeling.
🏁 A Growing Race Toward World Models #
AMI Labs is far from alone in pursuing this direction.
Several major research organizations and startups are investing heavily in AI systems designed to understand physical environments.
World Labs #
Founded by AI pioneer Fei-Fei Li, World Labs focuses on spatial intelligence and physically coherent 3D world generation.
Its Marble system aims to create interactive environments with realistic physical properties and has reportedly attracted a multibillion-dollar valuation.
Google DeepMind #
Google DeepMind’s Genie family explores another variation of world modeling through interactive, generative environments.
Although the underlying methodologies differ, both organizations share a belief that language alone is insufficient for building highly capable autonomous systems.
💰 Why Investors Are Backing AMI Labs #
The scale of investor interest reflects the growing conviction that post-LLM architectures could define the next phase of AI.
In March 2026, AMI Labs announced a seed funding round totaling approximately $1.03 billion at a reported valuation of $3.5 billion.
The financing was backed by a combination of venture firms, technology companies, and prominent individual investors.
Despite the enormous valuation, the company remains extremely early-stage:
- Roughly ten employees
- No commercial product
- Multi-year research roadmap
- Heavy emphasis on foundational science
The funding therefore represents a bet on scientific potential rather than demonstrated commercial execution.
🔬 What the May 2026 Papers Actually Prove #
Neither of the recently released papers demonstrates that deployable world models are imminent.
Instead, they provide a clearer picture of both the opportunities and obstacles ahead.
The identifiability research shows that learning meaningful environmental representations is mathematically possible under specific conditions. Meanwhile, the benchmark study demonstrates that existing systems remain far from achieving the robustness required for real-world deployment.
Taken together, the papers accomplish three important objectives:
- Define the theoretical requirements for successful world-model learning.
- Quantify the current performance gap.
- Establish measurable research targets for future development.
In other words, the research narrows uncertainty around the path forward without proving that the destination is close.
📌 Conclusion #
AMI Labs represents one of the most ambitious attempts to move beyond the current generation of language-centric AI systems. Led by Yann LeCun, the company is pursuing a vision in which machines learn the structure of reality itself, enabling planning, reasoning, and autonomous interaction with the physical world.
The research released in May 2026 offers both encouragement and caution. Theoretical results suggest that world models can learn meaningful latent representations under carefully defined conditions, while benchmark evaluations reveal substantial weaknesses in robustness and generalization.
Whether world models ultimately become the foundation of advanced AI remains uncertain. What is clear, however, is that some of the industry’s most influential researchers and investors are betting that understanding the world—not merely predicting text—will define the next major breakthrough in artificial intelligence.
FAQ #
What is a world model in artificial intelligence? #
A world model is an AI system that learns internal representations of how environments behave, enabling prediction, planning, and reasoning about future outcomes. Unlike language models that predict text sequences, world models aim to understand the underlying dynamics of reality.
What is JEPA? #
JEPA, or Joint Embedding Predictive Architecture, is an AI framework proposed by Yann LeCun that predicts abstract latent representations rather than generating raw outputs. The architecture is designed to learn high-level environmental structures and support planning capabilities.
Why does Yann LeCun criticize large language models? #
LeCun argues that LLMs are limited because they primarily learn from text and lack direct understanding of physical reality. He believes true intelligence requires systems that can model environments, predict consequences, and reason about actions rather than simply predicting the next token.
Has AMI Labs released a product? #
As of June 2026, AMI Labs has not publicly launched a commercial product. The company remains focused on foundational research and has indicated that its development roadmap extends over multiple years.