- Sources: arXiv 2608.13122, submitted 2026-08-13, HN discussion
- Summary: The paper reports a 5.1x speedup from an agent workflow that ports a 250,000-line Fortran weather code to GPU. The reported speedup rests on element-wise comparison against dumped reference states rather than on the agent's own claim. That harness flagged five of 162 kernels as numerically wrong, with floating-point and intrinsic-function divergences.
- Why it matters: The validation harness, not the agent, is what caught the wrong kernels, which sets the bar for anyone applying agents to numerical porting work.
send feedback on this story