AI News

Developer uses Claude Fable 5.2 to generate 97KB underwater simulator Abyssal Garden

cnBetaUpdated Nov 16 21:05

AI summary

Developer Noniv revealed on Reddit that an underwater simulator named Abyssal Garden was written entirely by Anthropic's Claude Fable 5.2, with only 97KB of code, using no external textures, maps, or 3D model assets, and without using Web engines such as Three.js or Babylon.js. The project is based on WebGPU and WGSL shaders, generating 2 million kelp leaves, tens of thousands of rocks, and tens of thousands of fish in real time at startup, with ambient sound effects synthesized in real time by the Web Audio API. In measured tests, it reached 180 FPS at 1440p on an RTX 5070 Ti, 248 FPS at maximum quality on an RTX 4090, and about 20 FPS at low quality on a Samsung Galaxy A16; development took three days, with actual code processing taking about 2 to 3 hours, and another AI sub-agent was introduced to audit rendered screenshots to identify visual defects.

Why read this

AI score 51

The original article presents a pure code generation path with no external assets and no mainstream engine, along with measured frame rates, which can be used to observe the efficiency limits of WebGPU procedural generation.

Topics

Tags

Body · AI translation

According to the developer Noniv on the Reddit community, this project named Abyssal Garden was written entirely by Anthropic's Claude Fable 5.2 model.

Unlike traditional game development, this project did not use any external textures, maps, or 3D model assets, and even completely abandoned mainstream web game engines such as Three.js or Babylon.js.

The core technology of this program lies in low-level calls to the WebGPU interface and the WGSL shader language. All visual elements, including 2 million kelp leaves, tens of thousands of rocks, and thousands of swimming fish, are all generated in real time by code when the program starts.

Even the ambient sound effects are synthesized in real time through the Web Audio API, rather than playing pre-recorded audio files.

Measured data demonstrates the operating efficiency of WebGPU. On an RTX 5070 Ti platform, a performance of 180 FPS at 1440p proves the advantages of procedural generation.

The developer also provided a screenshot of operation on an RTX 4090, showing that the frame rate can reach 248 FPS at the highest image quality. For mobile users, a Samsung Galaxy A16 phone can barely run at about 20 FPS at low image quality.

Despite the stunning visual effects, the developer said this was not achieved overnight. The entire development process lasted three days, with actual code processing time of about 2 to 3 hours, during which multiple iterations were made to add collision detection and an image quality settings menu.

Moreover, the developer also introduced another AI sub-agent to audit rendered screenshots, specifically for identifying visual defects and optimizing GPU time consumption.

Source

cnBeta