In my younger years, I was fascinated by non-western European languages and I wanted to learn as many as I could. I also thought that doing so would give me a perfect justification to travelling to the other side of the world. So logically, I joined the INALCO, a prestigious French language university, as a student in the late 1990s, and started learning Vietnamese, the language of my ancestors. Later, way before K-pop and K-dramas were what they are today, I added Korean to my curriculum. The reason? I was totally into linguistics and decided it would be thrilling to study the language of a country that created its own alphabet in the mid-15th century.
On top of that, I've also been trying to learn Icelandic on and off for years. I don't know why, but I've been attracted to this language for a very long time. Is it because it's an "old language" that hasn't evolved significantly for a thousand years? Or because there's a sense of enigma with all the Viking mythology and the magic lore attached to it? Or just because it sounds nice - and also very complicated for the language enthusiast that I am?
Anyway, whatever the reason is, my passion for complicated languages (from a totally subjective French point of view) remained and so did my interest in Icelandic. A few years ago, I even tried to learn Old Norse with the help of a book I found on the subject, Viking Language by the scholar Jesse L. Byock. Wouldn't that be brilliant to be able to decipher the Icelandic Sagas? A great new technical skill to add to my resume! Totally useless, so fundamentally essential, as the French sometimes say: "inutile, donc indispensable".
Unfortunately, a lack of time combined with the difficulty to find a teacher of Icelandic here in Paris put a premature end to my many attempts to learn this language. For years, I hoped that Duolingo would add Icelandic to their long list of available languages, but they preferred to add fictional languages instead like Klingon (created by linguists for Star Trek) and High Valyrian (created for Games of Thrones). How surprised and delighted I was when I then recently discovered that Icelandic was among the many rare languages taught by another language application, Ling! I created an account, started my first lessons and I was immediately hooked.
However, these languages learning apps always have the same problem: they seem to throw words and sentences at you but they are always scarce on grammar explanations - when they offer any. I was confronted with that immediately after my second Icelandic lesson on the Ling app, when I couldn't understand the difference between these 2 sentences:
- "The boy is 12 years old" was translated by: "Drengurinn er tólf ára gamall"
- "The girl is 14 years old" was translated by: "Stúlkan er fjórtán ára gömul"
Why was is "gamall" in one case, and "gömul" in the other? Was it even the same word? Out of frustration and curiosity, I copied these 2 sentences into a conversational AI, asking it to act like an Icelandic teacher and give me an explanation and, to my surprise, it worked! The AI gave me a full grammatical explanation, completed with conjugation for the verb and declinations. It was so satisfying!
But I soon realised that using the AI I'm already paying for as an Icelandic grammar teacher would not be very practical, as it might use up my monthly allowance, or I'd end up being quickly limited if I were to use other online AI models for which I had no paid plan at all. Then I remembered that it was actually possible to download open source LLM (Large Language Models) and it gave me the idea of building my own grammar teacher - personal and local.
Indeed, using a local open source model would mean zero ongoing cost: no subscription to pay, no API fees and no dependency on a third party service. It would also offer offline availability, as a local model doesn't need an internet connection to work. On top of that, there were also a few ecological arguments. First, there would be a smaller network footprint: remote queries travel through layers of infrastructure, but local ones don't leave my machine. Secondly, no infrastructural redundancy: cloud services maintain multiple copies of models across servers but a local model exists once, on my own hardware. I could also add that there would be a positive impact on the distributed energy consumption, but this one doesn't really work: my Mac Mini runs 24/7, so the comparison here isn't straightforward.
After thinking this through, I decided to build my own "Icelandic Grammar Teacher" application. It would be a local web app, hosted on my computer at home, on an open-source LLM installed locally. And because sitting on my sofa is more comfortable than sitting in front of my desk, I would make it accessible from my phone when at home, using my personal Wi-Fi network.
There is something quietly absurd about this whole journey. I started out with the ambition of reading medieval sagas in their original tongue, armed with Jesse Byock's textbook and what I can only describe as an unreasonable level of optimism. Several years and a few abandoned attempts later, I find myself running a language model on my Mac Mini just to understand why an adjective changes its ending depending on whether it describes a boy or a girl. I am not sure the Vikings would have been impressed. But the question had an answer, and I intended to find it.
In a follow-up article, I will dive into the technical details of this adventure, so stay tuned!
