World Labs
Fei-Fei Li's spatial-intelligence startup, building models that reason about three-dimensional space rather than text or flat images.
Mission
To give AI spatial intelligence — the ability to understand and generate three-dimensional worlds.
Founded By
Headquarters
San Francisco, California, USA
Founded
2024
Status
Active
Contribution to AI
Language models arrived with a settled idea of what their data was; three-dimensional space had no such thing, and the first useful move was assembling the people who had built the representations. Ben Mildenhall co-authored NeRF in 2020, which taught a field to store a scene's geometry and appearance in the weights of a network; Christoph Lassner's Pulsar renderer prefigured the Gaussian splatting that displaced it. Putting both inside one commercial group in 2024 shifted neural rendering from reconstructing scenes that exist to generating ones that do not. The December 2024 demo made the distinction visible: a single photograph became a scene you could walk through in a browser, everything outside the original frame invented, and — unlike the video-frame world models of that year, which forgot a room the moment the camera turned — it stayed put. Persistence became the technical claim the company was judged on. Marble, released generally in November 2025, took text, images, video or a rough 3D layout and returned worlds exportable as Gaussian splats or meshes, which meant generated environments entered Unreal, Unity and Blender as ordinary assets rather than remaining trapped in a vendor's viewer. RTFM, shown a month earlier, generated interactive frames on one H100 without any explicit geometry at all. The $230 million raised at a billion-dollar valuation on emerging from stealth also settled a funding question: spatial intelligence counted as a frontier, not a niche.
Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).