True. Due to its lack of verbosity, APL is not at all forgiving for any substitution or transposition in its symbols, which can cause great differences in program output, so it is quite improbable for LLMs to generate a working program, unless it is identical with one from its training set.

What I have said matches exactly what another poster said about his experience in using a LLM with APL: "Its corrected code almost never works".

The LLM recognizes the problem that must be solved by the code, but it fails to generate the right APL symbol string.

I doubt that here a coding agent that attempts to verify the generated code by compiling it can help, because the LLM will generate eventually some syntactically-correct symbol string, but which will implement a different function than desired.

Only a complete feedback loop, with a battery of varied tests for the executable program produced by the generated code, which can verify if it really implements the desired functionality, can be used to filter the results for a working program.

APL is greatly superior to almost all programming languages that are popular today, for writing expressions involving arrays (this includes expressions that do not involve arrays in other languages, but which could be made simpler by using arrays in APL).

However, the original APL has defects, due mainly to the fact that it was an incomplete programming language, e.g. when compared to the other contemporaneous IBM language, i.e. PL/I.

What one needs is a programming language with modern program structures, data types and data type definition facilities, but also with an expression syntax matching the power of APL expressions.

The fact that in 2026 most programmers continue tho write "for" loops for handling arrays, instead of using array expressions like it was possible in 1966 in APL, 60 years ago, seems an aberration of history. Even in the 1966 PL/I one could use array expressions, even if only expressions that were much simpler than those of APL.

Using symbols instead of keywords, like in APL, is not cryptic for anyone who uses such a language regularly. It is cryptic for those who have not used them. The English-based keywords are somewhat less cryptic only for English speakers, and even for them they can be misleading before they learn their correct meanings.

> “it is quite improbable for LLMs to generate a working program, unless it is identical with one from its training set”

This is a fascinating result. In some sense it’s like APL is actually the most human programming language, despite being one of the most difficult for ordinarily trained human programmers to pick up.

Whilst LLMs still perform weakly in APL, the situation is improving at pace, and giving it a “skill” to evaluate code makes a dramatic difference. I gave a conference talk about it recently (video): https://youtu.be/H_wdKeJ8gt4