Blind test: JPEG XL art before and after RLHF

Each round shows two images. Each one is a JPEG XL tree program of at most 256 bytes, written by a 3M-parameter model running in your browser and compiled by libjxl as WebAssembly. One image comes from the model before RLHF, the other from the same model after 820 RLHF steps on one person's 1,442 picks. Pick the one you like. After 10 rounds you see which was which.

How the model was trained, and what this test can and can't tell you: the write-up. Code: github.com/Nominy/jxl-art-rlhf.