> One downside of KANs is that they take longer per parameter to train—in part because they can’t take advantage of GPUs. This seems like a big gap. Anyone kno

> One downside of KANs is that they take longer per parameter to train—in part because they can’t take advantage of GPUs. This seems like a big gap. Anyone know if this is a fundamental architecture mismatch, or just no one has written the required CUDA kernels yet?