It always felt strange to me that the main implementation of something as niche and esolang-adjacent as APL is neither OSS nor casually usable commercially, but instead comes under an enterprise license.
Anyway, I had a fun time a while ago translating APL programs to NumPy. At some point you get what APL is all about, and you can move on with life without too many regrets. Turns out most of the time it's more like a puzzle to get an (often inefficient) terse implementation by torturing some linear algebra operators.
If you're after a language that's OSS, has terse notation, and rewires your brain by helping you think more clearly instead of puzzle-solving, TLA+ is the answer.
Edit: if you're curious to see at a glance what APL is all about:
APL code:
(2=+⌿0=∘.|⍨⍳N)/⍳N <- this computes primes up to N and is presented as the 'Hello world' of APL.
Thanks for the response. I'd interpret it as a valid technical caveat, but it feels somewhat orthogonal to what I was pointing out.
You focus on the 'often inefficient' parenthetical, yet, to me, your response highlights the puzzle nature of the thinking APL encourages. If anything, it shifts the question from 'how do I express this tersely' to a still narrower 'how do I express this tersely in a way the interpreter can also optimize'.
I think every programming language to a degree has some kind of puzzle aspect
I'm not sure APL has more or less of it compared to other languages
for example in Python, even though the language has a concept of "There should be one-- and preferably only one --obvious way" (PEP 20) it is quite multi paradigm, which I think is a strength of Python
oop, functional, imperative, …
and you get tons of libraries to choose from
e.g. numpy, pandas, polars, pytorch, keras, jax, … etc
but you still also have to figure out the algorithm and data structures you want to use (like in any language)
and you also kinda want to know (if you care about performance) how pytorch differs from numpy and how that differs from using a list with boxed values
Not saying this is not the case with APL
it definitely helps if you are familiar with the APL implementation you're using if you care about performance
I just don't think it's a disadvantage of APL over other languages
Agreed. I think I shouldn't put hard boxes around languages like 'puzzle language' vs 'abstraction/clear thinking language'.
What I was trying to point at was more specific: the way I experience APL thinking tends toward 'expression search' and 'notation compression', which feels, to me at least, somewhat at odds with clarity about the underlying problem. More often than not, I seemed to produce an APL-shaped model of the problem rather than a problem-shaped model expressed in a language.
When I first learned about APL, I was looking for new ways to think about computation. What I found was a language that rewarded deciphering APL programs and generating clever new ones. That is interesting and beautiful, but it was not quite the kind of brain-rewiring I was looking for. My original comment was targeting people in a position similar to mine and trying to set expectations about what they would learn best from APL. APL may change how you think about array expressions and how far they can go, but TLA+ is much closer to what I'd recommend if what you want is clearer thinking about programs, systems and state.
> At some point you get what APL is all about, and you can move on with life without too many regrets.
Unfortunately, this seems to be a common experience. A lot of smart people only engage with APL via toy puzzles, like you did, and bounce off because that gives no insight about how to use the language in real life. IME, to really start getting APL you need to write and rewrite a full application 20 times.
It helps to read code from the masters, too [0, 1, 2, 3, 4]. These all approach architecture in different ways: pedagogical FP style, OOP heavy, data-oriented design, event-driven state-machine, or a mix of the above.
> As you can see, the famous prime generator is not even the Eratostenes' sieve, but a simple N^2 divisor counting computation.
Well, that's because you wrote a divisor function, not a seive. Arguably, the ease of typing an outer product (i.e ∘.|⍨⍳N) can tempt us into writing quadratic algorithms unnecessarily, but this is just an experience issue, IMO.
If we want a seive, we can just write one directly:
p⊣{ω~n×1+⍳⌊N÷p⍪←n←ω↑⍨1⌊≢ω}⍣≡1↓1+⍳N⊣p←⍬
The algorithm is O(N log log N) as expected of a naive Eratosthenes implementation. You'll need ⎕IO←0 if you want to try it out.
There's also a faster seive by Roger Hui [0] in the dfns workspace as well as a family of prime number functions [1] for things more than just prime generation.
That's not O(N log log N), it's more like N^2. Prime sieves are hard to implement well with immutable arrays for obvious reasons; there are some cool methods but they're definitely harder. I'm ashamed to be part of a community that won't cop to this.
The algorithm iterates over numbers ⍺ from 2 to N, removing the multiples that are greater than ⍺ and no greater than N from p. If the removal with ~ has to inspect all of p, then all the primes are there so we have asymptotically at least N/log N entries by the prime number theorem and we get N^2/log N time (when ⍺ is over N/2, no multiples are in range so this can be skipped, but that just cuts a constant factor 2 from the time). Conceivably p could be marked sorted, so the entries to remove could be found with a binary search. This is a bit harder to analyze, but I think each prime under √N will cause the list to change, and incur N/log N data movement. So you get at least (N/log N)^(3/2) cost, still quite a bit worse than linear.
