Small Language Model AI

[Skip Cave]

The ACM recently presented a talk by Guglielmo Iozzia, covering recent developments using Small Language Models (SLMs). A recording of the talk is on the ACM website (members only). Here's the link:

https://events.zoom.us/ejl/Alb-GsS5Wmn4n8VIeBlAign4P80Xq243CpM9KvKWJRjRKciV4eCf~Ax3CPw-xRSaNpQ9s-_fiKeh6lv-44FVAJ-M3EQX7FZh4Ha0LR_-Z-BPBsRb8QBvB3H7-FdmUSR1yypye4N4w-sHJUXZ0fA4h0qQ/tbNK6IJVSVGsws5erEnSmQ/video/JIaZrNUPRbiyTqP-ges9Mw?_dlaccessid=RnH_sLLRSM-APJQ6Y5ZqmA&directLinkOpen=1

Guglielmo Iozzia also wrote a book on the subject, entitled "Domain-Specific Small Language Models: Efficient AI for local deployment," available on Amazon.

The basic concept is this: If the scope of the information covered by the language model is limited to a specific narrow domain, the model's size can be significantly reduced. Thus, a Small Language Model can execute on local processors with limited memory, instead of having to run remotely on huge machines in vast remote data centers. 

Guglielmo shows that an SML containing information about a programming language such as Pascal, can be loaded on the user's local machine. This allows questions about programming in that language to be answered during the programming task, by simply asking the local SLM about a specific programming issue.

  

So the question is, can the information about an array language like APL or J be compressed into an SML that will run on a local machine, (which is also running the language interpreter)? Better yet, can the SML be coded in that array language?  That capability could significantly reduce the startup time for a new J or APL user, to come up to speed!. 

Let me know what you think....

Skip

Skip Cave
Cave Consulting LLC

[LdBeth]

>>>>> In <CAJ8Lg_edLy8bTcP4iJ2YLjt9...@mail.gmail.com>

>>>>> "'Skip Cave' via forum" <fo...@jsoftware.com> wrote:

> So the question is, can the information about an array language like
> APL or J be compressed into an SML that will run on a local machine,
> (which is also running the language interpreter)? Better yet, can
> the SML be coded in that array language? That capability could
> significantly reduce the startup time for a new J or APL user, to
> come up to speed!.

It would no doubt be feasible to compress all the literature
about APL or J into a language model, if that can be done
for Pascal.

And it is already possible to do deep learning in "array languages",
it is just these are not APL or J. In other words, most of the deep
learning frameworks are already heavily influenced by concepts from
array languages.

Limited by the model size, Small Language Models can not go very far
into logic reasoning and despite been trained on array specific
language materials that would not enable them to do very complex
tasks. LLMs by big vendors already know J or APL relatively well
compared to two years before, and the advantage of SLM is mostly about
cost and privacy which does not mean SLMs are more friendly to user.