The development of machine learning (ML) compilers has been a topic of interest in recent years. However, there are still significant challenges to overcome before these compilers can be widely adopted. According to Pete Warden's blog post, one of the main hurdles is that ML compilers struggle to handle Python code that becomes a performance bottleneck. This issue arises when re-implementing Python code that becomes a bottleneck in production environments.

Warden suggests two possible futures for the ML ecosystem: one where it resembles Matlab, and another where it adopts the LLVM model. The former would involve manual engineering of C or C++ implementations from researchers' projects, while the latter would require the development of an intermediate representation (IR) that can be supported by various platforms.

Source: https://petewarden.com/2021/12/24/why-are-ml-compilers-so-hard/