The AI That Understands Physics: A Complete Guide to World Models in 2026
Your AI assistant can write a sonnet. It can debug code. But ask it why a basketball bounces off a backboard, and it guesses — it doesn't know. That gap between language fluency and physical understanding is exactly what World Models are designed to close.
Yann LeCun, who just left Meta to found AMI Labs with a €3 billion valuation, put it plainly: "LLMs predict the next word. World Models predict the next state of the world." In 2026, NVIDIA, Google DeepMind, OpenAI, Meta, and Fei-Fei Li's World Labs are all racing to build them. Here's what that means — and what you should do about it.
What Is a World Model?
A World Model is a neural network that simulates how the physical world behaves. NVIDIA defines it as "an AI model that understands how the physical world operates — gravity, inertia, collision dynamics." It takes video, images, sensor data, and text as input, then generates accurate simulations of what happens next.
Think of it as a computational snow globe. An AI system builds this simplified internal representation of an environment to test predictions and decisions before acting in the real world — the same way a racing driver mentally rehearses a track before qualifying.
World Models vs. LLMs: The Key Difference
| Dimension | LLMs | World Models |
|---|---|---|
| Trained on | Text patterns | Physical world dynamics |
| Core ability | Language generation | Future state prediction |
| Input | Primarily text | Video, images, sensors, text |
| Key weakness | No physical intuition | High compute requirements |
LLMs excel at "the world as described in words." World Models learn the world as it actually operates — causally, spatially, temporally.
Three Core Capabilities
- Representation Learning — Compress high-dimensional sensor data (video, lidar) into meaningful abstract concepts, not raw pixels.
- Prediction — Forecast future environmental states, including object permanence (knowing an object still exists when it's out of view).
- Planning — Simulate thousands of action sequences internally to select the optimal path before executing a single real-world movement.
Combined, these three capabilities transform AI from a question-answering tool into an autonomous agent.
Who's Building World Models in 2026?
NVIDIA Cosmos: The Physical AI Platform
NVIDIA launched Cosmos World Foundation Models at CES 2025 and has since become the leading open-source platform for Physical AI development. Cosmos consists of three models:
- Cosmos Transfer — Converts 3D simulations into photorealistic training data using structured inputs like depth maps and lidar scans.
- Cosmos Predict — Generates virtual world states from text, image, or video prompts. Given a start and end frame, it predicts plausible intermediate motion.
- Cosmos Reason — A spatiotemporal reasoning model that ranked #1 on the Hugging Face Physical Reasoning leaderboard with over 1 million individual downloads.
The full Cosmos WFM series has been downloaded over 2 million times under an Apache 2.0 open-source license. Robotics companies including 1X, Agility Robotics, and Figure AI use it to train humanoid robots. Autonomous driving companies Uber and Waabi use it for simulation.
Google DeepMind Genie 3: Real-Time Interactive Worlds
Released in August 2025, Genie 3 is the first general-purpose World Model capable of generating explorable 3D environments at 24fps, 720p resolution from text prompts alone. Waymo used Genie 3 as the foundation for its autonomous driving-specific World Model — a preview of how general models become industry-specific tools. According to Google DeepMind's official blog, Genie 3 demonstrates emergent behaviors including object interaction, complex physics simulation, and multi-character animation.
Meta JEPA and Yann LeCun's AMI Labs
Meta's JEPA (Joint Embedding Predictive Architecture) takes a fundamentally different approach: instead of predicting pixels, it predicts abstract representations. This allows the model to build an internal model of the external world without memorizing raw visual data. In late 2025, Meta released VL-JEPA (Vision-Language JEPA), challenging the dominant autoregressive paradigm in vision-language AI.
LeCun departed Meta in December 2025 to found AMI Labs (Advanced Machine Intelligence), targeting €500M at a €3B valuation. AMI Labs' explicit goal is AGI through World Model architectures.
OpenAI Sora 2 and Runway Gen-4.5
Sora 2 (released September 30, 2025) positions itself as a "world simulator," not just a video generator. OpenAI's documentation states Sora models are "a foundation for models that understand and simulate the real world." Its key advances: consistent spatial relationships across entire videos, physics-compliant motion (a basketball that misses a shot bounces correctly off the backboard), and synchronized audio.
Runway Gen-4.5 (November 2025) outperformed both Google and OpenAI video models on independent benchmarks, with particular strength in physics, human motion, and causal understanding. Runway simultaneously announced its first official World Model strategy targeting healthcare and energy.
World Labs Marble: First Commercial World Model
Fei-Fei Li's World Labs, founded with $230M in funding, released Marble in November 2025. Marble generates downloadable 3D environments (not temporary previews) from text, photos, video, or panoramic images. Pricing runs from Free (4 generations/month) to Max at $95/month (75 generations, full commercial rights). According to TechCrunch, Marble is already seeing rapid adoption in game development, VFX, and XR content production.
Real-World Applications Right Now
Autonomous Driving: Billions of Virtual Miles
In February 2026, Waymo published its Waymo World Model, built on Genie 3. It offers three control mechanisms:
- Driving action control — Simulates counterfactual scenarios ("what if we'd braked earlier?")
