The Map Is Becoming a Copy of the World

A single phone camera can now rebuild a room in 3D at 20 frames per second. Live, over long sequences, end-to-end. No lidar. No rig. One camera, streaming.

That sentence would have sounded absurd two years ago. Spatial capture meant expensive lidar-plus-RGB scanners, careful sweeps, and hours of post-processing before you had anything explorable. As recently as this spring, the honest assessment in the field was that distribution of 3D Gaussian splats was pretty much solved (you could stream a big indoor scan into a browser and walk through it in real time) while capture remained the blocker. Getting good results still required hardware most people will never touch.

That blocker is dissolving. And it's worth slowing down to see what happens when it's gone.

The capture curve, compressed

Here's the trajectory, all within roughly the last eighteen months.

Niantic put collaborative splat capture on phones and a shared map, letting anyone capture anything in 3D and pin it to a place. World Labs' Marble went from generating 3D scenes you could stitch into expansive environments to reconstructing real-world locations from a handful of images — sixteen iPhone photos of a Shinjuku alley became a fully explorable 3D world in minutes. NVIDIA's Lyra removed the manual stitching step for large-scale explorable worlds entirely. An open-source repo now takes any city name and spits out a 3D model with buildings and streets pulled from OpenStreetMap, MIT-licensed, no login. SuperSplat added automatic voxel collision, turning any splat into solid, walkable geometry at 5 cm resolution — more compressed than a mesh, but solid enough to move through.

And in June, Apple announced at WWDC that Gaussian splatting is coming to Apple Maps.

Each of these is a modest engineering step. Together they describe one curve: the cost of turning a real place into an explorable digital one is falling toward zero, and the device doing it is the one already in a few billion pockets.

What people do with free capture

You don't have to speculate about demand. People are already crowd-sourcing real places into explorable splats — cathedrals, alleys, plazas — and sharing them in tools like SuperSplat. Walk into St. Stephan's from your desk. It's uneven and hobbyist right now, the way early Wikipedia was uneven and hobbyist. But the shape is visible: a next-gen map you can step inside, built bit by bit by whoever happens to be standing there with a phone.

The historical rhyme is Street View. Google spent years driving camera cars down every road on Earth to build a photographic index of the planet. That was a centralized, capital-intensive project because capture was expensive. When capture costs one camera and a few seconds, the same index gets built as a side effect of people existing in places. Apple putting splats into Maps is the institutional signal that this is where mapping goes next: from pictures of places to the places themselves, in geometry.

The second customer: machines

There's a quieter implication under the consumer one. A live, single-camera 3D reconstruction of a space is exactly what a robot needs to understand the room it's moving through. Perception has been one of the stubborn bottlenecks in embodied AI, and the hardware answer has often been lidar: precise, expensive, power-hungry. If a commodity camera stream can produce dense, persistent 3D structure in real time, the perception stack for household and warehouse robots gets radically cheaper.

And the crowd-sourced map feeds back into this. A robot entering a building that has already been captured doesn't start from zero. It loads the space. Machine-readable 3D models of cities are the connective layer between the world of bits and the world of atoms — I wrote in 2024 that I could no longer walk around a city without seeing that model in my head. The difference now is that the model is starting to actually exist, and ordinary phones are building it.

From picture to copy

A useful way to hold this: maps have always been compressions. A road atlas compresses geography into lines. Street View compresses it into photographs. A splat map barely compresses at all — it keeps the geometry, the light, increasingly the physics. At some point the representation is dense enough that the distinction between "a map of the place" and "a copy of the place" stops mattering for most uses. Tourism, real estate, insurance, robotics, games, memory itself: all of them run on the copy once it exists.

We've watched this pattern before. First we map, then we predict, then we act. Spatial computing spent a decade stuck at expensive, partial mapping. The capture cost just collapsed, and the mapping phase is about to finish itself, one phone at a time.

The interesting question is no longer whether a walk-in copy of the world gets built. It's who holds it, and what gets permission to move through it.