A language model tells you what usually happens. A world model tries to tell you what will happen if you act.

In March 2026, AMI Labs—the Paris-based startup co-founded by Turing Award winner Yann LeCun after he left Meta—raised $1.03 billion at a $3.5 billion pre-money valuation to build world models. Its chief executive predicted that within six months every company would call itself a world model to raise money. He was not far wrong. “World model” now appears in pitch decks for robotics, autonomous vehicles, video generation, supply-chain software, and even customer-service agents.

For business leaders, the term has become hard to evaluate. This guide separates the substance from the marketing: what a world model is, how LeCun's JEPA approach differs from video generators and language models, what has been demonstrated as of October 2026, which industries should run pilots now, and how to scope a world-model project that produces evidence instead of a slide deck.

What Is a World Model?

A world model is a learned internal representation of an environment that predicts how that environment will change—usually as a function of an action. Given the current state and a candidate action, it answers: what happens next? That lets an AI system simulate consequences before acting, compare alternative plans, and choose the one most likely to reach a goal.

The idea is not new. Control engineers have used models of physical systems for decades, and game-playing systems search through simulated futures. What changed is that world models can now be learned from large amounts of raw video and sensor data rather than written by hand, which makes them applicable to messy, open-ended environments such as kitchens, warehouses, roads, and factory floors.

World Models vs Language Models vs Video Generators

A language model predicts text. It can describe what usually happens in a situation, because it has read many descriptions, but it has no direct representation of physical state and does not natively predict the consequence of an action in the physical world.

A video generator predicts pixels. It can produce a convincing clip of a ball rolling off a table, but generating every pixel is computationally expensive, and visual realism does not guarantee physically correct dynamics. Pure diffusion-based video models typically need many denoising passes per sample, which makes them slow for the rapid simulated rollouts that planning requires.

A JEPA-style world model predicts meaning. It encodes observations into abstract representations and predicts the representation of the future state, ignoring unpredictable detail such as the texture of a wall or the motion of leaves. Because it plans in this compact space, it can evaluate many candidate actions cheaply. That is the architectural bet behind AMI Labs.

"A language model tells you what usually happens. A world model tries to tell you what will happen if you act."

Corazor Editorial

AI Product Engineering

❞

What Has Actually Been Demonstrated

V-JEPA 2 (Meta, 2025) is the most cited result. Pre-trained on more than a million hours of internet video, it reached state-of-the-art results on several video-understanding benchmarks. Its action-conditioned variant was post-trained on less than 62 hours of unlabelled robot video and then deployed zero-shot on robot arms in two different labs, performing pick-and-place tasks specified by goal images—without robot-specific data collection or task-specific training. That is the strongest public evidence that video-trained latent world models can transfer to real robot control.

H-JEPA (AMI Labs with NYU, INRIA Paris and Brown University, October 2026) is AMI's first public research release. It stacks several world models, each predicting further ahead in its own abstract space, and plans top-down: the highest level sets subgoals, lower levels work out the movements. The authors report 73.3% success on Visual AntMaze for a three-level model versus 18.0% for a single-level baseline, with less planning compute. On real robot manipulation video, an extra training objective was needed to stop the model ignoring the moving robot. Code and 93 checkpoints are MIT-licensed. The results are paper-reported across three seeds and not yet independently replicated, and the authors note that more hierarchy is not always better.

Hybrid research such as ThinkJEPA (2026) pairs a vision-language model, which provides broad semantic knowledge, with a JEPA branch for fine-grained motion, and reports better long-horizon forecasting than either alone. This pattern—language model for meaning, world model for dynamics—is where most practical systems are heading.

What has not been demonstrated: a world model running a business process, a general-purpose world-model API with service-level commitments, or peer-reviewed deployments in safety-critical industrial control. AMI Labs itself told TechCrunch in September 2026 that it is in a research and building phase and is not discussing product plans or timelines.

Which Industries Should Pilot World Models Now

Robotics and warehouse automation are the clearest fit. Pick-and-place, navigation, and manipulation are exactly the tasks where latent world models have shown transfer from video to real hardware. Teams with existing robot fleets and logged video can test whether a pre-trained world model reduces the data needed for a new task or improves planning in cluttered environments.

Manufacturing and quality inspection benefit from models that understand normal process dynamics. A world model trained on line video can flag when what happens next diverges from what should happen next—a more general signal than a classifier trained on a fixed list of defects.

