If LLMs Are Calculators for Words Where Does That Leave the Language Services Industry?

Lukas Werner

A few weeks ago I read a blog post from Simon Willison about how LLMs are calculators for words. I’ve been thinking about how I should plan for this new “normal” world, and what sorts of things will exist and wont in it.

  1. One thing is for sure we will need to have lots of checking/validating of these models.
  2. For the second I’m not so certain about the language services industry still existing.

What even is the Language Services Industry?

Well basically, it has to do with translating things across languages. But they also do more such as helping with author editing, copy editing, proofreading, globalization and localization. In 2022 it was estimated to have had a $60.63 billion market value and is projected to have valued at $96.21 billion in 2023.

So now we shall apply the ‘LLMs are calculators for words lens’

and WOW, they are screwed.

First off lets start with Neural Machine Translation (NMT). It is essentially the exact same idea behind LLMs which the language services industry has been using this for YEARS and only used them to translate languages with higher speed and accuracy. Grammarly was the only one using similar tools to do the smarter stuff some of the “word calculator” stuff LLMs are doing.

And now for my “unique” insight into this…

Language Service Providers (LSPs) have been not taking smaller, lower budget clients, them away. So where have these clients been going to get their translations and help with language?

drum roll please

Tech companies that have been releasing tools to do it like Google Translate. Basically cutting out the LSPs in the process. But much like the calculator replaced tons of jobs and industries relating to math, LLMs likely will replace tons of jobs and industries relating to word processing.

However there is a tiny bit of hope for this massive industry.

Small resource languages like Swahili will continue to need people to do the word processing. So the people working on those small resource languages will end up basically being our modern mathematicians working on harder problems instead of on your getting your ledger updated for tax season.

Update(May 23, 2023): Meta AI

So today Meta AI released a new model MMS (Massively Multilingual Speech) which touts higher accuracy than whisper while having 10x more languages supported. This further points out a trend from Meta (and previously FaceBook) AI research. Wherein they have been building models that have large multilingual capabilities, e.g. NLLB (No Language Left Behind).

Meta has even started offering services based on these translations for free to business users under its translated advertisements offering, which is work the LSPs used to do ALL the time. Of which they have stagnated and left the customers to fall to a tech company to do for free as an after thought.