World models might outweigh LLMs in the next decade

LLMs (along with Reinforcement Learning) have been the biggest disruptor in the artificial intelligence landscape so far, powering chatbots, coding assistants, and frontier AI labs like Anthropic, OpenAI, Gemini, and others. However, I like to believe there is a technology that might outweigh LLMs in the next 10 years: World Models. I will dive into them in this brief article.

To put it very simply, World Models focus on understanding the mechanics of an environment, i.e., simulating a sequence of events and simulating the physical consequences to plan the next action. I believe this is a direct replica of how humans learn from the real world; for example, let's say I am holding a glass, and I know if I drop the glass, it'll shatter.

I know this because I understand basic physics and gravity. Basically, before the human brain acts, it runs a split-second simulation in its brain to predict what the next state might be. This is exactly what World Models do: build a representation of an environment (following the law of physics) and predict how the environment will change given specific actions.

The first time I learned about World Models, I wondered why one should use them since reinforcement learning does the same. Needless to say, I did not understand what World Models exactly were back then.

In traditional RL, the model goes through millions of iterations to reach the optimal state. For example, without a world model, a drone might need to crash into trees hundreds or even thousands of times before it learns how its actions affect flight. But with a world model (such as an internal physics engine), the drone can perform most of its learning in simulation.

By refining its policy within the virtual environment, it enters the real world with a much better understanding of how to act, which reduces the amount of costly and risky trial-and-error required in reality.

World Models use 3 distinct components to work; they are the Encoder, the Predictor, and the Controller.

Firstly, the Encoder takes raw sensory data (like video frames) and compresses it to build an abstract space that only focuses on objects and physics. Frequently uses Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), or Joint Embedding Predictive Architectures (JEPAs).

Next, the Predictor utilizes the current abstract state and predicts what the next state of the space might look like, given a specific action. The Predictor layer is often built with RNNs (Recurrent Neural Networks) to understand temporal and spatial relationships.

Lastly, the Controller uses the internal simulations to pick the optimal flow of actions (without fear of failure). This is the layer that utilizes the framework of Reinforcement Learning.

Just as LLMs have completely disrupted the tech landscape by 2026, the commercialization of World Models will soon do the same for the physical landscape. Equipping autonomous robotics with the ability to simulate and predict their environments will unlock massive advancements in defense, manufacturing, logistics, and healthcare. I look forward to the continued progress in self-supervised learning and world modeling.

Anyways, thanks for reading, and have a great day.