CUDA Agent is the first known RL-trained model to surpass Claude Opus 4.5 (and Gemini 3 Pro) in CUDA kernel generation! ByteDance's CUDA Agent uses **agentic reinforcement learning** to train a model that automatically generates high-performance CUDA kernels, rewarding directly with real GPU profiling speed—breaking away from conventional approaches. Key benchmark data on KernelBench: - Simple/medium kernels (Level-1/2): **100%** faster than torch.compile - Complex kernels (Level-3): **92%** faster rate - Overall: **96.8%** faster rate vs torch.compile, with ~2.11× geometric mean speedup - Outperforms strongest proprietary models like Claude Opus 4.5 and Gemini 3 Pro by about **40%** on the hardest Level-3 tasks (where those models only beat torch.compile ~66–69% of the time) The true ceiling for AI hardware isn't the silicon itself—it's the "**software unlocking + optimization closed loop**" capability. Combining this with the simultaneous Apple ANE breakthrough: - Apple M4 ANE: **6.6 TFLOPS/W** (~80× more efficient than A100), with hundreds of millions of devices sitting idle; the bottleneck is Apple's closed APIs + abstraction layers (CoreML hides 2–4× real throughput) - NVIDIA GPUs: RL agents learn "**extreme optimization under real hardware feedback**," proving learned strategies can crush static rules/compilers The performance moats of hardware giants (Apple/NVIDIA) are being **double-killed by AI**: reverse engineering smashes closed APIs (turning idle chips into compute farms), while RL squeezes every last drop from existing GPUs. In the future, the real choke point won't be compute hardware—it's who masters the "**hardware-native feedback + autonomous learning optimization**" closed loop first. By combining soft and hard tactics, whoever flips existing device performance 2×, 10×, or more can progressively dismantle the giants' walls. This compound growth creates speeds beyond human intuition: from 10× → 100× → 1,000× within a few years. The era of **on-device training** (ANE side) + **cloud/edge extreme inference** (CUDA Agent side) is accelerating fast. AI can now "**self-optimize**" close to theoretical peaks. The untapped potential in hundreds of millions of idle Apple devices + massive NVIDIA cards is being collectively kicked open by independent hackers, companies, and researchers.(常为希 🔸🚢币安人生（Ai奇点）)