Language models are reaching the edges of their training diet. After years of feeding them the internet, performance gains now arrive in fractions, not leaps. Synthetic data helps extend their reach, models generating training material for other models, but this mostly refines familiar capabilities. It’s sharpening the same knife, not forging a new tool.
World models point elsewhere. They don’t try to predict the next word, they try to predict the next state of the world. That means physics, causality, spatial relations, the connective tissue of reality. Humans don’t learn language in isolation. We learn it while moving through environments that constantly reinforce cause and effect. World models aim for that same grounding.
The most surprising progress is coming from video. Tools like the newly released Sora 2 by OpenAI and Google’s Veo 3 make polished clips, but the breakthrough is in coherence: liquids flow like liquids, gravity works as gravity, collisions look right. These models are internalising dynamics.
Google’s Genie 3 pushes this even further. It takes a rough sketch or a few frames of video and spins them into an interactive game world, complete with agents that move, react, and collide in believable ways. Genie 3 is wild precisely because it blurs the line between video generation and world simulation. It demonstrates that models can learn not only how the world looks, but how it can be inhabited.
Scaling this kind of learning is the real challenge. A studio’s film library, or even the whole internet of video, can’t cover the infinite variety of physical interactions. That’s why simulation platforms matter. NVIDIA Omniverse is an example: a synthetic environment where infinite scenarios can be generated, edge cases tested, and models taught how reality operates without waiting for the data to exist in the wild.
Language models will remain essential, but world models open the next chapter. When linguistic skill is fused with physical, spatial, and causal reasoning, AI shifts from text mimicry to deeper comprehension. The slowing progress of LLMs may be less an ending than the forcing function that drives us into this more ambitious terrain.
