Cyber 'Brain Platinum' — Can It Cure AI's Amnesia?

Overview

Fei‑Fei Li (李飞飞)'s AI world‑model outfit World Labs today open‑sourced Spark 2.0, a browser‑first renderer that it has been reported can stream scenes with more than 100 million 3D Gaussian "splat" primitives to ordinary devices. The announcement targets a practical bottleneck in photogrammetry‑driven 3D: how do you shove a multi‑gigabyte, tens‑of‑millions‑of‑elements scan into a phone browser and still keep it smooth? Spark 2.0 answers with engineering rather than magic — continuous Level‑of‑Detail, progressive streaming and a GPU virtual‑memory layer.

How Spark 2.0 works

3D Gaussian Splatting (3DGS) turns photos into millions of translucent ellipsoids — "splats" — instead of triangles. That soft, additive representation reproduces foliage translucency and granular textures better than hard‑edged meshes, but files can easily hit tens of millions of splats and gigabytes in size. Spark 2.0 sits on THREE.js/WebGL2 (not WebGPU) and combines a continuous LoD tree, a chunked, random‑access .RAD format for progressive download, foveated rendering, and a GPU page‑table that swaps 64K‑splat blocks in and out of a fixed memory pool. The result: a stable splat budget (e.g., ~500k on mobile, ~2.5M on desktop) regardless of scene size, and background Web Worker threads that fetch and decode the blocks your view needs next.

Why it matters

The tech matters for more than prettier browser demos. High‑quality 3DGS scans are heavy both in bandwidth and GPU memory; Spark's streaming format and virtual memory approach aim to let many independent scans coexist and be stitched into much larger worlds — theoretically without a hard scene size limit. Against a backdrop of export controls, accelerator shortages and uneven WebGPU support that constrain access to high‑end GPUs, solutions that degrade gracefully to WebGL2 and prioritize streaming become strategically important. World Labs frames Spark 2.0 as an open‑source contribution to the ecosystem rather than a closed product push — a nod to the reality that 3DGS adoption requires tooling across capture, compression and rendering.

Will open‑sourcing Spark end the "amnesia" of large, photoreal 3D scenes in AI and AR workflows? It reduces the engineering barriers, but wider adoption — standards, editors, and real‑world bandwidth and device testing — will decide whether billion‑splat worlds become everyday rather than experimental.