An AI Agent Just Beat NVIDIA's Own Engineers at Optimizing NVIDIA's Own GPU

An AI coding agent just beat NVIDIA's own engineers at optimizing NVIDIA's own GPU.

The paper is called AVO — Agentic Variation Operators. Researchers gave a coding agent access to CUDA documentation, Blackwell B200 architecture specs, and a scoring function. Then they let it run for 7 days.

The agent produced 40 committed kernel versions. Explored over 500 optimization directions internally. And the final result: 1,668 TFLOPS on multi-head attention — outperforming cuDNN by 3.5% and FlashAttention-4 by 10.5%.

cuDNN is NVIDIA's crown jewel. FlashAttention is the most important kernel innovation of the transformer era. Both represent months of hand-tuning by senior GPU kernel engineers — the kind of people who think in register allocation and warp synchronization.

The agent beat them in a week.

What makes this different from the usual "AI writes code" story: the optimizations weren't superficial. The agent discovered branchless accumulator rescaling that eliminated warp synchronization overhead. It restructured the dual-stage pipeline so correction and GEMM operations overlap. It rebalanced registers across warp groups — 184/88/56 instead of FlashAttention's 192/80/48 — reducing spilling on the critical path.

These are optimizations that require reasoning simultaneously about memory ordering, pipeline scheduling, and hardware architecture. The kind of work that takes senior engineers weeks to even conceptualize.

And then the agent adapted its MHA kernel to grouped-query attention in 30 minutes. Not 30 days. 30 minutes.

The paper's framing is the most interesting part. They argue that confining AI to a "generate code" step inside a human-designed pipeline is fundamentally limiting. The real unlock: give the agent full autonomy over the entire process — when to consult documentation, what to test, when to change strategy, when to backtrack.

Sound familiar? That's the same shift Karpathy described last week. Stop typing code. Start directing agents. The agents don't just execute faster. They explore spaces humans can't reach.

Paper: AVO: Agentic Variation Operators for Autonomous Evolutionary Search