02 Dec 2025 — 4 min read

Photo by Andres Garcia / Unsplash

I started writing this Newslttr in a hotel in Hong Kong where I was on a mixed working/chilling trip. Between the amazing cuisine and bustling street markets of this Chinese Special Administrative Region I was able to find an actual use for some of the AI hype. Not the generative LLM bluster that seems to be underpinning the whole of the USA stock market at the moment, but the much simpler and more reliable technology of machine translation.

As a former British colony, English is widely spoken in Hong Kong as a professional and business language but I still encountered many people who could speak only Cantonese. In response to my blank face when I clearly don’t understand them, they would whip out their smartphones, speak into the microphone and almost instantly they can show me the English translation. It’s a lot less flashy than ChatGPT producing your next pitch deck but machine translation was clunky ten years ago and the stuff of science fiction a decade before that.

Spoken and written language are what divides humans from the rest of the animal kingdom. So much so that we imbue the likes of ChatGPT with apparent human-like intelligence just because we communicate with it through human language.

Human beings are much the same the world over but with over 7000 languages we are often divided by an inability to communicate. The idea of a universal translator has long been a popular trope in science fiction. Probably the most famous of these is the Babel fish from Douglas Adams’ The Hitch-hikers Guide to the Galaxy. Named after the biblical story of the Tower of Babel, where God cursed men not to understand each other’s speech, the Babel fish is described as absorbing speech and transmitting it again as brain waves.

"The practical upshot of all this is that if you stick a Babel fish in your ear you can instantly understand anything said to you in any form of language" - Douglas Adams, The Hitch-hikers Guide to the Galaxy

Google Translate doesn’t yet read your brainwaves (as much as Google might want to do that) and thankfully you don’t have to stick your smartphone in your ear to get it work. Instead it relies on a machine learning model trained on large corpuses of text. If that sounds familiar, it’s because it‘s much the same technique as the more en vogue LLMs use to build their models. In both cases, what you’re building is a model of language. LLMs essentially use theirs to predict the next most likely word in a sequence; translation models predict the most likely foreign translation given a sequence in the original language.

In order for the model to know how a phrase translates, it needs a large set of documents which convey the same meanings in different languages. One of the earliest and most reliable corpuses of such documents came courtesy of the European Union. The EU has twenty four official languages and all its documents are available in all the official languages. They’re produced by professional translators and are therefore of high quality. The translation models are essentially able to learn how a given phrase in one language should be translated to another by comparing the content in each of the different languages.

This is all well and good but there are a couple of potential snags. For one thing, EU documents are official publications which use quite formal language and jargon. People in the real world don’t speak or write in such a formal tone. Colloquialisms and slang pervade; languages have idiomatic phrases which don’t make any sense if translated literally. Secondly, European languages are predominant in Europe (obviously) and in countries which were previously part of European empires. There are far fewer examples of translated text datasets in other languages and the translation models are consequently much less reliable.

And this is where the worlds of LLMs and translation meet up. Transformers (no, not the robots that turn into cars) are a type of machine learning model that can do more than just represent how one phrase relates to another. By utilising an encoder and decoder architecture, they translate a phrase into an intermediate representation which captures its meaning. The decoder stage then translates that back into human language. The clever bit is that you can swap in whichever encoders and decoders you like to create arbitrary translators. So it doesn’t matter if you don’t have, say, an English-Hindi translation dataset as long as you have encoders and decoders for both languages. The universal intermediate representation of the content acts as a kind of digital Babel Fish enabling translation between language pairs.

Add in a speech-to-text engine and a text-to-speech function, both of which are probably also using a transformer architecture, and you’ve got yourself a real life universal translator. Model sizes and processor speed are now at the point where all this can be done locally on your smartphone without having to send anything to a remote data centre for processing.