amusing

amusingkernels is an agentic loop that turns SOL-ExecBench problems into fast GPU kernels. Starting from a problem definition, it writes Triton versions, validates their signatures, evaluates correctness and latency natively on a GB10 (with B200 checks via Modal), and iterates — every eval scored as time vs. GB10 roofline SOL, with durable lessons carried across runs. Two L1 problems solved end to end so far: GEGLU at 84% of SOL, an attention matmul at 67%.