Core argument: Spatial intelligence is the ability to perceive, reason about, and act in three-dimensional space and time. Every person who has ever navigated a city has it. Almost no AI system does — because language models learn from text, and text flattens the world into words. Fei-Fei Li calls teaching machines this ability the next frontier of AI, and her company’s $8.2 billion acquisition is the biggest bet yet that she is right.
This is the ninth chapter of an ongoing series on world models. Previous chapters: why LLMs hallucinate, world model vs LLM, how world models work, who is building them, what artificial general intelligence really means, when it might arrive, and the forces shaping AI’s decade. The series hub holds the full map.
1. The short answer
Spatial intelligence, in the AI context, means a machine’s ability to perceive, reason about, and act in three-dimensional space and time — to understand where objects are, how they relate to each other, and what happens when things move.
The term is borrowed from psychology. Howard Gardner’s theory of multiple intelligences, introduced in 1983, listed spatial-visual intelligence as one of the core human capacities: the ability to think in images and pictures, to visualize accurately and abstractly. It is the intelligence of architects, surgeons, pilots, and anyone who has parallel-parked a car.
Fei-Fei Li — the creator of ImageNet, often called the godmother of AI — argues that this human capacity is what AI is missing. Language models learned to talk about the world from text; her stated goal, in her essay “From Words to Worlds,” is machines that understand and act in the world itself. Her company, World Labs, builds systems that generate and reason about 3D environments. In September 2026, AMD agreed to acquire the company for roughly $8.2 billion.
The gap she is pointing at is real. Ask a language model how to rotate a sofa through a doorway and it will produce confident, grammatical advice — because words about sofas are everywhere in its training data. Whether the advice survives contact with an actual sofa and an actual doorway is a question the model has no way to check.
2. The person who read the map versus the person who walked the city
There is an old distinction in how humans hold spatial knowledge, and it divides people cleanly into two kinds.
The first kind has read the map. They can name the districts and describe the city’s layout accurately — from above, as lines on paper.
The second kind has walked the city. They know which corner is loud, which alley smells of rain, how the light falls at five in the afternoon. Their knowledge is not diagrammatic. It is a feeling for the place, built from moving through it.
A language model is the map reader. Everything it knows about cities came as text — flattened, secondhand. Spatial intelligence is the attempt to build the walker: a system whose knowledge of space comes from three-dimensional experience rather than description.
The analogy breaks in one place that matters. The walker’s knowledge updates itself: every step returns feedback from the world, whether the walker wants it or not. A model’s spatial knowledge only updates when someone feeds it new data. Building a machine that walks — or something functionally like it — is the entire problem, and nobody has fully solved it.
3. Why this is hard for AI
The reason is the same one this series keeps returning to: the data channel.
Language models learn from text, and text is a lossy compression of the world. “The cup is next to the laptop” carries almost nothing a machine can act on — where is the laptop, on what surface, what happens if the cup is pushed? Human readers fill those gaps automatically because they have bodies. Models cannot; for them, the sentence is all there is.
Getting spatial ability into a machine requires data that carries structure, not just description: images and video with depth, 3D reconstructions of real places, sensor traces of objects moving and colliding. That data is scarcer and heavier than text, and harder to collect — which is why text intelligence took off a decade before spatial intelligence, even though the need for the latter is older than writing.
It also explains a subtle failure of today’s models. Asked about objects and space, they answer from statistics — what words usually go together — rather than from any representation of a room. The answers sound right because word patterns track reality most of the time. When the patterns miss, the model has nothing underneath to catch the mistake.
4. How machines get spatial intelligence
Three routes to machine spatial ability are being built right now, and they map onto the technical approaches this series covered in chapter 3.
Generating 3D worlds. World Labs’ first product, Marble, launched in November 2025: text or images in, a persistent, explorable 3D world out — stored in a format game engines and design tools can use. Its successor Atlas, announced in September 2026, folds generation, reconstruction, and simulation into a single model. The wager is that building worlds teaches systems what worlds are like.
