World Models: How AI Is Learning to Understand Reality Instead of Text

At its May conference, MIT Technology Review discussed world models, which may be the next big leap in the development of artificial intelligence. Editor-in-chief Mat Honan and senior AI editor Will Douglas Heaven examined how companies are trying to teach neural networks not just to process text, but to truly understand the world around them.

What are world models

A world model is not just another version of an LLM. It is a fundamentally different kind of system that can watch video, analyze images, interact with its environment and predict the consequences of actions. Think of a person who sees a cube at the edge of a table and understands that it will fall. The neural network has to acquire that understanding without explicit instructions, by learning it from observing the physical world.

Such models change the training paradigm. Instead of the classic setup, "here is a text, answer the question," a new one appears: "watch the video, predict what happens next." That requires a completely different architecture, a different dataset and a different way of measuring the model's errors.

Why text is clearly not enough

Modern large language models are champions at processing information, but they are blind in the literal sense. They know about gravity only because people have written about it millions of times online. But they have never seen an object fall, never felt inertia, never experimented with physics. This creates specific blind spots:

• They cannot predict physical interactions from first principles
• They get confused by spatial relationships between objects in video
• They cannot grasp cause-and-effect links across a sequence of frames
• They cannot plan actions grounded in real physics
• They make errors when predicting trajectories and collisions

The limitation is especially visible when AI tries to control a robot, plan logistics or predict the consequences of manipulations in the real world.

Who is working on world models

OpenAI, DeepMind, Tesla and other major companies are actively investing resources in developing world models. Their approaches vary. OpenAI and DeepMind work with video datasets from YouTube and with synthetic simulations. Tesla uses millions of hours of video from its cars' cameras to teach its system to see the world the way people on the road do.

Some companies start with supervised learning on labeled video. Others use reinforcement learning in controlled simulations, where the model can make a million mistakes without real-world consequences, gradually refining its understanding.

What this means

If companies manage to scale world models as successfully as they scaled LLMs, AI moves to a new level: from symbolic processing of information to something closer to a genuine understanding of physical reality. Robotics will leave the lab. Autonomous systems will become more reliable. Planning of complex processes will speed up.

But this is still the beginning of the road. MIT Technology Review is drawing attention to this because world models are probably the most important direction in AI for the next several years. The companies that are first to teach neural networks to see and understand the world will gain a huge competitive advantage.