There are well-known signs that today’s “smart” models do not truly understand the world—they merely talk about it skillfully:
1. Chess
While LLMs are familiar with chess theory, they often make completely illegal moves: leaving their own king in check, jumping over pawns, or simply suggesting moves that the rulebook does not allow.
2. Maps and Routes
Vision-Language Models (VLMs) and LLMs were given the task of planning a route on a building’s floor plan: “Go to the backyard while avoiding obstacles.” The models regularly plotted paths through walls and pillars, or started off in entirely the wrong direction.
3. High School Math
When tested on graduation-level word problems, several leading LLMs—even when they arrived at the correct final answer—produced solutions full of faulty steps, logical leaps, and nonsensical assumptions. They lack a coherent “internal model” of the subject.
What’s the Difference? Words vs. World Model
The issues above do not stem from malfunction; rather, today’s LLMs predict words, not model the world.
A true world model works like an internal simulator: it learns how the environment changes in response to actions and “runs” possible future scenarios ahead of making decisions.
Yann LeCun’s (ex-META) JEPA (Joint Embedding Predictive Architecture) uses sensor data (images, video, audio, measurements), to develop compact, abstract representations of the real world, predicting their future states, and the agent plans on this internal model—not reproducing pixels or words, but states and consequences.
Consensus Among Visionaries
It’s no coincidence that LeCun, Schmidhuber, Bengio, Friston, and Gary Marcus all say: the next big leap will not be bigger LLMs, but a better world model.
In the short term, these are mainly being built by universities and major research labs, but companies in manufacturing, logistics, robotics, and simulation-based decision support should already be designing their digital twins and data strategies as if tomorrow they would want to connect their company’s “physics” to such a world model.
Conclusion
LeCun is now betting billions that the future is not about ever-larger talking parrots (LLMs), but about world models that understand physics—academia, big labs, and capital are already lining up behind this idea.
The question is not whether they will be right, but whether your company’s data and digital twin strategy is ready for the moment when models that understand the world arrive alongside language models.
Leave a Reply