Model trained without natural language writes kernels passing 28% of tests, up to 66x faster than Triton

cephaloform · x · 2026-09-27

BenKoska reports a surprising result: a model that was never trained on natural language or natural-language reasoning could still produce GPU kernels that passed all tests 28% of the time, achieving speedups of up to 66x compared to Triton kernels. "Turns out, it actually kind of works?"

Original post →

More from coding & agent

coding & agent channel →