An area that's still pretty different between J and APL (IMO) is that ⋄ makes APL much easier for procedural programming while J is easier for functional code. As a result, I (a journeyman in J and beginner in APL) tend to procedural much more in APL. I also get this impression from APLcart where ⋄ and ← are not at all uncommon. Of course, you can do that with J also using [x=. A second area of difference is that J is arguably more math-oriented, which is ironic considering the origins of APL, but I mean look at the primitives! J's 'higher-math' primitives are really useful for preparing datasets.

Also, I'm mainly doing preparatory data wrangling with J (i.e. quick&dirty work) and not having to activate the APL keys everytime is a nice feature. The ability to write a simple #! script and dispensing with the 'APL machine' of workspaces also counts (a feature that just appeared in APL). As a data wrangler though, the true killer feature of J is Jd, the integrated columnar store. Of course you can mmap through it all in APL, but having a completely integrated solution does the trick for me! Actually, I find the concept of 'fetch your data SQLish and model it with J' such a good idea that I'm playing with the idea of extending the concept to the Racket data science world (which is mostly non-existant) :-)

So if you're doing serious programming as done in Fortune 500 companies I'm pretty sure APL wins due to consistency and tooling integration, but for one-off work J is a killer!