Autonomous vehicles, drones, and mobility already use learned predictive models of other agents and occupancy. Research such as diffusion-based occupancy forecasting for driving shows how domain-specific world models are being built; JEPA-style approaches offer a cheaper representation for planning.

Healthcare and sports video are emerging areas. AMI Labs' first announced partner is Nabla, a clinical AI company, and motion analysis in physiotherapy, surgery video, or athletic performance is a natural match for models that understand movement. These domains add regulatory and validation requirements that make careful pilots essential.

Pure software businesses—SaaS, fintech workflows, support operations—should mostly wait. The world-model framing is useful for thinking about agents that predict the consequences of actions, but today those agents are better built with LLMs, deterministic simulators of your own systems, and strong evaluation than with physical-world models.

World Models and AI Agents: Separating the Idea From the Product

Many vendors now describe enterprise agents as having a world model. Usually they mean the agent keeps structured state about the task, or simulates an action before taking it. Those are good engineering practices, but they are not JEPA-style learned world models. A dry-run mode that previews what a refund or deployment will change, a sandbox copy of a database, or a rules engine that checks preconditions all give agents a form of consequence prediction today, with far more reliability than a learned model. Read our production governance playbook for AI agents for the controls that matter before an agent takes real actions.

Where learned world models will matter for software agents is long-horizon planning in environments that change: supply chains, network operations, or markets, where you cannot write down all the rules. That is a credible future direction, not a 2026 procurement item.

How to Scope a World-Model Pilot (Without Burning Budget)

Start with a decision, not a model. Name the operational decision the world model should improve—grasp selection, route choice, anomaly escalation—and the metric that proves it: success rate, cycle time, false alarms, or data needed to learn a new task.

Audit your data before anything else. World models learn from video and sensor streams, ideally with the actions taken recorded alongside. Many companies have cameras but no synchronised action logs. Fixing data capture is often the first month of real work, and it is valuable regardless of which model you eventually use.

Begin from open checkpoints. V-JEPA 2 and H-JEPA publish code and weights, so a first experiment can fine-tune or probe an existing model on your data rather than pre-training from scratch. Set a fixed compute budget and a decision point.

Compare against a strong baseline. A tuned classical controller, a task-specific model, or a vision-language model with rules may already meet the target. The pilot should show that the world model beats the baseline on the metric that matters, at acceptable cost, not just that it works.

Keep humans and deterministic safety in the loop. In physical systems, the world model proposes and a certified controller or safety envelope decides. Log every prediction against the observed outcome; that log becomes your evaluation set and your evidence for the next stage.

A Realistic 90-Day Plan

Days 1–30: define the decision and metric, audit and fix data capture, assemble a held-out evaluation set from real operations, and establish the baseline. Days 31–60: probe or fine-tune an open world-model checkpoint, compare against the baseline offline, and analyse failure cases by condition—lighting, object type, speed. Days 61–90: run a shadow deployment in which the model predicts but does not act, measure prediction accuracy against reality, estimate the operating cost, and make a go, iterate, or stop decision based on evidence.

Frequently Asked Questions

What is AMI Labs?

AMI Labs (Advanced Machine Intelligence) is a world-model research company co-founded by Yann LeCun after he left Meta in late 2025. It raised $1.03 billion in March 2026, has offices in Paris, Montreal, Singapore and New York, and released its first open research, H-JEPA, in October 2026. It has not announced commercial products.

Is a world model the same as a digital twin?

Not exactly. A digital twin is usually an engineered simulation built from known physics and system data. A learned world model infers dynamics from observation. The two are complementary: digital twins are precise where the rules are known, and learned world models can generalise where they are not. Many industrial teams will combine them.

Can a small company use world models?

Yes, for scoped problems. Open checkpoints mean you do not need a billion-dollar training budget to experiment. The main costs are data capture, GPU time for fine-tuning, and engineering to integrate predictions safely into operations.

Conclusion

World models are a real and important direction in AI, not just a funding buzzword—but they are early. The credible evidence today is in video understanding and robot planning, with open code and checkpoints that make serious pilots affordable for companies that work in the physical world. For software-only businesses, the useful lesson is the principle rather than the product: give your agents a way to predict and verify consequences before they act. Pilot where you have data and a measurable decision, compare against strong baselines, and let evidence, not the buzzword, decide what you scale. Plan a world-model or physical-AI pilot with Corazor — we scope the decision, the data, and the evaluation before any model is trained.

Corazor Editorial