Julia has really lost the differentiable programming mindshare to JAX. I’ve spent weeks or months getting tricky gradients to work in Julia, only to have everything “just work” in JAX. The quality of the autograd is night and day, and goes down to the basic design decisions of the respective “languages” (in the sense that JAX jit compiles a subset of Python) and their intermediate representations.

Fundamentally, when you keep a tight, purely functional core representation of your language (e.g. jaxpr’s) and decompose your autograd into two steps (forward mode and a compiler-level transpose operation) you get a system that is substantially easier to guarantee correct gradients, is much more composable, and even makes it easier to define custom gradients.

Unfortunately, Julia didn’t actually have any proper PLT or compilers people involved in the outset. This is the original sin I see as someone with an interest in autograd. I’m sure someone more focused on type theory has a more cogent criticism of their design decisions in that domain and would identify a different “original sin”.

In the end, I think they’ve made a nice MatLab alternative but there’s a hard upper bound on what they can reach.