I’m not convinced by the loops proposal. TensorFlow had this kind of lazy evaluation (except I guess TF was the worst of both worlds) and it makes debugging ver

I’m not convinced by the loops proposal. TensorFlow had this kind of lazy evaluation (except I guess TF was the worst of both worlds) and it makes debugging very difficult to the point that I believe it’s the main reason PyTorch won out. Such systems are great if they work perfectly but they never do. NumPy definitely has some rough edges, I sympathize with the frustrations for sure.