Today in AI: World Models Have No Recipe Yet

There is no consensus on what a world model actually is. Justin Johnson, co-founder of World Labs, told TWIML that the industry still lacks an established recipe for building them.

This admission is significant. While language models eventually converged on transformers and next-token prediction, spatial AI has not found its footing yet.

The full discussion can be found on TWIML's YouTube channel.

Johnson views the field as an attempt to build systems that can understand, generate, and simulate environments rather than just describing them. This represents the frontier beyond language that Fei-Fei Li has been championing as spatial intelligence for robots and agents.

What defines a world model?

Johnson describes world models as theory builders. They maintain a state of the world and predict how that state evolves when an action is taken. He bases this on POMDPs, a framework where an agent can't see the absolute ground truth and must instead infer it.

The field is currently split between explicit and implicit approaches. Explicit models use geometry, such as Gaussian splats, to ensure consistency by design. Implicit or generative models learn physics and rendering without a fixed 3D representation. While the latter scales more easily, it often suffers from drift.

How Marble creates 3D worlds

Marble processes images and other inputs to produce navigable 3D worlds. Instead of just providing novel views, it creates spaces you can actually move through. Johnson noted that the system relies on explicit representations like Gaussian splats for consistency, then uses generative priors to fill in the gaps where data is missing.

Training data and output representations are still major bottlenecks. Evaluation is even more difficult. There is currently no ImageNet equivalent to judge whether a generated world is physically plausible or stable over long planning horizons, which remains a significant hurdle.

World Labs isn't the only player in this space. Runway, Luma, and DeepMind's Genie line all claim progress in world models, though most of these rely on video demos. Marble is attempting to ship actual geometry that you can simulate in.

The ultimate goal is a unified model. The industry is looking for a single system that can render, plan, and simulate without needing to swap architectures or loss functions for every individual task.