Fei-Fei Li's World Labs Unveils Atlas: A World Model for Spatial Intelligence
The woman who taught AI to see is now teaching it to inhabit worlds.
In 2009, Fei-Fei Li released ImageNet -- a dataset of 3.2 million labeled images that became the spark for the deep learning revolution. It taught AI to recognize a cat, a car, a chair. It was a breakthrough in teaching machines to see the 2D world.
Seventeen years later, Li has unveiled something that represents a fundamentally different ambition.
Her startup, World Labs, has released Atlas -- a multimodal world model that generates photorealistic, fully navigable 3D environments from a single photograph. It is not a better image recognizer. It is not a faster video generator. It is a model that attempts to understand the physical world in three dimensions, with physics, perspective, and time baked in from the start.
This is the difference between recognizing a chair in a photo and understanding that the chair sits in a room, has a back side, casts a shadow, and would look different from across the room.
From ImageNet to spatial intelligence
Li launched World Labs in February 2024 with a bold thesis: artificial general intelligence is impossible without spatial intelligence. An AI that can ace every text benchmark but cannot reason about 3D space, object interactions, and physical consequences is, in her view, fundamentally incomplete.
The startup raised $1.2 billion from Nvidia, AMD, and Autodesk before releasing a product. The thesis attracted capital because it identified a genuine gap in the AI landscape. The current generation of AI models excels at processing text and 2D images. They struggle with the physical world -- the kind of reasoning a toddler develops by knocking over a block tower.
Atlas is the first public proof that Li's thesis might be right.
What Atlas actually does
Atlas is a multimodal autoregressive diffusion transformer -- a description that packs a lot of technical ambition into a few words. It processes text, images, video, and 3D data natively, combining them into a shared "spatial context." This is not a video model that has been retrofitted to accept camera coordinates. It is an architecture built from scratch around the idea that every input belongs somewhere in 3D space.
The results are striking:
- Camera-controlled generation: From one to six reference images, Atlas generates new views at any camera angle, with pixel-perfect geometric consistency. It can imagine what lies behind an object, around a corner, or above a roofline -- and it gets the geometry right.
- Spatial reconstruction: Atlas reconstructs real-world scenes from as few as two or three photographs. It outperforms specialist 3D reconstruction models on every standard benchmark, despite being a generalist model.
- Space-time simulation: From a handful of ordinary cell phone cameras, Atlas can freeze time and reframe shots -- the "bullet time" effect without a Hollywood budget. More importantly, it enables Real-to-Sim workflows for robotics.
- Explicit 3D outputs: Atlas can output point clouds and 3D Gaussian splats, making its generations usable in game engines, CAD software, and robotics simulators.
The robotics play
While Atlas generates stunning visual results -- the company's demo reel shows cathedral interiors, candy cities, and medieval town squares -- its real target is robotics training.
The standard approach to training robots involves either expensive physical data collection or manually crafted simulations. Atlas offers a shortcut: film a space with a cell phone, and Atlas reconstructs it as a 3D simulation. A robot can then navigate that space, manipulate objects, and encounter varied lighting conditions and object arrangements -- all generated by the same model that built the world.
This is a fundamentally different approach from the dominant paradigm in AI robotics, which relies on massive curated datasets and hand-coded simulators. Atlas proposes that a single model can generate the environment and simulate the robot's sensor readings within it.
Benchmark performance
World Labs tested Atlas against several of the industry's top models. In blind human evaluations for camera-path adherence, Atlas was preferred over:
- FLUX 3: 93% preference
- Gemini Omni Flash: 81% preference
- MiniMax H3: 75% preference
- Seedance 2.5: 94% preference
On 3D reconstruction benchmarks, Atlas outperformed specialist models including Pi3X, VGGT-Omega 1B, and Depth Anything 3 across six standard datasets (DTU, ETH3D, KITTI, NRGBD, 7-Scenes, T&T, and ScanNet).
The competitive landscape
Atlas enters a crowded and rapidly evolving world model market. Odyssey is building interactive world simulations. Yann LeCun's AMI Labs is focused on physical planning architectures. Niantic is building geospatial mapping systems. What distinguishes Atlas is its ambition to be a single "omni model" that unifies generation, reconstruction, and simulation in one architecture.
The real test will come with wider availability. Atlas is currently in early access for select enterprise partners. The company has not announced a general release date.
What it means
The narrative arc of Fei-Fei Li's career is one of the more interesting stories in modern AI. She created the dataset that taught AI to see. Now she is building the model that teaches AI to inhabit the world it sees. The transition from 2D to 3D understanding is not incremental -- it represents a fundamentally different conception of what intelligence requires.
If Atlas delivers on its promise, the impact will extend far beyond better video generation. It could change how robots are trained, how simulations are built, and how AI systems learn to interact with the physical world. That is the kind of ambition that the $1.2 billion in funding was betting on.