Gemma 4 Hits 255 tok/s on WebGPU via Agentic Kernel Writing

Custom WebGPU kernels reportedly pushed Gemma 4 inference to 255 tok/s in-browser, up from 84 tok/s, using an agentic optimization loop. The demo and kernels are claimed to be released publicly. The surrounding narrative about safeguard rollbacks is satirical, but the kernel optimization approach and performance numbers are the substantive claim.