Last time, 5,000 particles in live force fields. This time, over 1,000,000 strands. Two ranch alpacas, a mother and her cub, in a meadow where every strand of fur and every blade of grass answers the same live wind, in real time. Built by AI agents on Wave Engine, our framework on top of @threejs. The strand count isn't the story. How little the agents had to write, and how little they had to figure out, is. To capture that, we coined two measures of our own: FLTC, framework-level token compression: far less code to write. FLRC, framework-level reasoning compression: far less to reason through. Details below.

The numbers: about 2,000 lines of Wave code for the whole scene, with every model and material generated in code except the imported alpaca. By our estimate, roughly 80% less reasoning (FLRC) than the same scene in raw three.js. Most tools buy savings like that by packaging features for one genre. It works until you step outside the genre. Wave has no fur system and no alpaca feature. One wind moves the fur, the grass and the weathervane; the same camera direction would shoot any subject. The agents composed this from general building blocks, and this demo uses a small slice of them: thousands of generalizations, refined through years of accumulated experience and reinforcement learning. That is why it compounds. Every scene an agent builds on Wave starts with that experience behind it. More soon. #Wave3DAI #threejs #Web3D #WebGPU