Carragee Consulting

Hard for humans and LLMs – The Rise of SLMs

Some words can have lots of meanings. Playing with language and appreciation of its complexity is the basis for plays and song lyrics and grandmas puns.

Playing with double entendre at work could have serious consequences.

Is a Tender the same as a Contract? How does that differ from an SOW?

Is POS a Point of Sale or junker car?

LLMs don’t solve for what words mean in a specific business context very well. They work probabilistically and if your meaning is a minority use case in the training set of an LLM the results might not be reliable enough to use them.

To address this folks are re-training the models on specific datasets, particularly internal documentation, adding a fine tuning step with targeted inputs. Small Language Models used in conjunction with base LLMs are a solution to bring domain, industry and tailored expertise.

Kids learn language this way too — by building a base of 2000-3000 words in their first years and ~10000 by the time they are 8-10years old and “fluent”. English has hundreds of thousands of words though, almost all of which have at least two nuanced meanings.

As with employees, the usefulness of many of these AI tools is measured not with breadth, but depth and reliability in specialized domains.

“I don’t need my accountant to quote French poetry”.

Armand Ruiz post about SLMs