LLM AssBench: AIs build 3D buttocks to compare skills

Common AI benchmarks lean on math problems or coding tasks, but widely known problems may already sit in training data, making it hard to measure ability accurately. LLM AssBench sidesteps that by asking for something with no standard answer: a single HTML file containing a realistic, operable 3D buttocks model built with Three.js, complete with skin texture, drag-to-rotate and click-to-jiggle behavior.

Luigi points out that small AIs can sometimes produce 3DCG close in quality to large ones, and argues that judging performance requires looking beyond model size to how accurately a model handles a task suited to its purpose. The planned 40-point scoring is intended to make it easier to pick an AI that delivers sufficient quality while keeping costs down.