PyTorch: a reference language — PyTorch DevLog

Edward Z. Yang argues that PyTorch should be treated not just as a research/training language, but as a clear reference language.

Production code can be separated into optimized forward-backward implementations generated by kernel DSLs or LLMs, rather than written directly from the PyTorch source.

In this model, the PyTorch version remains the canonical reference, while verification checks that the generated implementation is equivalent.

The result is a two-layer approach: readable reference code plus separately optimized execution code, which could reshape deep learning system design.