CUDA Agent is the first known RL training model, surpassing Claude Opus-4.6 and Gemini 3 Pro in CUDA kernel generation! CUDA Agent trains models with agentic RL to automatically generate high-performance CUDA kernels, using real GPU profiling speed as a reward signal, breaking conventions Take a look at the following data: KernelBench benchmark: Simple/medium kernels are 100% faster than (torch. com) pile, and complex kernels are 92% faster Overall 96.8% faster rate vs (torch. com) pile, far exceeding Claude Opus 4.5/Emini 3 Pro (about 40%) The true ceiling of AI hardware is the ability to "unlock software+optimize closed loops", not just the chip itself. Combined with the simultaneous occurrence of the Apple Ane incident: Apple M4 ANE: 6.6 TFLOPS/W (≈ 80 times A100), hundreds of millions of devices are idle, and the bottleneck is the closed API+abstraction layer (CoreML blocks 2-4 times throughput) NVIDIA GPU: RL Agent learns' ultimate optimization under real hardware feedback ', proving that the learned strategy can defeat static rules The performance moat of hardware (Apple/Nvidia) is being killed by AI's "reverse engineering+RL optimization" - the former breaks open closed APIs to turn idle chips into computing farms, while the latter uses reinforcement learning to squeeze every drop of existing GPU performance. The bottleneck of the future is not computing hardware, but who first masters the closed-loop of "hardware native feedback+self-learning optimization", using both software and hardware. Whoever can double the performance of existing devices can gradually break through the walls of giants. This compound growth will create a speed that is difficult for human intuition to easily perceive: within a few years, it can expand from 10 times to 100 times → 1000 times The era of on device training (ANE side)+cloud/edge extreme inference (CUDA Agent side) is accelerating, and AI itself can "self optimize" to near theoretical peak. The potential of billions of idle Apple devices and massive NVIDIA cards is being collectively kicked open by independent/corporate hackers/researchers.