World Models: What They Are and Why Everyone Is Working on Them in 2026

February 2026 — Yann LeCun is leaving Meta to focus on World Models. Fei-Fei Li is launching World Labs. Google DeepMind is advancing Genie. Runway is working on video generation. What do all these projects have in common? World Models — the next frontier after language models.

What World Models are

The simple explanation

A World Model is an AI system that:

1. Understands the basic laws of the world — gravity, inertia, cause and effect
2. Can run simulations — "what happens if..."
3. Plans actions — based on its predictions

An analogy: When you see a ball heading toward the edge of a table, you know it's going to fall. That's your "world model" — an internal simulation of physics.

LLMs (large language models) don't have a model like this. They can describe a ball falling, but they don't understand why it happens.

The technical definition

A World Model is a learned simulator that:

• Takes in the current state (observation)
• Takes in an action
• Predicts the next state
• Works in an abstract space (latent space), not in pixels

How it differs from an LLM

Why this matters now

The limits of LLMs

By 2026, it had become clear that LLMs have real limitations:

1. Hallucinations — LLMs confidently make things up because they don't understand reality
2. No common sense — they don't know basic physics
3. Poor planning — they can't think many steps ahead
4. No grasp of causality — correlation ≠ causation

World Models address these problems

• Fewer hallucinations — the model checks its predictions against a simulator
• Common sense — an understanding of physics is built in
• Planning — you can "play out" actions in a simulation first
• Causality — the model learns from interventions, not just observations

Who's working on World Models

AMI Labs (Yann LeCun)

• Focus: The JEPA architecture, a general understanding of the world
• Funding: €500M at a €3B valuation
• Status: Startup, January 2026

LeCun believes the path to AGI runs through World Models, not LLMs.

World Labs (Fei-Fei Li)

• Focus: Spatial intelligence, 3D understanding
• Funding: $1B+
• Status: Launched in 2024, in active development

Fei-Fei Li (the creator of ImageNet) is working on AI that understands 3D space.

Google DeepMind (Genie)

• Focus: Generative Interactive Environments
• Product: Genie 2 (announced in 2025)
• Status: In active development

Genie can generate interactive worlds from text or an image.

Runway

• Focus: Video generation as a World Model
• Product: Gen-3
• Status: Production-ready

Generating video is essentially predicting the next frames — which is exactly what a World Model does.

Tesla (FSD)

• Focus: Autonomous driving
• Product: Full Self-Driving
• Status: In production

Tesla's self-driving system is a World Model that predicts how traffic will behave.

How World Models work

Architecture (simplified)

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   Encoder   │ ──▶ │ World Model │ ──▶ │   Decoder   │
│ (Perception)│     │ (Dynamics)  │     │ (Action)    │
└─────────────┘     └─────────────┘     └─────────────┘
      ▲                   │                   │
      │                   ▼                   ▼
   Observation      Latent State          Action

Key components

1. Encoder (Perception) Turns raw data (images, sound) into an abstract representation.

2. World Model (Dynamics) Predicts how the state will change when an action is taken.

3. Decoder (Action) Picks the best action for reaching the goal.

Example: a robot pouring water

The LLM approach:

1. Instructions: "Pick up the kettle and tilt it over the cup"
2. The robot doesn't know how far to tilt it
3. The water spills

The World Model approach:

1. The model simulates the physics of the water
2. It predicts the trajectory at different angles
3. It picks the optimal angle
4. It pours successfully

JEPA: LeCun's architecture

Joint Embedding Predictive Architecture

JEPA is LeCun's approach, and it differs from traditional methods:

The traditional approach:

• Predict pixels/tokens
• Work in a high-dimensional space
• Lots of noise, slow training

JEPA:

• Predict abstract representations (embeddings)
• Work in latent space
• Less noise, more efficient

I-JEPA (Image JEPA)

An example that already works:

1. The model sees part of an image
2. It predicts the embedding of the hidden part
3. It doesn't generate pixels — only an abstraction

The result: an understanding of structure without unnecessary detail.

V-JEPA (Video JEPA)

The extension to video:

• Predicts future frames
• Learns dynamics and physics
• Lays the groundwork for World Models

Applications of World Models

1. Robotics

Why it matters: Robots need to understand physics to manipulate objects.

Examples:

• Figure 02 (a humanoid robot)
• Boston Dynamics Spot
• Tesla Optimus

2. Autonomous systems

Why it matters: Cars and drones need to predict how everything around them will behave.

Examples:

• Tesla FSD
• Waymo
• DJI drones

3. Games and simulations

Why it matters: Generating realistic, interactive worlds.

Examples:

• Google Genie
• NVIDIA Omniverse
• Unity ML-Agents

4. Video generation

Why it matters: Generating video = predicting physically accurate sequences.

Examples:

• Runway Gen-3
• OpenAI Sora
• Google Veo

5. Scientific simulations

Why it matters: Modeling physical, chemical, and biological processes.

Examples:

• AlphaFold (proteins)
• Climate models
• Drug discovery

Technical challenges

1. Data

The problem: Where do you get data about the physics of the world?

Solutions:

• Simulators (but there's a gap between simulation and reality)
• Video (YouTube, but there's no ground truth)
• Robots (expensive and slow)

2. Scaling

The problem: Does JEPA work at GPT-4 scale?

Status: Not proven yet. So far, experiments have been limited to small tasks.

3. Evaluation

The problem: How do you measure "understanding physics"?

Approaches:

• Physics benchmarks (PHYRE, Physion)
• Robotics tasks
• Video prediction

4. Compute

The problem: Simulating the world takes a lot of resources.

Solutions:

• Efficient architectures
• Specialized hardware
• Hierarchical models

World Models vs. LLMs: who wins?

The case for LLMs

• Scaling works — going from GPT-4 to GPT-5 brings real improvements
• Versatility — text covers almost everything
• Infrastructure — it's already built
• OpenAI, Anthropic — the top companies are betting on LLMs

The case for World Models

• Physics is essential — for robots, cars, and the real world
• LLMs have hit a ceiling — hallucinations aren't going away
• LeCun, Fei-Fei Li — leading researchers, including Turing Award winner LeCun, are betting on World Models
• AGI demands it — it requires understanding the world, not just text

The likely outcome

A hybrid: LLMs for language and reasoning + World Models for physics and planning.

For developers

What to study

1. V-JEPA papers — the architecture fundamentals
2. DeepMind's Genie — a practical implementation
3. MuJoCo/Isaac Gym — simulators for experiments
4. RL fundamentals — World Models are closely tied to reinforcement learning

Resources

• Papers: "World Models" (Ha & Schmidhuber, 2018), "JEPA" (LeCun, 2022)
• Code: GitHub huggingface/world-models
• Courses: Deep RL courses from Berkeley and Stanford

Conclusion

World Models aren't hype — they're the next logical step after LLMs. When Yann LeCun stakes €500M and his reputation on a technology, it's worth paying attention.

Key takeaways:

1. World Models understand physics — unlike LLMs
2. Key players: AMI Labs, World Labs, DeepMind, Runway
3. Applications: robots, cars, video, games, science
4. Challenges: data, scale, evaluation
5. The future: a hybrid of LLMs + World Models

Keep an eye on this field — it may be where AGI is born.