amk

amusingkernels is an LLM-driven agent that writes, debugs, and optimizes GPU kernels for NVIDIA's SOL-ExecBench benchmark, producing leaderboard-ready submissions. Muse drives a tool-using loop — read the problem, write a kernel, validate its signature, evaluate on a local GPU or a Modal B200, read back the trace — while Gemini acts as second-opinion critic on failures and fresh-eyes proposer when the search stalls, and KernelWiki contributes retrieved optimization techniques matched to the run's measured symptoms. At its heart is an exercise in software archaeology: a GPU-less analytical model of the NVIDIA B200, reconstructed without access to the vendor's proprietary timing models — a timed Petri net of the TMA → tcgen05.mma → epilogue pipeline assembled from datasheet peaks and CUTLASS pipeline semantics, bit-exact functional models of narrow-precision tensor arithmetic, and hard architectural gates for shared memory, TMEM, and instruction shapes, all pieced together from public shards. Every Modal B200 evaluation is logged as a predicted-vs-measured pair, so each real trace carbon-dates the reconstruction: agreement is reported as latency ratios, rank correlation, and bottleneck confusion, and the residuals drive calibration of the model's uncalibrated constants. Early results validate the approach end to end: seven frozen L1 solutions, a 10-turn dual-model FlashInfer-Bench GEMM campaign, and a first calibrated comparison (123µs predicted vs 50.5µs measured) closing the loop between excavation and evidence.