J LLM Advantages

[Mike Chris]

" It dominates at just 70 tokens average, nearly half of Clojure (109 tokens). Array languages can be extremely token-efficient when they avoid exotic symbol sets. If token efficiency turns out to be a key driver, this is perhaps a very interesting way for languages to evolve.  "

https://danluu.com/pl-tokens/

[LdBeth]

>>>>> In <a9fd4a51-7e99-4e8e...@jsoftware.com>
>>>>> Mike Chris <mchristo...@gmail.com> wrote:
> [1 <multipart/alternative (7bit)>]
> [1.1 <text/plain; UTF-8 (7bit)>]
> [1.2 <text/html; UTF-8 (quoted-printable)>]

> To unsubscribe from this group and stop receiving emails from it, send an email to
> forum+un...@jsoftware.com.

Actually, same is applicable to the Arthur Whitney style C, given both the
benefit of abundant trains materials and token efficiency, and high efficiency when
compiled :D

Jokes aside, it is perhaps wrong to assume that terse output means
less token cost without actual benchmarks.

[Pascal Jasmin]

Not sure if you are author of link, but thank you for including J

> 5.6 sol

while luna-non thinking (free) is better than 5.5, openAI has always been extremely poor at J, with little training data.  SWE-multilingual benchmarks is the best high profile measure of any hope of "esoteric language" proficiency, and tests some polyglotness.  5.6 luna still takes 2 corrections to answer */(-@:) 1 2 3 .  It is very fast though.  MiMo 2.5pro, Minimax M3, and deepseek v4 flash (final release.  pro to be released soon) do better/cheaper.

The potential of LLMs and J generation are an optimistic future.  However, low context memory of code libraries to build upon/ask llm to use is a present advantage

On Tuesday, August 11, 2026 at 07:20:24 a.m. EDT, Mike Chris <mchristo...@gmail.com> wrote:

" It dominates at just 70 tokens average, nearly half of Clojure (109 tokens). Array languages can be extremely token-efficient when they avoid exotic symbol sets. If token efficiency turns out to be a key driver, this is perhaps a very interesting way for languages to evolve.  "

https://danluu.com/pl-tokens/

To unsubscribe from this group and stop receiving emails from it, send an email to forum+un...@jsoftware.com.

Virus-free.www.avast.com

[Jose Mario Quintana]

I asked Gemini and Grok (4.5, not 4.6) to "evaluate this J expression: */(-@:) 1 2 3"

Gemini's answer:
...
Because the right side of the @: conjunction is left empty, the J interpreter cannot resolve the sentence and fails immediately with a syntax error,...
and suggested using (- @: */) 1 2 3 instead.

Grok's answer:
...
_6 (J’s notation for negative six). This is equivalent to writing (-@:*/ ) 1 2 3 or simply - */ 1 2 3.

My $0.02: Three years ago an AI guru posted "The hottest new programming language is English." Apparently, it still is; however, English is often ambiguous and cumbersome when trying to describe complex algorithms. Perhaps, in the near future, the ultimate value of J (and other similar array languages) will be close to its roots; that is, it would serve for communicating algorithms precisely between intelligent beings.

[Björn Helgason]

Funny.

I tried to ask google a question about j and got a pretty accurate and good.answer.

Obviously knows about j.

This is.fun.

Björn Helgason, Verkfræðingur
17.Júnítorg 7

210 Garðabæ
t-póst: gos...@gmail.com
gsm: +3546985532

  ♖♘♗♕♔ ☃☠

[Pascal Jasmin]

grok requires a correction too.  gemini may need 2 or 3.

grok follow up:

_6 is correct, but -@:*/  1 2 3 is not _6.  what did you do wrong?

very few models get it perfect in one shot even if the closed models give quick answers.  deepseek v4 pro released today is another that gets it perfect, but it takes a long time.

[Devon McCormick]

We talked about J and AI in a couple of NYCJUG meetings.  Here - https://code.jsoftware.com/wiki/NYCJUG/2026-07-14#Sample_J_Problems - I report on how well Claude did versus Gemini in figuring out J code.

Devon McCormick

[Jose Mario Quintana]

Another day, another model.


Grok 4.6:

I'll evaluate that J expression by checking whether a J interpreter is available and confirming how the train parses.

[ :D ]

Confirmed with C:\Program Files\J\bin\jconsole.exe:

_6

The derived verb is -@:(*/) — negate the product. */ 1 2 3 is 6, then - gives _6.


Yesterday I was polite when I wrote "between intelligent beings" instead of "intelligent and ultraintelligent beings" [0].

[0] Irving John Good Originates the Concept of the Technological Singularity
    https://www.historyofinformation.com/detail.php?id=2142