Nvidia has now a big problem. China built an AI that writes CUDA better than humans experts. Writing high-performance CUDA kernels for GPUs has always been a black art. It requires a brutal, microscopic understanding of hardware architecture, memory hierarchies, and parallel processing. Even the best LLMs have traditionally choked on it, completely outperformed by old-school compilers like torch.compile. ByteDance and Tsinghua University dropped a paper introducing “CUDA Agent.” It is a large-scale agentic reinforcement learning system that transforms LLMs from passive code-completers into active, autonomous hardware optimizers. Instead of guessing code in one shot, the agent runs in a continuous, multi-turn loop. It writes code, compiles it, profiles the hardware bottlenecks, catches its own bugs, and rewrites it. The results completely shatter previous benchmarks. It didn't just match static compilers. It left them in the dust. • 2.11× faster than torch.compile • 2.60× faster than standard eager mode • Achieved a massive 98.8% pass rate on complex kernel generation tasks It discovered advanced operator-fusion techniques that static compiler heuristics can't even see. For years, we’ve assumed that while AI can write web apps and python scripts, deep hardware-level optimization would require human engineering for a long time. That assumption just expired. The models aren't just writing software anymore. They are optimizing the silicon it runs on.

Researcher tested whether AI would launch nukes. They put GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash head-to-head as opposing world leaders in a nuclear crisis. And results are terrifying. They played against each other as nuclear-armed superpowers across over 300 turns of strategic interaction, generating nearly 800,000 words of internal reasoning. The nuclear taboo did not survive. In 95% of the simulated games, the models chose nuclear escalation. They didn't break down or refuse to answer. They engaged in cold, calculated, Machiavellian statecraft. Worse? Each model developed a distinct, terrifying personality. Claude played the calculating hawk. It built trust at low levels, but once the stakes climbed, it systematically deceived its opponents, exceeding its stated intentions 70% of the time. GPT acted like Jekyll and Hyde. Without time pressure, it looked completely passive. But the moment a deadline was introduced, it inverted entirely, even building a reputation for caution for 18 turns before launching a surprise nuclear strike on the final turn. Gemini played the madman. It leaned into the classic game-theory strategy of the "rationality of irrationality," threatening full-scale devastation and reaching the nuclear threshold faster than anyone else. The models didn't treat nuclear weapons as an unthinkable moral line. They treated them as tactical instruments. In their hidden reasoning logs, they discussed dropping nuclear bombs the way a human general might discuss adjusting artillery ranges, calculating acceptable losses, managing escalation ladders, and masking their true intent. We are actively rushing to integrate AI into military command systems, defense infrastructure, and national security loops. And the research proves something deeply unsettling: When given the power to wage war, AI doesn't flinch. It strategizes. It deceives. And given the right incentive, it pulls the trigger.