So far, AI has not been great at writing APL code. But Grok appears pretty good at reading and evaluating APL code. We asked Grok to evaluate this:

40 f 'Once upon a time in the west.'

given:

f←{ 
     ⎕IO ⎕ML←0 1
     segs←{¯1+⍵{(⍵,⍴⍺)-¯1,⍵}⍸⍵}
     split←{((⍵|⍺)>⍳⍵)+⌊⍺÷1⌈⍵}
     lf sp←(⎕UCS 13 32)=⊂⍵
     sizes←segs lf
     ⍺←⌈/sizes
     blanks←segs~(lf∨sp)/sp
     required←blanks+⍺-sizes
     breps←required split¨blanks
     last←1⌈¯1+⍴sizes
     brep←∊(last↑breps),×last↓breps
     ((~sp)+sp\∊brep)\⍵
 }

Grok proceeded to analyze and simulate execution of the function line by line, in gory detail and with voluminous output, explaining every step, all of which appeared to be correct. It got the right answer, and it knew and reported on the value of every variable along the way. It recognized the purpose of the function.

Here is a very small extract of Grok's output:

apl
sizes ← segs lf
lf is all zeros (no line feeds).
segs computes segment lengths:
⍸lf is empty (no 1s).
⍴⍺ = 29 (length of ⍵).
(⍵,⍴⍺)-¯1,⍵ = (empty,29)-¯1,empty = 29-¯1 = 30.
¯1+30 = 29.
So, sizes = 29 (one segment of length 29).

Grok claims it only used its knowledge of APL primitives and syntax to evaluate the function, and that it did not look on the internet to find if this function existed somewhere and get information about it. I think I believe it.

So then we asked Grok to evaluate the expression:

      'MM-DD-YYYY' Text2Date '12-31-1999'  '02-29-2001' '01-01-2000'

given this code base. It got almost the right answer, getting hung up in the Scale function thinking ⍺-6-⍺ was (⍺-6)-⍺ rather than ⍺-(6-⍺), which is a rather odd error given everything else it got right. But that was its only error. Other than that one glitch, it worked though numerous functions and explicated everything in profuse detail.