Reconstructing the world from images. A parallel effort teaches systems to infer full 3D structure from flat photographs — the way you can look at one snapshot of a room and know what is behind the camera. Deep 3D reconstruction from limited views is one of World Labs’ core research claims.
Learning to act in simulation. The robotics route: train systems inside simulated 3D environments until their spatial skills hold up in the physical world. This is where world models — systems that learn how the world behaves by predicting it, the subject of chapter 3 — do the teaching, and where the reliability gap between simulation and reality remains the open problem.
The three routes are converging. A system that can generate a 3D world, reconstruct real ones from photos, and act inside simulations has, piece by piece, assembled something like spatial intelligence.
5. Who is betting, and how much
The spatial intelligence bet now has a price tag. AMD’s agreement to acquire World Labs — announced in late September 2026, with Fei-Fei Li joining AMD as chief scientist — values the idea of machine spatial understanding at roughly $8.2 billion before the deal has even closed. The rationale is physical AI: robots and vehicles that operate in the world need systems that represent the world.
The bet is also a vote about sequence. Li’s position has been consistent since her ImageNet days: first give machines the right perceptual foundation, then the capabilities follow. ImageNet did that for visual recognition fifteen years ago — the dataset that made computers see. Spatial intelligence is her argument that the same move, played in three dimensions, will make machines act.
6. My take
Three readings, in descending order of confidence.
First, high confidence: the human-versus-machine distinction here is real and underrated. Text knowledge and spatial knowledge are different materials, and the AI industry spent a decade building only the first. The second is now being built, and the people doing it are credible.
Second, medium confidence: the term will stay contested. “Spatial intelligence” in AI borrows prestige from a human-psychology concept while naming something narrower — machine representations of space. That is not dishonest, but readers should notice the borrowing. The machine version is, for now, a long way from the human one.
Third, medium confidence: the near-term payoff is industrial, not personal. Robot arms that load dishwashers, vehicles that navigate loading docks, design tools that turn photos into buildable spaces — the first customers are factories and warehouses, with design studios close behind. Consumer uses will come later, and probably through the game and design tools rather than through any single assistant.
7. Questions people ask
Is spatial intelligence the same as a world model? They overlap. A world model is the general idea — a machine’s internal picture of how the world behaves. Spatial intelligence is the specific, human-inspired name for the part of that picture dealing with space, distance, and structure. Think of world models as the field, and spatial intelligence as one of its central competencies.
Do current chatbots have any spatial intelligence? Fragments. They can describe spaces, reason about simple arrangements from text, and pass some visualization tests — all from word patterns. Under pressure, the pattern knowledge fails in ways a child’s would not: mirror-image confusion, invented layouts, objects that teleport between sentences. The description of space is not the possession of it.
Why did AMD agree to pay $8.2 billion for this? Because physical AI — robots, autonomous systems — needs environments to train and act in, and World Labs builds the environments. For a chip company, owning the workload is the point: the machines that run spatial AI are machines that buy chips.
Sources
- a16z, “The Frontier of Spatial Intelligence with Fei-Fei Li”, podcast discussion with Fei-Fei Li and Justin Johnson — “perceive, reason, and act in 3D space and time”.
- Simply Psychology, “Gardner’s Theory of Multiple Intelligences”.
- Northern Illinois University CITL, Gardner’s multiple intelligences teaching guide — spatial-visual definition.
- Fei-Fei Li, “From Words to Worlds”, Substack essay.
- World Labs, “Marble: A Multimodal World Model”, November 2025; “Atlas: A World Model for Spatial Intelligence”, September 2026.
- TechCrunch, “AMD will acquire Fei-Fei Li’s World Labs for $8.2 billion”, September 28, 2026.
- CNBC, “AMD to acquire Fei-Fei Li’s World Labs”, September 28, 2026 — coverage carrying the “godmother of AI” designation.
Next in this series: the Sora question — does video generation count as understanding the world? Or start from the series hub.