World Models: The Next Big Shift in Artificial Intelligence BY FOM GYEM
Artificial intelligence systems like ChatGPT, Gemini and Claude have become household names. They write emails, answer questions and generate content at impressive speed. But despite the hype, these tools all share a major limitation: they don’t actually understand how the real world works.
What they do is predict words now, a new class of AI is emerging that could change everything. They’re called world models, and according to researchers and industry leaders, they may become one of the most talked-about technologies before the end of the year.
Traditional AI models learn by consuming massive amounts of text books, articles, websites and social media posts. From this, they learn statistical patterns that allow them to predict what word or sentence should come next. That’s why a chatbot can accurately describe what happens when a coffee cup falls off a table: gravity pulls it down, the coffee spills and the cup might break. But the AI doesn’t understand gravity. It has never seen an object fall or collide with the ground. It’s simply repeating patterns it has seen in text.
World models work differently. Instead of learning primarily from text, they are trained on videos, images and simulations, allowing them to understand 3D space, motion, cause and effect, and physical interactions. Rather than predicting the next word, they predict what happens next in the real world. A common analogy used by researchers is this: Text-based AI reads the driver’s manual. World models actually get behind the wheel.
The world’s largest AI companies are now investing heavily in world models.
Google recently unveiled Genie 3, a system capable of generating interactive 3D environments from a single image. Runway, another AI company, launched a world-model-based system that is already being used in Hollywood to generate realistic video scenes.
OpenAI is reportedly working on a successor to its video model, internally referred to as Sora 2, designed to simulate complex physical environments and behaviours.
Yann LeCun, one of the pioneers of modern AI and a Turing Award winner. He recently left his role at Meta to launch a new startup focused entirely on world models, called Advanced Machine Intelligence (AMI Labs). According to reports, the company attracted as much as $5 billion in funding before releasing a product.
Industry experts believe world models could be the key to making AI truly useful beyond screens and chat boxes.
Robots powered by world models could operate safely in unpredictable environments, such as warehouses, hospitals and homes. Self-driving cars could better understand complex situations like school zones or busy intersections.
Tesla has been quietly training its own world models using billions of miles of driving footage. Robotics company Figure AI is applying similar technology to humanoid robots designed to work alongside humans without causing damage.
Even smartphones could soon benefit. With embedded world models, a phone camera might not just identify a leaking pipe, but guide a user step by step through fixing it because the AI understands how plumbing systems actually work.
For years, AI progress has focused on language: chatbots that talk, write and explain. World models represent a shift toward AI that acts systems that can reason about the physical world and interact with it safely.
As competition intensifies and products begin to ship, analysts expect world models to move quickly from research labs into consumer devices, robots and vehicles.
If current trends continue, “world models” may soon be as familiar a term as “chatbot” or “generative AI.” And when that happens, the conversation around artificial intelligence may change for good.
