This kind of sounds like just giving up on making the compiler better.
The idea makes sense because researchers already think in PyTorch. The tooling should adapt to them, not the other way around.
This kind of sounds like just giving up on making the compiler better.