I've used Julia a bit and I fail to see that "light years ahead" point. Julia is sure a cool language to work with, with a very nice type system and a quite smart compiler, but well, the stuff I've used is still much of meta-programming (like Python does, except here it's JITed right). But maybe my epxerience is so limited I don't see the interesting bits...

Were you doing anything with:

1. custom units 2. custom GPU kernels 3. Custom array types 4. custom bayesian priors. 5. AD through custom types 6. Task based parallelism 7. symbolic gradients with modeling toolkit 8. Agent based modeling 9. physics informed neural networks 10. abstract tables types ...

or various combinations of the above?

I think the biggest thing that is "light years ahead" is that numpy is built right in and insanely extensible. Libraries can be very lightweight / maintainable because they don't need to roll their own numpy like tensorflow pytorch etc. do (which makes them depend on bigcorps to improve and maintain). In julia somehow everything is compatible and composable. On the other hand counting from 1 is is quite a big issue for me and leads to constant errors when I switch between python and julia. Also the startup time it takes is a huge bummer. They should have an interpreted mode for debugging (as long as the debugger is unusable for e.g. breakpoints)

I think the key is less that it has good multi-dimensional arrays built in, and more that Multiple dispatch makes Julia more composable than python. For an example of this, consider that BandedMatrices is a package that can be used in conjunction with any library that is expecting a Matrix like object, despite the fact that most of them weren't designed to do so.

I haven't tried it yet as it's still a release candidate, but startup time is said to be much improved in v1.6.0.

"Please don't think that Julia is only useful for 1. custom units 2. custom GPU kernels 3. Custom array types 4. custom bayesian priors. 5. AD through custom types 6. Task based parallelism 7. symbolic gradients with modeling toolkit 8. Agent based modeling 9. physics informed neural networks 10. abstract tables types ..."

Is Julia better for these use-cases? I'd interested and would love some links/examples.

> 2. custom GPU kernels

You can write them in Julia, whereas in Python you have to write them in C/C++ and then use the FFI to call them.

You can write them in python using numba

Unfortunately not. The Python ecosystem has much more already in both symbolics and numerics, and a much bigger community.