a16zWhy World Models Could Change Robotics, 3D, and Creativity
a16z
Introducing Atlas 0:00
World Labs has launched Atlas, a new frontier world model capable of generating, reconstructing, and simulating the world. Given an image and a camera trajectory, it can produce video frames from any chosen perspective. It can also perform sparse 3D reconstruction, taking anywhere from one to a hundred views of a real space and turning them into either a flythrough video or an explicit 3D model. One striking demonstration involves recreating the famous bullet time effect from The Matrix, where Neo appears frozen in midair while the camera circles him. That shot originally required hundreds of cameras arranged in a ring on a green-screen set. With Atlas, the same effect can be achieved using just three ordinary cameras, such as iPhones on tripods, filming something like a basketball shot or a strawberry dropping into milk, then freezing time and flying the camera around the scene.
New view prediction as a core idea 3:00
The most fundamental thing Atlas does is called new view prediction. Just as language models are built on predicting the next word and video models are built on predicting the next frame, Atlas is built on predicting a new viewpoint. Multiple views or descriptions of a scene are fed in to form what the team calls a spatial context, an implicit understanding of what the world looks like. From there, a virtual camera can be pointed at any position in space and time, and Atlas will generate what that view should look like.
Why this differs from other video models 4:00
Unlike typical video models that loosely interpret a single image or a start and end frame, every image fed into Atlas carries an exact three-dimensional camera pose. This grounding means four views of a room, one from each corner, can be reconstructed with high precision rather than guessed at. It also allows for creative flythroughs built by deliberately staging images from different sources, giving directed control over where the camera travels, rather than relying on repeated text-prompt attempts as with typical generative video tools.
Merging generation and reconstruction 5:00
Atlas is described as new partly because it performs generation and reconstruction jointly within a single model, something computer vision has historically treated as separate subfields. To achieve this, the model was built to be natively multimodal from the start, working across text, images, video, camera poses, and depth maps, which describe the spatial structure of a scene. Fei-Fei Li emphasizes that this unification of pixel generation and pixel reconstruction is unprecedented in a field that has existed for more than half a century, historically split into separate tracks at computer vision conferences for generation, recognition, and reconstruction.
Spatial intelligence and the road ahead 8:30
Spatial intelligence, as described here, must eventually let a system generate a space, reason within it, and edit or interact with it, ultimately extending into a fourth dimension of time. The core requirement is understanding geometry, structure, and physics well enough to render, simulate, and plan actions. Atlas marks a major step because every frame now carries an estimated viewpoint or camera pose, the most critical piece of geometric information, though higher fidelity simulation and dynamics over time remain further down the roadmap.
From Marble to Atlas 11:00
World Labs has existed for two and a half years, and last year released an earlier world model called Marble, which powers a product generating 3D worlds from images, video, or text, outputting Gaussian splats. While useful for rendering efficiently on devices and integrating with game engines, splats became a bottleneck, so Atlas was redesigned around new view prediction as its core primitive, generating RGB frames and 3D data that can still produce splat worlds when needed but aren't limited to that format.
Dense versus sparse reconstruction 13:30
Traditional dense 3D reconstruction requires capturing nearly every surface from multiple angles, since anything unphotographed becomes a hole in the model. Capturing a single room properly this way might take an hour for a trained person, or two hours for a first-time user covering a multi-room space, often needing hundreds of photos. Atlas aims to cut that down by fifty to a hundred times, reconstructing a scene from as few as three images by combining triangulation from visible points with generative filling-in of unseen gaps. A Stanford quad demo showed a full aerial reconstruction built entirely from ground-level photos, with everything beyond direct sightlines generated according to reconstruction principles.
Scaling context like language models 19:00
The team compares this to how language models expanded context length over the years, eventually handling long documents reliably. Reconstruction is described as generation with a very long context: Atlas can take a 64-image capture of a house and produce a flythrough nearly indistinguishable from one built with 2,000 images, something the earlier Marble model could never do since it only handled a couple of images at once.
