Solving Poker with Custom WebGPU
Summary
A developer recounts building an in-browser poker solver using custom WebGPU kernels instead of relying on a tensor library. The post covers the CFR-based approach, the motivation to avoid large libraries, and how Codex helped generate efficient kernels, achieving significant speedups. The project is open-source and demonstrates that bespoke, verifiable kernels can outperform general-purpose libraries in specific tasks.