I found that a big issue with it in practice is that the numbers are just not great, and the repo is very vibe-coded. When the numbers are great, it's sometimes cheating with the setup. For example some GEMMs only compute each 8th value and reach close to peak TFLOPs because of this, but it's a vibe-coded reward hack and the verifier does not check all numbers of the target.

Also, compilers have a lot to prove in 2026. Why not just hill-climb a Triton or Cuda kernel if you need perf?