In 4 Hours, an Agent Livestreams Kernel Optimization to 87.3% of Expert-Level Performance

In 4 hours, an Agent on FORGEKERNEL LIVE pushed the FlashMLA Prefill kernel from scratch to 0.873x of FlashMLA, which is 87.3% of expert-level performance. The screenshot clock stops at T+03:55:28, just a few minutes short of 4 hours.

The task is written very specifically: FLASHMLA PREFILL - DEEPSEEK-V4.1 - B200, with parameters s_q=4096, h_q 128 / h_kv 1, d 512, topk 640, bf16>fp8. The work it does is not ordinary business code. The kernel runs directly on the GPU and changes throughput, latency, and hardware utilization.

Several numbers in the screenshot are more useful than slogans: the current version is VERSION 3, 593 μs; the fastest candidate kernel on the chart is 0.873x; the center shows -12.7%, with a note beside it saying it is still below the expert line; compute resources are 3.9 GPU-h, 3 versions, and 1 failure across 6 rounds; the top stats bar lists GPU-HOURS, ITERATIONS, KERNEL CODE BY HUMANS. The line chart runs from T+0 to T+4, with the vertical axis from 0.3x to 1.00x; the blue line is the speed change of the fastest version, marked with a -12.7% performance drop.

Saman Ahmed's tweet says it is crazy for AI to build and optimize its own kernel in real time, going from zero to 87.3% of expert-level performance in only 4 hours, and sums it up as moving from writing code to improving the system that runs the code. Yuxuan Li's original post is more restrained: “Within only 4 hours, our agent is getting really close to the frontier from scratch.” In other words, within 4 hours, the agent had already come very close to the frontier starting from zero.

In the replies, Synthia wrote only “That time,” accompanied by a “TIME TO COOK” video.

The real informational value of this post lies in the closed loop: generate candidate kernels, benchmark, select, and iterate. The 4 hours and 3.9 GPU-h are costs, not zero cost. It is still 12.7% away from the expert line, which is usually hard to close in kernel optimization; the closer you get to a hand-written library, the more fragmented the gains become. The screenshot does not hide the failed rounds, nor does it remove “KERNEL CODE BY HUMANS.” AI is clearly moving from writing code to modifying systems; but treating 0.873x as the endgame is premature.