Video generation models like Google DeepMind's Veo are laying the groundwork for embodied AI that can operate in the physical world, not just digital spaces. While most people see AI-generated videos as entertainment, researchers argue these same technologies are essential for developing robots and AI systems that can learn from real-world interactions the way humans do.

How Did Video Generation Become the Key to Building Smarter Robots?

The breakthrough came from a fundamental shift in how video models work. Until recently, AI video generators like Google DeepMind's Veo 1, released in May 2024, could only create short clips of fixed length from a single text prompt. This meant longer videos had to be stitched together from multiple clips, creating visible inconsistencies. In July 2024, researcher Diego Marti Monso and collaborators at MIT published Diffusion Forcing, a new training method that allowed video models to generate videos of any length, removing that constraint entirely.

This breakthrough sparked a cascade of improvements. DeepMind released Veo 2 in December 2024 and Veo 3 in May 2025. OpenAI followed with Sora 2 in September 2025. Other open-source models like Sand AI's MAGI-1 and Skywork AI's Skyreels v2 were trained directly using Diffusion Forcing. The entertainment industry quickly adopted these tools to produce films, advertisements, and even video games.

But the real impact extends far beyond entertainment. The infinite-length breakthrough sparked a revolution in world modeling, a concept that transforms video models into interactive simulators. A world model is a video model trained to respond to inputs like video game controls or robot commands. Marti Monso demonstrated this with Open Dreamer, a model trained on recordings of humans playing Minecraft that learns to simulate the game in real time. The key insight: if you train such a model on the physical world instead of a digital game, you create a reactive intelligence capable of adapting and operating in real environments.

Why Are Video Models Better Than Language Models for Building Physical AI?

Today's most familiar AI systems are Large Language Models, or LLMs, like ChatGPT. These models learn from vast amounts of text data, absorbing facts and patterns of human knowledge. However, researchers argue that video models are fundamentally better suited for understanding the physical world. The reason is simple: text is expensive to produce and limited in what it can convey.

"Text is actually quite expensive to produce. Imagine how long it would take you to write a full description of a day in your life without skipping any detail. Now, consider how much easier and richer it would be to use a camera to record your day instead," explained Diego Marti Monso, one of the inventors of algorithms powering AI video models.

Diego Marti Monso, Researcher at MIT Computer Science and Artificial Intelligence Laboratory

Videos pack physical information far more densely than text. As the saying goes, a picture is worth a thousand words. This architecture is much better suited for the real world. It's far easier to predict where a baseball will go by watching its arc than by reading a written description of its trajectory. Videos capture the cause-and-effect relationships that matter for physical interaction.

Steps to Understanding How Video Models Lead to Embodied AI

• Video Generation Breakthrough: Models like Veo 3 can now generate videos of unlimited length, moving beyond the short, stitched-together clips that plagued earlier systems and enabling more realistic, continuous simulations.
• World Modeling Integration: Video models are being transformed into interactive simulators that respond to inputs like robot commands or game controls, turning passive video generation into active decision-making systems.
• Physical World Training: When world models trained on digital environments like Minecraft are adapted to learn from real-world robot recordings, they become reactive intelligences capable of understanding and navigating physical spaces.
• Vision as Foundation: Video models prioritize visual information the way humans and animals do, incorporating life experiences into decision-making through observation of interaction results, which is essential for embodied AI development.

The connection between AI video generation and physical robotics reveals something profound about how intelligence works. Marti Monso argues that true artificial general intelligence, or AGI, requires more than digital capabilities. AGI refers to AI systems that can solve any general task the way humans do. Today's AI excels in white-collar digital work but falls short in managing businesses or performing blue-collar tasks. There are no generally useful robots that humans can reliably deploy.

The missing piece is embodiment. Giving AI systems a physical form to interact with the real world is not optional for true AGI; it's essential. The AI-generated videos people scroll past on social media are actually byproducts of AI learning how the physical world works. As researchers continue improving the algorithms that generate these videos, the path toward embodied AI that inhabits human spaces becomes clearer.

What appears to many as "AI slop" may actually represent the foundation of the next generation of intelligent systems. The same techniques that create entertaining deepfakes and viral videos are simultaneously teaching machines to understand physics, causality, and interaction. This dual purpose means that every improvement in video generation brings us closer to robots and AI systems that can learn, adapt, and assist humans in the physical world.