Confidence in scaling further 21:00
The team held early conviction in two hypotheses, the general scaling law and the new-view-prediction paradigm, though they weren't certain how well or how fast it would work. Each time the model was made bigger and trained longer on more chips, it improved significantly, and the version shown publicly is the largest trained so far. They believe they are still at the beginning of this scaling curve, with training compute rather than architecture currently the main limiting factor.
The Overnight Discovery That Sealed The Decision 23:31
During training, Justin's team fed an early, smaller version of the model into viewpoint generation using the famous garden table scene familiar from Nerf papers. Overnight, Ben sent a Slack message showing the camera flying under the table and revealing a soccer ball that was genuinely part of the original scene, not something invented. The next morning, the three of them looked at each other and made the decision within five seconds to build this into a full product, because no one had ever produced a result like it before.
From Marble To Grounded Creative Workflows 25:01
Ben explains that the earlier product, Marble, was often used by people who wanted a full 3D scene from a single image just so they could grab a few extra viewpoints as screenshots, accepting that quality would degrade. That use pointed directly toward generative view synthesis as a real need. Creative professionals rarely rely on one monolithic model; they pull images from favorite tools, combine them as key frames in video software, then edit everything together. What they consistently want is a stable, persistent 3D world to ground their work in, something people are used to from decades of building physical stages, props, and sets for film, marketing, and games. Ben also points to industries like architecture, construction, and even conference booth fabrication, where translating verbal feedback or sketches into a 3D design remains slow and labor-intensive, something he sees AI as well suited to fix.
Why Atlas Matters For Robotics 29:00
Fei-Fei explains that the company's robotics division, formerly Sceneics, specializes in going from the real world into simulation and back again, so that a robot can be trained safely before deployment. Reconstructing real environments densely from photos is slow and laborious, and this bottleneck limits how fast robotic simulation can proceed. Atlas offers a faster path for that reconstruction step. She stresses that the biggest problem in robotics right now is data, not chips, and that training requires not just captured environments but randomized variations of them, such as cables bending differently or boxes changing size and color.
Simulation As The Missing Ingredient 33:00
Ben notes that robotics policies are fundamentally different from other AI outputs like images, code, or video, because a policy acts in the world and must handle unexpected responses rather than simply producing a static artifact. Training such policies requires exposure to every plausible scenario, which is where simulation becomes essential, either through classical physics engines coded by hand or through data-driven neural simulators trained on real data. Fei-Fei adds that once a learned simulator truly understands how the world responds to actions, it could become the planner itself, since predicting a response to an action is closely tied to imagining which action produces a desired response.
Dynamics Already Live In The Model 35:31
Responding to feedback that the model needs more motion, the team reveals that Atlas already supports dynamics in its architecture and training data, unlike the fully static Marble. Subtle motion already appears in released examples, including water waves and small moving cars in aerial views. They explain that even to produce clean static 3D reconstructions, it helps to train on dynamic data so the model learns to factor out motion, and the current release simply emphasized spatial over temporal behavior during post-training, with more dynamic capability planned ahead.
Control And Editability As The Real Frontier 38:00
Ben says the most exciting future direction is deeper multimodal control, or editability, letting people specify changes intuitively without sacrificing output quality, a balance that image models are approaching but video and world models have not yet reached. Fei-Fei frames this as closing the loop between seeing, experiencing, and interacting, calling it the natural next rung for spatial intelligence.
Next Viewpoint Prediction As AI Complete 40:30
Fei-Fei introduces the idea of AI completeness, comparing it to Turing completeness, where solving one task in full generality would solve any intelligence problem, citing next token prediction as the classic example from language models. She argues that generative next viewpoint prediction in Atlas is similarly AI complete, framing it in evolutionary terms: animals evolved eyes because movement reveals new viewpoints, while trees, which do not move, never needed them. She concludes that next viewpoint prediction stands as the spatial equivalent of next token prediction.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
