The second wave of AI is reaching the factory floor

Artificial intelligence is entering a new phase—one that extends beyond language and into the physical world. Recent announcements from leading technology companies highlight a major shift in industrial AI, with the focus moving from text-based assistants to systems capable of understanding, predicting, and interacting with real-world environments.

In June, NVIDIA introduced Cosmos 3, a physics-based world model designed to enable AI-powered robots to better understand and navigate their surroundings. Around the same time, Siemens unveiled Intelligence Center X, a platform that connects AI agents with production teams to scale industrial AI beyond pilot projects. Earlier, Google revealed its own world model initiative, Gemini Omni, alongside plans for major investments in AI infrastructure.

While these announcements came from different companies, they point to the same direction: the future of AI depends on its ability to understand the physical world—not just generate text.

From Language to Physical Intelligence

Large language models (LLMs) such as ChatGPT, Gemini, and Claude have transformed how people interact with AI. They excel at summarizing information, answering questions, and generating content by learning from enormous collections of text.

Manufacturing, however, presents a different challenge.

Factories operate in dynamic environments where machines move, materials shift, and tiny variations can determine product quality. In these settings, AI must understand spatial relationships, motion, force, and cause-and-effect—not simply describe them.

This has accelerated interest in world models, AI systems designed to build an internal understanding of physical environments. Rather than predicting the next word in a sentence, world models predict what is likely to happen next in the real world.

Why Traditional AI Has Limits in Manufacturing

A language model can explain how a production line works or recommend process improvements based on historical data. But when unexpected events occur—such as a misaligned component, a slipping robotic gripper, or a machine operating outside normal conditions—it lacks a genuine understanding of physical interactions.

In other words, it has learned about physics through text but has never experienced it.

World models address this limitation by continuously observing their surroundings, learning how objects behave, and simulating possible outcomes before actions are taken. This allows AI systems to anticipate problems instead of merely reacting after they occur.

For example, if a component begins falling onto a conveyor belt, a language model may simply describe the incident. A world model could predict the object's trajectory, recognize that production will be disrupted, and adjust the robot's movements to prevent the error before it happens.

A Growing Debate in AI

The emergence of world models has also sparked debate among AI researchers.

Yann LeCun, one of the pioneers of modern artificial intelligence, argues that today's language models have significant limitations when it comes to reasoning, planning, and understanding the physical world. He believes future AI systems require entirely new architectures capable of modeling real-world interactions.

Others, including Greg Brockman, maintain that increasingly capable language models remain the most promising path toward more general AI capabilities.

For manufacturers investing in AI today, the distinction matters. Choosing between language-based AI and physics-aware AI could influence automation strategies for years to come.

What Makes a World Model Different?

Unlike traditional language models, world models create an internal simulation of reality.

They learn how objects move, how forces interact, and how actions produce consequences. Instead of responding only after receiving new information, they continuously predict future states of the environment and adjust accordingly.

This predictive capability is particularly valuable for robotics, autonomous manufacturing, and industrial automation, where real-time decision-making is essential.

Early research already suggests significant benefits. Studies indicate that robotic systems trained using world-model approaches can achieve performance improvements of up to 30% compared with systems that learn directly from raw sensor data, offering gains in efficiency, accuracy, and adaptability.

Defining the Next Generation of Industrial AI

As interest grows, researchers are also working toward standardizing what constitutes a true world model.

One recent open-source initiative, OpenWorldLib, proposes that a genuine world model must do more than generate realistic images or simulations. It must perceive its environment, interact with it, retain memory of previous states, and continuously learn from real-world feedback.

By this definition, current text-to-image and text-to-video generators, while impressive, do not qualify because they cannot adapt to changing physical environments.

The research also suggests that today's large language models may eventually evolve toward these capabilities, although significant technical challenges remain.

The Factory of the Future

The next generation of industrial AI will not simply answer questions or generate reports. It will understand machines, predict physical events, and collaborate with robots and workers in real time.

As companies invest heavily in world models and physical AI, manufacturing is entering a new era—one where intelligence is measured not only by what AI can say, but by what it can anticipate, understand, and do on the factory floor.