I've tried it for modeling tasks but I struggled with the sort of basic data manipulation that can be done within pandas/data.table. I can quite comfortably work with 100-million line CSVs using data.table on a standard laptop but Mathematica wasn't even able to ingest the file. I don’t disagree its technically very impressive but there's no point in having these amazing features if it stumbles with such basic tasks.
I've used Mathematica since about the year 2000, and I think Wolfram "missed the boat" of AI, big data, and machine learning.
They were in the unique position of having one of the best symbolic differentiation engines and one of the best numeric engines and a Lisp-like REPL that allows one to write terse, elegant code.
What they were always missing was efficient bulk data structures.
In recent versions they've added a handful of "special cases" where some types of data are stored as a plain data array like in C-derived languages, but this is hit-and-miss.
Similarly, they've dabbled with GPU acceleration and parallelism, but it's half-baked. It feels like a proof of concept, not something you'd ever actually use.
Julia and the like will slowly but surely eat their lunch.
the neural network stuff is very good and relies on GPU acceleration. it is the "computation graph" paradigm -- have to define the whole architecture up front -- but it's actually quite pleasant to use.
