Fei-Fei Li, the Stanford computer scientist who essentially created modern computer vision, has quietly launched a $230 million startup dedicated to solving one of AI's biggest blind spots: the inability to understand three-dimensional physical space. While the world focused on chatbots and image generators, Li recognized that the next frontier in artificial intelligence requires something fundamentally different from large language models (LLMs), which are AI systems trained on vast amounts of text. Her new company, World Labs, is building what researchers call "world models," generative systems that can perceive, reason about, and predict how the physical world actually works.

What's the Difference Between ChatGPT and Spatial AI?

The distinction matters enormously for what AI can actually accomplish in the real world. A traditional language model can describe the steps to perform surgery in perfect detail. But it has no internal simulation of how tissue resists a scalpel, how blood flows, or how a surgeon's hand should adjust in real time based on what they feel. A spatial AI system, by contrast, builds a continuous 3D model of the environment and can predict physical interactions as they happen.

Consider a simple test: ask ChatGPT to describe what happens when you tilt a glass of water on a table 45 degrees. The model can write a paragraph about gravity and physics. Now ask it to predict the trajectory precisely enough that a robot hand could catch the glass mid-fall. It cannot. Current language models lack what researchers call "spatial intelligence," the ability to perceive 3D structure, reason about how objects relate to each other in space, and predict how the physical world will change as things move, fall, collide, or interact.

"Building spatially intelligent AI requires something even more ambitious than LLMs: world models, a new type of generative model whose capabilities of understanding, reasoning, generation and interaction with the semantically, physically, geometrically and dynamically complex world are far beyond the reach of today's LLMs," stated Fei-Fei Li, founder of World Labs.

Fei-Fei Li, Founder and CEO of World Labs

How Is Spatial AI Being Built Right Now?

The progress in this space over the past 18 months has been startling. What was mostly academic research two years ago is now appearing in production systems that real people are using. Multiple companies are racing to deploy generative world models, each taking a different approach to the same fundamental challenge: teaching AI to understand and simulate physical reality.

• Google DeepMind's Genie 3: This system generates not just images but entire explorable 3D environments from text descriptions. Give it "a misty rainforest with ancient stone ruins and soft morning light filtering through the canopy," and it creates a physically consistent world where objects behave realistically, lighting changes as you move, and surfaces have appropriate properties.
• Niantic Spatial's Large Geospatial Model: The company behind Pokémon Go built the world's largest real-world spatial dataset from years of augmented reality gameplay. Their Large Geospatial Model uses this unprecedented corpus of location data to give AI precise, verified understanding of actual places, combining simulated worlds for training with real-world data for deployment.
• 4D World Models: The most technically ambitious work adds time as a fourth dimension. Current video AI struggles with "object persistence," where a dog might lose its collar mid-scene or a chair might change size between frames. 4D models maintain a persistent internal representation of every object, its identity, physical properties, and position across the entire duration of a generated scene.

Niantic projects that by the end of 2026, the most capable AI systems will "navigate our streets, factories, and homes using a shared understanding of space." This is not marketing language; the technology is already in production.

Why Does This Matter Beyond Research?

The economic implications are staggering. According to Citigroup, 80 percent of global economic activity depends on the physical world, including logistics, construction, and transportation. Spatial AI unlocks applications that language models simply cannot handle: robots that navigate dynamic factory floors, AI glasses that understand your surroundings in real time, autonomous vehicles that predict how pedestrians will move, and surgical robots that adapt to tissue resistance during procedures.

Li's decision to raise $230 million specifically for world models signals where the AI industry believes the next breakthrough will occur. The talent and capital moving in this direction simultaneously is a signal worth taking seriously. Li herself is uniquely positioned to lead this effort. She earned a bachelor's degree in physics from Princeton University and a Ph.D. in electrical and electronics engineering from the California Institute of Technology. She is the inventor of ImageNet and the ImageNet Challenge, a critical large-scale dataset and benchmarking effort regarded as one of the three driving forces of the birth of the modern AI and deep learning revolution.

Beyond her technical credentials, Li serves as the inaugural Sequoia professor of computer science at Stanford University and is the founding co-director of Stanford's Human-Centered AI Institute. She was also named one of the most compelling voices at the intersection of AI, innovation, and the future of business, recognized as a keynote speaker for academic conferences, healthcare and technology events, and audiences interested in AI ethics and responsible development.

How to Understand the Practical Impact of Spatial AI

• Robotics and Automation: Boston Dynamics robots now use spatial world models to understand spatial relationships, predict collisions, and perform complex tasks in dynamic environments. A robot that understands 3D space can navigate a warehouse, avoid obstacles, and manipulate objects without constant human supervision.
• Augmented Reality and AI Glasses: Spatial AI enables AI glasses to understand your physical surroundings in real time, providing contextual information, navigation assistance, and interactive overlays that respond to the actual 3D environment you are in.
• Autonomous Systems: Self-driving vehicles and delivery robots require precise 3D understanding of streets, pedestrians, and obstacles. Spatial AI provides the foundation for these systems to predict how the world will change as they move through it.
• Virtual and Gaming Environments: Generative world models like Genie 3 can create interactive, physically consistent 3D environments from text descriptions, enabling game developers and content creators to generate entire worlds rather than building them manually.

The shift from language-based AI to spatially intelligent AI represents a fundamental change in what artificial intelligence can accomplish. While chatbots excel at pattern matching and text generation, world models enable AI to interact with the physical world in ways that require genuine understanding of how objects move, collide, and interact. Fei-Fei Li's World Labs is betting that this capability will define the next era of AI, just as ImageNet defined the era of deep learning that preceded it.