AI NewsDev toolsAnnouncement

Claude Code wrote a WebGPU ocean renderer, 61 KB single HTML file

A single-file WebGPU ocean renderer written by Claude Code Opus 5.5, with FFT waves, whitecaps, physically based sky and clouds, now at 854 GitHub stars and shipped as a 61 KB HTML file.

AI News

Editorial2 min read

LinkedInX

Why it mattersA coding agent can produce a 61 KB file with Tessendorf FFT waves, LEAN mapping and a CDLOD quadtree inside it, which raises the bar on what counts as a hobby render and what counts as graphics engineering.

A WebGPU ocean that opens in Chrome from a single 61 KB HTML file has crossed 850 GitHub stars, and the repository README says every line was written by Claude Code running on Opus 5.5. Noniv/web-ocean, released under the MIT licence, draws Tessendorf FFT waves, whitecaps, sun glitter, physically based sky with clouds and a free camera, and reports 854 stars with 133 forks on 22 October.

The author, who signs as Noniv, says the full version is 102 KB and the minified build is 61 KB, and that the file runs directly from a file:// URL with no server.

What the model wrote

The technical list the README gives sits well above a typical demo. It names six FFT cascades of 256 by 256 samples each for the wave spectrum, LEAN mapping for the slope roughness that keeps the surface readable from a distance, a CDLOD quadtree with 17 levels for mesh detail that falls off with distance, Hillaire's atmospheric model with transmittance lookups for the sky, a two-layer procedural cloud volume, and HDR output with 4x MSAA and bloom. Wavelengths run from about 1 km down to millimetres.

Each one of those names a paper. None of them is a plug-in library in the WebGPU ecosystem, so the ocean is written out in shaders in that single HTML file.

How the author kept Claude Code honest

The README calls out a loop the author built around the agent. Rather than let Claude Code evaluate its own screenshots, which the author says the model is bad at, the project uses what it calls critic subagents: separate prompts whose only job is to look at the current frame and say what is wrong with it. The author then fed the critic's answer back to the writer. That is the first time this AI News has reported a hobby graphics project naming self-evaluation of a model's output as the specific thing that failed, with the fix stated in the same README.

What the repository does not say

The README does not state a build time, a number of prompts, or a token bill, so this is a visible result rather than a measured engineering comparison. It also does not benchmark against an engine such as Three.js or Babylon.js on the same wave spectrum, so the 61 KB file size is a size figure rather than a performance figure.

For a team wondering how far a coding agent can push a self-contained demo this year, open the page: the ocean is what fits.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project