- Scene layout control — Customizes road layouts, traffic signals, and pedestrian behavior
- Language control — Generates scenes from text prompts like "3 AM snowstorm at a four-way intersection"
The core value: Waymo's autonomous driver can experience billions of virtual miles — including scenarios from tornadoes to elephant encounters — before encountering them on real roads.
UK startup Wayve (backed by NVIDIA) built GAIA-1, a multimodal World Model taking video, text, and driving actions as input. In South Korea, Wearable AI is running World Model-based autonomous vehicles at Incheon Airport Terminal 2 without pre-built maps.
Robotics and Physical AI
Physical AI refers to AI systems that autonomously perceive, understand, and act in physical environments — robots, autonomous vehicles, and smart spaces.
Two systems work in tandem in advanced robots:
- World Model (the brain): predicts consequences of actions before execution
- VLA — Vision-Language-Action model (the engine): translates visual and language context into physical actions
Together they enable genuine autonomy. Real deployments using NVIDIA Cosmos include 1X's NEO Gamma humanoid, Agility Robotics' warehouse Digit robot, and Figure AI's manufacturing systems. Alibaba entered the competition in February 2026 with RynnBrain, competing directly with Cosmos and Google's Gemini Robotics-ER 1.5.
Healthcare: Simulating Patient Futures
This is World Models' most consequential emerging application.
- Johns Hopkins' MeWM (Medical World Model) — Predicts tumor progression or regression under different treatment decisions, giving oncologists a visual simulation before committing to a protocol.
- NVIDIA x GE HealthCare — Announced collaboration to develop autonomous diagnostic imaging, using Cosmos to generate synthetic training data for autonomous X-ray and ultrasound systems.
- Suturing World Model — Diffusion-based model that simulates the biomechanical dynamics of robotic surgical suturing, enabling virtual practice of thousands of sutures before operating.
A 2025 survey paper (arXiv 2511.16333) defines four maturity levels for healthcare World Models, from temporal prediction (L1, widely achieved) to autonomous treatment planning (L4, extremely rare). Most clinical systems today operate at L1–L2. The global healthcare digital twin market is projected to grow from $1.17B in 2022 to over $38B by 2032, with World Models as a core driver.
Current Limitations Worth Knowing
World Models are powerful but still have significant gaps:
- Intuitive physics: A 2025 benchmark (IntPhys 2) showed that even the best vision-language AI models perform only marginally above chance when judging whether a video obeys physical laws.
- Compute cost: Video generation and physics simulation require 8–32x more GPU resources than text processing — a significant barrier to enterprise adoption.
- Data bottleneck: High-quality physical world training data is becoming scarce. Starting in 2026, data supply is projected to be a limiting factor.
Frequently Asked Questions
What is a World Model in simple terms?
A World Model is an AI system that learns how the physical world works — not by reading about it, but by processing video, sensor data, and interactions. It can then simulate "what happens next" before any real action is taken, enabling robots, autonomous vehicles, and AI agents to plan safely.
How are World Models different from video generation models?
Video generation models produce visually plausible outputs. World Models are trained to simulate causally accurate physical dynamics — not just "looks real" but "behaves physically correctly." Sora 2 and Genie 3 sit at the intersection: they generate video and are designed to encode physical understanding.
Which industries will be most disrupted by World Models?
Autonomous driving and robotics are already seeing active deployment. Healthcare (surgical simulation, diagnostic imaging, patient digital twins) will see significant impact by 2026–2028. Manufacturing, gaming, and VFX are adopting commercial tools like Runway Gen-4.5 and World Labs Marble right now.
Can my company use World Models today?
Yes. World Labs Marble (from $20/month), Runway Gen-4.5, and OpenAI Sora 2 are commercially available today. NVIDIA Cosmos is open-source under Apache 2.0. The barrier is not access — it's knowing which use case to start with.
What Businesses Should Do Now
Short-term (2026): Integrate commercial World Model APIs into your content and product workflows. Runway Gen-4.5 for video production, World Labs Marble for 3D visualization, NVIDIA Cosmos for synthetic data generation in R&D. Start building internal World Model literacy in your AI team.
Medium-term (2027–2028): Build domain-specific World Models on top of Cosmos or Genie 3 foundations. Pilot synthetic data platforms for robot or vehicle training. Evaluate Physical AI partnerships for automation initiatives.
Long-term (2029+): As World Models approach AGI-level capabilities, decision domains currently requiring human judgment will automate. Organizations with existing World Model infrastructure will have a decisive advantage.
The Bottom Line
2026 is the inflection point where AI stops reading about the world and starts understanding it.
LLMs taught machines to communicate. World Models are teaching them to think spatially, causally, and physically. NVIDIA, Google, Meta, OpenAI, and a generation of well-funded startups are simultaneously crossing the same threshold. The first commercially deployed World Models are already generating revenue.
The companies that build World Model fluency now — in their teams, their data pipelines, and their product strategies — will be positioned for the next wave of AI disruption.
For more AI trends and analysis, visit aboutcorelab.blogspot.com.