I've always wondered about code like this and what inspires people to write it. How is this any different from just directly writing machine code?
Or is this the exact opposite, operating at the highest level of abstraction?
Or is this the exact opposite, operating at the highest level of abstraction?
It's not as abstract as possible or anything. Most directly, there are abstractions like algebraic datatypes or macros that APL doesn't have at all. And there's a sense in which APL is very concrete: it tends to use first-order functions instead of higher-order ones like Replicate[0] instead of filter. For me the fact that an array oriented program tends to handle just a few arrays at a time makes it easier to understand what's going on. I think there are also ways in which APL programming is low-level. Going back to Jelly, it's actually a stack machine, same as Java object code, or Forth. Modern APL has tacit programming[1], which has a similar mix of mathematical purity and machine-like shuffling. I think the way APL handles array axes is similar. They are numbered, so to work with them you have to keep track of the ordering. Something like einsum that names them would be higher-level.
Next, array languages abstract away the execution strategy of loops. You as a programmer do not pick whether you use an accumulator or divide-and-conquer in your loops. Some good parallel libraries like Rust's rayon arguably have this feature in the imperative model, but it's not exactly a language feature of Rust.
W.r.t. machine languages, array programs do not have typing information which I think is the case for all modern intermediate representations.
Machine code can be practical. To me, code golf is just pure programming, for no practical purpose. It's not for everyone, or even most people. It's more like art than practical programming. Except it can be competitive? Not sure really.
It can also be fun just to build something, without the risk that you might accidentally get a significant number of users.