Everything published about KDA so far has been on NVIDIA hardware with excellent documentation, a mature profiler and a large public corpus of prior kernels. That is a reasonable place to start and a bad place to stop, because it leaves open the obvious objection: perhaps the agent is retrieving, not reasoning.
So we pointed it at hardware where there is much less to retrieve. On ASM, HIP and ROCm — including AMD parts with substantially thinner documentation and far fewer published kernels — KDA still finds the optimization opportunity and still lands the speedup. Two of them are in FlyDSL, as pull requests #711 and #685.
This is the property that decides whether any of this is durable. New accelerators arrive with no corpus at all: NVIDIA's Rubin parts and Groq's hardware will both, at some point, be something an agent has read nothing about. A loop that works only where the answers already exist on the internet is a search engine with extra steps. A loop that works on the thin documentation is a method.