The 4-Bitter Lesson: NVFP4 4-bit RL training recipe

Blog post details a stable 4-bit RL training recipe using NVFP4 quantization on NVIDIA Rubin GPUs, achieving up to 9x more ops/sec than BF16. It addresses the throughput-stability tradeoff by using NVFP4 for both weights and activations, unlike prior methods that only quantized weights.