Why World Models Might Shape the Future of AI

If you regularly use AI tools like ChatGPT or Gemini, you’ve probably marveled at their power … and then been surprised when they make a major, easy-to-spot mistake.

Like telling people that it’s a good idea to eat one small rock every day.¹ Or being unable to count the number of r’s in the word strawberry.² But there’s a reason why such advanced technology sometimes does something jaw-droppingly dumb.

The most popular AI tools today operate on large language models (LLMs). While LLMs can be extremely useful for coding, writing and other tasks, some experts believe this type of AI will always face certain limitations due to its intrinsic design.

Yann LeCun, a leading AI developer and critic of LLMs, has argued they aren’t even smarter than a cat.³

“We are used to the idea that people or entities that can express themselves, or manipulate language, are smart — but that’s not true,” said LeCun, who helped lead Meta Platforms’ AI efforts for about a decade before leaving in 2025.

“You can manipulate language and not be smart, and that’s basically what LLMs are demonstrating.”

Instead, more companies are developing “world models” designed to understand and interact with the real world. Think of it as AI with common sense.

Proponents of world models think they will be critical for embodied AI, or AI that interacts with physical environments. World models could serve as the brains for self-driving vehicles, humanoid robots and other devices, allowing them to navigate streets and factories effortlessly.

Major technology firms such as Alphabet and NVIDIA are dedicating resources to world model research, as are several well-funded startups.

World models are still in their early days, and their developers are working to solve challenges around power and capacity. However, we believe this emerging type of AI may hold significant potential for investors, especially if world models accelerate the development of next-generation technology.

To gain a better sense of this trend, let’s take a closer look at how world models and LLMs function.

How Do Large Language Models Work?

Large language models, the most popular type of AI in use today, are fed (or trained on) massive amounts of text.

By analyzing all that material, an LLM learns what words and phrases typically appear together and in what context. It looks for patterns in its giant storehouse of training data.

Ask an AI-powered chatbot a question – “What is a cat?” – and it will almost always provide a correct answer. “A cat is a small, furry animal that’s typically kept as a household pet.”

While an LLM can define “cat,” it doesn’t truly understand what a cat is. The model simply knows that a string of characters like “small, furry animal” is a statistically likely answer, based on the data it has consumed. (Image generators like Midjourney and DALL-E work a little differently, but they operate on the same basic principle. They’re trained to recognize patterns in existing images and video, which enables them to create similar output.)

LLMs possess significant capabilities, especially in language-related work. They can draft detailed reports, dash off emails and help build apps.

However, because they lack practical knowledge of the world, LLMs may struggle in other areas, such as reasoning and planning.

Fei-Fei Li, an AI pioneer and co-founder of World Labs, argues that AI needs to develop spatial intelligence, which she calls the ability to understand and interact with the world.

“For computers to have the spatial intelligence of humans, they need to be able to model the world, reason about things and places, and interact in both time and 3D space,” Li wrote in a 2025 Economist essay. “In short, we need to go from large language models to large world models.”⁴

How World Models Work and How They Differ from LLMs

If LLMs are book-smart, then world models are street-smart.

World models simulate an environment using video, 3D scans, sensor data and other information. This environment could be something as small as your living room. Given enough data and computing power, it might be as sprawling as an entire continent’s weather system.

World models analyze massive collections of training data to learn (or infer) the physics of a simulated environment — the cause-and-effect rules of how things work.

And that’s really important because, once a world model understands the rules of an environment, it can also apply them to new or unexpected situations.

Given a goal, the world model can create multiple simulations to plot out the steps required to achieve it. It can plan and react to changes in the environment.

How Would a World Model Perform in Real-World Scenarios?

Imagine a humanoid robot trained to lift packages from a shipping pallet and carry them across a factory floor. Then the robot places these items on a conveyor belt.

One day, the robot finds a tall stack of packages piled haphazardly and looking like they’re about to fall over.

The robot hasn’t faced a situation exactly like this before. However, because the bot operates on a world model, it has been trained on thousands of similar simulations. It knows that it must rotate the top box a certain way, or the stack might tip over.

The robot never freezes or waits for assistance. Its world model helps it adapt and continue.

That’s also the goal for autonomous vehicles. World models could help self-driving cars respond more effectively to surprises such as traffic jams, aggressive drivers and severe weather.

Improving the performance of embodied AI could be critical for boosting consumer and business adoption of robots, self-driving cars and other technologies.

Where World Models Could Be Applied Next

World models could help create virtual copies, or “digital twins,” of real-world streets, factories and other assets. Users can then use those twins to predict problems, experiment with changes and optimize operations.

For example, Cadence Design Systems has introduced a tool that allows users to build digital twins of data centers.⁵

Designers and engineers can use this tool to create plans for real-world data centers and, even better, forecast how they will operate once built. This could help operators ensure that any new centers meet performance standards even before construction starts.

Video game publishers are also experimenting with world models. World models’ ability to create virtual environments might speed up the development of new games.

Google, for example, recently released Project Genie, a world model that lets anyone create their own interactive environments with basic prompts.⁶

This puts pressure on established game publishers, although firms like Roblox and Unity also have their own expertise in this area.⁷ New world models might make them more productive.

What Could Hold Back World Models?

Resource Demand

World models appear to be more resource-intensive than LLMs. They may require more data, more computing power and more electricity.

Many world models need to be updated in real time, which is another resource demand. For example, self-driving cars must constantly adjust their models as they roll down the street, encountering other vehicles and pedestrians. It’s a huge computing task.

A world model’s data is also multimodal, encompassing different types of data. Getting video, 3D scans and other inputs to work together poses a significant technical challenge.

Data Quality

Like LLMs, world models depend on the quality of their underlying data. They could still make mistakes or hallucinate (presenting false information as fact).

If a model’s training data is incomplete or if it overemphasizes certain situations, the model won’t provide precise, accurate answers. A weather-focused model trained only on rainy days, for example, would be useless on sunny days.

Competition from LLMs

LLM proponents argue that LLMs show signs of learning how to reason as they absorb more data and scale up their computing power.

This phenomenon, also known as emergent reasoning, could help correct some of LLMs' flaws, moving them closer to world models over time.

As they take on new types of data or add memory, newer LLMs might learn from experience in a way that earlier LLMs haven’t. If LLMs continue to make progress, its backers say, world models may not be as important.

It may not be an either-or situation. Instead, consumers and companies might use both types of AI models for different kinds of jobs.

What’s Next for AI World Models?

Nobody knows exactly when world models will break through as LLMs have.

We believe this technology could help other products and services reach critical mass in the marketplace, especially in robotics and autonomous vehicles. We also see potential in companies that help users create digital twins and virtual environments.

In our view, world models are an important part of the overall AI story. They could have a major impact on AI’s ability to generate results for consumers, businesses and investors.