The three distinct bets behind AI "world models"

Everyone in AI suddenly agrees that the next big thing is world models. The catch is that nobody agrees on what the phrase actually means.

Over the past few weeks, Fei-Fei Li's World Labs launched Marble, DeepMind showed off Genie 3 and SIMA 2, and reports emerged that Meta chief AI scientist Yann LeCun is preparing to leave and start his own world model company. Same buzzword, three completely different bets.

Three visions pretending to be the same thing

The phrase world model actually dates back to 1943, when Scottish psychologist Kenneth Craik argued that the human brain runs a small-scale simulation of reality to plan ahead. Today, it has become AI's favorite marketing umbrella.

World Labs, which raised 230 million dollars, dropped Marble as a multimodal world model. In practice, it takes text prompts, single images, short videos, or blocky layouts and turns them into walkable 3D scenes. Users can tweak environments inside a browser editor called Chisel and export assets as 3D Gaussian splats or traditional meshes, backed by their own Spark Three.js renderer.

Hacker News and Reddit users caught on immediately, pointing out that Marble is essentially a polished 3D Gaussian splatting asset pipeline for game engines, not an embodied robot brain.

LeCun is chasing the exact opposite. Rooted in control theory and his 2022 paper on autonomous machine intelligence, his JEPA architecture does not care about rendering pretty graphics. Its world model lives entirely inside latent representations, predicting how compressed internal states will evolve so an agent can reason and plan ahead.

DeepMind sits in the middle. Genie 3 generates 720p interactive video on the fly without a game engine, while SIMA 2 provides an embodied agent that can play and learn inside these virtual sandboxes.

Why it matters

If you build games, VR experiences, or interactive web pages, World Labs gives you a genuinely useful asset pipeline today. Exporting splats straight to Three.js cuts down environment prototyping time. But if you are building autonomous robotics, Marble will not help you solve real-world physical reasoning.

For the industry, this split proves that top researchers believe LLMs are hitting a wall. Language alone cannot model the physical universe. If LeCun indeed leaves Meta to launch an independent startup, it signals that leading minds are ready to bypass generative tokens entirely in favor of latent prediction architectures.

We think branding a 3D Gaussian splat pipeline as a world model is a brilliant fundraising move that comes with serious technical debt. Looking photorealistic is completely different from understanding commonsense physics. Right now, it looks like the industry is selling high-end 3D graphics under the promise of artificial brains, and closing that gap will take much more than neat browser demos.