Google DeepMind just dropped a model that, in my opinion, is a real step-change for “world models”: D4RT (4D reconstruction through time). In this video I break down what it actually does—taking a plain 2D video and reconstructing a 3D scene over time (that’s the 4th dimension). I walk through the core capabilities: point tracking (following pixels across frames even when they get occluded), point cloud reconstruction (freeze time and rotate the 3D scene), depth understanding, and camera pose estimation (changing viewpoint beyond the original camera path). For robotics, AR, and self-driving, this is exactly the kind of persistent scene understanding we’ve been missing. Then I get into the “secret sauce” behind the speedup—up to 300x faster than prior approaches. Older pipelines stitched together separate models for depth, motion, segmentation, plus extra optimization, which is slow and can introduce artifacts like ghosting. D4RT instead encodes the whole video once into a global scene representation, and then answers thousands of independent queries like: “Where is this pixel at time t in 3D, from this camera angle?” Because queries run in parallel and space is decoupled from time, you can scale from one point to a million points without compute exploding—while keeping sharp details (they even embed a 9x9 patch per query for crisp boundaries).