The idea makes sense because researchers already think in PyTorch. The tooling should adapt to them, not the other way around.