Open-Source Agent Beats NVIDIA FlashMLA Benchmark in Black-Box Test

According to a discussion on Reddit's LocalLLaMA subreddit posted on November 2, 2026, OpenBMB has wrapped up ForgeKernel Live, an autonomous GPU kernel optimization run on NVIDIA B200 that lasted roughly 22 hours and was carried out by an autonomous AI agent. The defining detail of the experiment is that the agent operated against a black box. It was allowed to execute the official FlashMLA implementation and benchmark its output, but it could not read the source code, documentation, or profiling information for what it was optimizing against. The numbers cited in the discussion include 22 hours of continuous autonomous optimization, 32 kernel versions produced across 42 iteration rounds, and a first pass over the official baseline within under 10 hours. By the end of the run, the agent had lifted FlashMLA benchmark performance by 9.8%. The discussion frames the speedup as less interesting than the black-box setup itself, since the agent had to reason from measurement signals alone rather than from internals it could inspect. At this point the event exists only as a Reddit thread, and the figures quoted above come from that discussion; no other reports were available to corroborate them.