Edit: changed the algorithm while I was writing... the new one is better, it keeps one list p of primes and one list ω of not-yet-marked-out numbers. However, for the first √N primes generated, ω changes, and it contains at least N/log N - √N/log√N primes, where the √ part drops out because it's not asymptotically relevant. You end up with at least (N/log N)^(3/2) cost again.
>A lot of smart people only engage with APL via toy puzzles
I think part of this is because that is how most (possibly all) sources teach APL and array languages, solving puzzles and manipulating arrays. If you learn to write programs in an Algol derived language, you can write programs in most common languages without having to learn how to write programs, you just need to learn the language. Modern array languages sort of allow us to use them like the Algol derived languages, but this does not seem to work out so well and often does not work to the strengths of array languages.
I basically feel the same way. In a way it is very liberating. All of those esoteric languages that were on my ever-growing todo list are now things I can let go of. Ultimately we have to ask ourselves how we want to spend our time, and now it is much harder to justify spending countless hours studying one programming language after another. We still can, of course, but we are now more "free" to do other things instead.
It's sort of sad, but really I think it is a weight off my shoulders.
I had a little excursion into Dyalog APL recently and wound up writing an emacs mode to evaluate Dyalog APL [1]. It was a pretty nice experience using Claude to extract the small subset of features I wanted from gnu-apl-mode [2] to work with Dyalog APL.
I’d really like to properly get into APL though. My plan is to solve a bunch of problems on Kattis [3].
I'm really enjoying this way of learning a new language in the age of LLMs - starting with easy problems on an online code judge website and work with an LLM to come up with/explain simple solutions. It gives me dopamine hits, lots of reps, allows me to start coding right away, and is a nice way to slowly ramp up difficulty and get practice with different features of the language.
Nice to see this getting the Jupyter Notebook treatment. The original book was already one of the better introductions to APL. Interactive examples make a huge difference for a language where half the learning curve is just building muscle memory with the symbols
Given you could even use it commercially (it requires an enterprise license, but I suppose Matlab does too), moderately useful conceptually, weakly useful mechanically. APL is very limited in what offers you. I did a ML course in Matlab a while ago and I remember I could scalar loops and procedural scripts, had nice tables and object-ike structures. You'd give that all up in APL so it wouldn't help you there, but you'd see how far you can get only with creative 'array language semantics'.
It does, over time. It changes the way you think about computational problem solving. It's like the difference between designing objects in 2D on a drafting table and moving to 3D CAD. It changes your brain visualizes, explores and solves problems.
That said, learning APL isn't about learning the symbols any more than learning mathematics is not about learning the meaning of the various symbols it uses. To continue with that parallel, it also isn't about memorizing formulas. It is about using the tools to solve problems and, over time, changing the way you solve problems...now in 3D.
I learned APL in the early 80's and used it professionally for about ten years. The way I think of solving problems is fundamentally different in many ways because of this experience.
In the modern world there is no place for the commercial compiler. They should have made it free (and open source) and only IDE (maybe) paid one. Even better - push into GCC or LLVM.
I agree with you in principal, but I don't think that that's realistic at all, since that would completely destroy Dyalog's entire business model. I personally have little interest in learning a non-FOSS programming language, but Dyalog's paying customers are clearly okay with it, so I see little reason for them to change.
But why? Commercial APL implementations have existed since decades ago. They always had customers. They were always at least partly owned by employees. Dyalog was not created in a void, but continuing this tradition.
Anyway, I had a fun time a while ago translating APL programs to NumPy. At some point you get what APL is all about, and you can move on with life without too many regrets. Turns out most of the time it's more like a puzzle to get an (often inefficient) terse implementation by torturing some linear algebra operators.
If you're after a language that's OSS, has terse notation, and rewires your brain by helping you think more clearly instead of puzzle-solving, TLA+ is the answer.
Edit: if you're curious to see at a glance what APL is all about:
APL code:
(2=+⌿0=∘.|⍨⍳N)/⍳N <- this computes primes up to N and is presented as the 'Hello world' of APL.
Equivalent NUMPY code:
```
R = np.arange(1, N + 1) # ⍳N
divides = (R[None, :] % R[:, None]) == 0 # 0=∘.|⍨⍳N
divisor_counts = divides.sum(axis=0) # +⌿
result = R[divisor_counts == 2] # (2=...)/⍳N
```
As you can see, the famous prime generator is not even the Eratostenes' sieve, but a simple N^2 divisor counting computation.
reply