J/AI meeting 3 - presenters?

[Henry Rich]

I have been out for a month of vacation and a month of detailed coding -
that's a full technology generation at AI-speed.  Should we have another
Zoom meeting on the state of J-with-AI?

If anyone has anything new to share, email me.  Please give an
indication of how long you expect to talk.  If there is enough material,
I will send out a poll for meeting times.

From our previous sessions I have accumulated a list of a few things
that Jsoftware could do to help AI:

1. Help for the AI's tokenizer, to make sure that J tokens are properly
recognized

2. More J code for the AI to train on.  I myself have ~100KLOC of J that
I could make public.  Where is the right place to put it?

3. English instructions on how to write J

I'd like to add more to this list, and to start implementing it.

hhr

[Michal Wallace]

Im in.  I'd also like to share some brief thoughts about a role for J in a world where AI writes most of our code.

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

[Pascal Jasmin]

I've mentioned J's terseness as an advantage for codebase using little context room in agent/harnesses.

A new advantage is deterministic code to assist LLMs/harnesses.  A popular new category of models is fast cheap classification (Jev is first breakthrough), which can quickly solve 1-50 classification questions about a prompt/text.  A deterministic program can analyze the collection of classification results to produce a final decision.

A big J advantage is that its cold start is much faster than python or node/JS, as well as fastest program load.  That a J function to grade one question "automagically" works with a list of questionss is another advantage.

As always, I'll be availabe at the end of the session if there is time available.

[Thomas McGuire]

We had discussed previously about training a small model to do J coding. I have models up on hugging face that have been trained on the V3Stack data of J code from public repositories and the jwikisource (the text snapshot of the J website). I also have the datasets there as well. I can give a quick update and I'll share a link to the q8_0 quantized model that I will fund for a while so J people can try it. I'll give a short report if you want to include that in this next meeting. 

Tom McGuire

[Devon McCormick]

I could spend 5-10 minutes talking about trying to get Claude to write more idiomatic J.  TLDR; it's an uphill slog.

Devon McCormick