thoughts on slm’s for businesses

slms, has anyone even heard of them? you probably heard of llms, slms is exactly what you are thinking. small language models.

i just came across an X post a week before talking about monetizing this particular aspect of llms and it got me thinking.

the problem with llms is that they are tailored for a wide variety of work. they can write emails, solve math, explain physics, write code, and probably tell you how to build minecraft from scratch.

but most businesses probably don’t need all of that.

imagine a trading company gets 1000s of invoices every month, and they replace the manual labor with an automated task force. they use an llm to sort out this mess.

but does it really need a model which has also been trained on solving physics? math? or how to build minecraft from scratch?

probably not.

this got me thinking, we could use slm’s and specialize them particularly on that specific company data and workflow, and it will perform that specific task.

for example:

invoice comes in → understand it → extract important information → report it.

a similar idea can be seen in project nanda by MIT professor ramesh raskar. i like his work pretty well. the general idea of having different agents doing different tasks is interesting. more on this on future posts (maybe)

but then i started thinking, what is even the difference between a specialized slm, a sub agent using an llm, and just giving an llm tools?

no clue.

anyway, you might wonder, but hey utkarsh, isn’t an llm gonna be better trained on how to report, understand invoices, and all that?

well you are right.

and i don’t know the answer to that.

but i think one of the biggest reasons could be cost.

if a smaller model can perform the same narrow task, then maybe it is cheaper to run, needs less compute, and could even be deployed locally for companies that care about privacy.

we could just license that particular slm or system to the company and make a profit out of it.

but then again, maybe the company doesn’t even care if it is an slm.

they probably just care that something which took employees hours now takes a few seconds.

so maybe the business isn’t really selling slm’s.

maybe it is finding boring expensive workflows and using the smallest model that can automate them.

no clue if this is actually a good business idea.

but i want to look into it more.

peace.