World models represent a fundamental shift in AI research, moving beyond language-focused systems to create AI that can simulate and interact with physical environments in real time. Over the past year, major AI labs and startups have announced breakthroughs in this emerging category, with companies like World Labs and Advanced Machine Intelligence (AMI) each raising approximately $1 billion in funding, signaling serious investor confidence in the technology.

What Exactly Are World Models, and How Do They Differ From ChatGPT?

World models are fundamentally different from the large language models (LLMs) that have dominated AI headlines since ChatGPT's release. While LLMs process and generate text through turn-based interactions, world models aim to create continuous, real-time simulations of environments. The distinction matters because it opens up entirely new applications.

"The key things that would distinguish it from an LLM are demonstrating degrees of spatial and, maybe for lack of a better word, continuous understanding," said Ben Mildenhall, co-founder of World Labs. "A very distinguishing aspect of interacting with an LLM is they are turn-based. Something that would define a world model is the degree of freedom that you have in interacting with the spatial world, where you do not have this mediated linear journey."

Ben Mildenhall, Co-founder of World Labs

Vincent Sitzmann, an assistant professor at MIT who leads the Scene Representation Group within MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), describes a world model more simply: any model that takes in an interaction and enables you to simulate what would happen next in some environment. Fei-Fei Li, computer vision pioneer and co-founder of World Labs, has outlined three criteria that define a world model: they can generate worlds with perceptual, geometrical, and physical consistency; they are multimodal by design; and they can output the next states based on input actions.

Why Are Major AI Researchers Betting Against Language Models?

The shift toward world models reflects growing skepticism about the long-term potential of LLMs among some of AI's most influential figures. Yann LeCun, former chief AI scientist at Meta, has been particularly vocal about this perspective.

"The idea that you're going to extend the capabilities of LLMs to the point that they're going to have human-level intelligence is complete nonsense," said Yann LeCun, former Meta chief AI scientist.

Yann LeCun, Former Chief AI Scientist at Meta

Even Clem Delangue, CEO of Hugging Face, a major platform hosting LLM repositories, has expressed similar concerns about the current trajectory of language-focused AI. He noted that while LLMs have captured attention, they represent only a subset of AI's potential applications.

"I think we're in an LLM bubble, and I think the LLM bubble might be bursting next year. But 'LLM' is just a subset of AI when it comes to applying AI to biology, chemistry, image, audio, and video," said Clem Delangue, CEO of Hugging Face.

Clem Delangue, CEO of Hugging Face

What Recent Breakthroughs Have Emerged in World Models?

The past several months have seen a flurry of announcements from major AI organizations entering the world models space. These developments have moved the technology from pure research into commercial applications with real-world use cases:

• Google DeepMind's Genie 3: Unveiled in August, this model builds real-time interactivity on top of a video generation foundation, allowing users to interact with simulated environments dynamically.
• World Labs' Marble: Introduced in November, this model and toolset allows users to generate immersive environments that can be exported as 3D assets, accepting input in the form of text, images, video, or other assets.
• Runway's GWM-1: Announced in December by video generation and filmmaking AI company Runway, this is a trio of specialized world models built on the company's existing video model expertise.
• Advanced Machine Intelligence (AMI): Yann LeCun's newly founded company is betting that the future of AI lies in models that interact with or simulate the physical world, rather than processing language alone.

The funding rounds supporting these efforts have been substantial. World Labs and AMI each raised approximately $1 billion in February and March respectively, while Runway secured $315 million in February. This level of investment reflects confidence that world models represent a genuine shift in AI's trajectory.

How to Understand the Practical Applications of World Models

While some of the excitement around world models centers on building foundations for artificial general intelligence (AGI) or superintelligence, most researchers are focused on concrete, near-term applications that could deliver value within the next few years:

• Robotics Training and Testing: World models can simulate environments where robots learn to navigate and perform tasks before being deployed in the physical world, reducing training time and costs.
• 3D Asset Generation: Game developers and filmmakers can use world models to generate detailed 3D environments and assets from text descriptions or images, accelerating creative workflows.
• Scientific Simulation and Modeling: Researchers can use world models to simulate complex physical systems, chemical reactions, and biological processes, enabling faster hypothesis testing and discovery.
• Virtual Environment Creation: World models enable the creation of interactive, navigable virtual spaces that respond to user input in real time, opening new possibilities for training, entertainment, and design.

Why Is "World Models" Such a Confusing Term?

One challenge in understanding the world models landscape is that the term itself lacks a universally agreed-upon definition. Vincent Sitzmann acknowledged this ambiguity directly, noting that "world models" is definitely an overloaded term. Different researchers and companies emphasize different aspects of what makes a model a "world model," which can create confusion about what exactly is being discussed.

Runway's definition emphasizes the internal representation aspect: "an AI system that builds an internal representation of an environment and uses it to simulate future events within that environment." The company further clarified that general world models aim to represent and simulate a wide range of situations and interactions, like those encountered in the real world. This broader framing helps explain why the category encompasses such diverse applications, from robotics to creative tools.

The phrase itself is not entirely new; it has long appeared in reinforcement learning and robotics research to describe models that predict environment dynamics. What is genuinely new is the attempt to scale that concept into general-purpose, generative systems trained on massive visual and multimodal data, combined with the massive funding rounds and commercial interest now surrounding the field.

As world models continue to mature and attract investment, they may fundamentally reshape how AI systems interact with the physical world. Whether they ultimately deliver on their promise remains to be seen, but the convergence of research breakthroughs, substantial funding, and practical applications suggests this is a trend worth watching closely in the coming years.