Tried out Co-dfns APL compiler for the first time today. For a simple softmax calculation, it generated 3.7M CUDA kernels... that is... pathologically bad 😳 @arcfide 🤔🤔🤔