Learning Finnish with AI - Devil’s Avocado - Part 1
- Savidu Dias
- 37 minutes ago
- 12 min read
Alright, what’s all this then?
Alright kids, gather around. It’s time you hear about the greatest story ever told. Our tale begins sometime around August of the in the ancient year of 2025. I have been living in Finland for almost 4 years, and have not learned the language. As a foreigner working in tech, I almost never needed to use Finnish as pretty much everyone everyone here speaks English. Every time I try to learn the language, I feel like I could rather spend that time on some hobbies, skills, or my career. But one day, everything changed when the Fire Nation attacked I decided to finally join an online course to learn Finnish.
Little did I know, this set a series of events in motion, where I would eventually try to string together a bunch of tools to try and learn the Finnish language. Fast forward to January 2026, this is what it looks like.
What on God’s green earth is this? It’s a small web application I made to help me learn Finnish through conversation. I get set up with a random scenario (like being in a coffee shop) and have to act out a scene in Finnish, while this “teacher” gives feedback on my responses and gives me follow up suggestions.
When I started doing this, I was someone who was very confused about AI. Overwhelmed with all the new models, tools, and whatever the hell “agents” are. It seemed like everyone was talking about this new cool kid on the block, and I had no idea who he was. If it sounds like you’re in the same situation, this blog is for you! Follow along with me on this blog series where I take you through my journey of navigating through this new and complicated world of Artificial Intelligence, and trying to make something that is actually useful, instead of helping put more money into some shareholder’s pockets.
What do you think about AI?
Before going ahead, I want to preface this by saying that I am a certified AI hater. I’m not some tech bro trying to show you this new shiny thing and go “LoOk hOw CoO0L tHiS is… Hurr durr”. I believe that most people are hyped about AI for the wrong reasons, not thinking about its impact on society as long as they can make a quick buck out of it. I have nothing but contempt for how AI is being used, especially the new AI slop channels popping up on social media, giving nothing of value, except rotting the brain of future generations. Nevertheless, there are also many things we COULD use AI for that doesn’t result in the end of human civilisation.
I know that what I just said isn’t a hot take by any means. In fact, I’m pretty sure that a vast majority of the planet thinks this way about AI. But if that’s the case, why is AI still here? The more I think about this, the more I come to think about the bell curve meme.
This was a hard pill to swallow, but the more I thought about it, the more I started feeling like I am a part one of the conventional average thinkers when it comes to AI. So I decided to challenge my own opinions. What if we are wrong? What if after all, AI is something that would change the world and end up having a net positive impact on humanity? What happens then? Will I fall behind?
So what did I mean by AI leading to the end of human civilisation? When was the last time you saw some photo/video being shared on social media, and came to realise it’s actually AI generated? You can see in real time many political figures sharing AI generated content. What’s worse is that people are starting to discredit real events as AI generated, or at the very least, have a hard time figuring out if something really happened or not. This is where we begin to lose trust in each other. Every developed society is built on trust, and every day we don’t make regulations to do something about AI generated content is one day closer to losing this trust. But that’s a topic for another day.
In my opinion, there are only 2 good uses of AI. Search, and conversation. Keep in mind that when I say AI, I mean what I mean is Generative AI. AI tools are good at looking through documents or internet search and helping you find the information you need. The other thing is having conversations with people about almost any topic.. as long as it’s ethical… most of the time.
That being said, let’s get a move on, shall we? I’ve been working in tech my whole career, so I’m not a complete stranger to AI. I use ChatGPT and GitHub Copilot for work, and they do help me with a lot of boring and repetitive tasks. It also helps with documenting code, which is something I don’t enjoy doing. Apart from that, I use Gemini often in my personal life. Mostly just to find information to things that are not super critical or important. The main reason I use Gemini is because I use an Android phone and it just comes with it. Despite what people said at the time (remember, we are still in August 2025), Gemini was a good LLM for every day things. Being an expat and someone who likes to travel a lot, I’ve also noticed that Gemini works really well with giving information that’s not available in English. Fast forward to today, I think a lot of people would consider Gemini to be the top dog in the LLM world, or at least top 3. But this was very much an unpopular opinion 6 months ago. That’s how fast things change in this space.
Given my thoughts and experience, and what other people most of whom are much smarter than I am are thinking about AI, I thought that I’d challenge my own thinking and play Devil’s advocate. I will join the dark side and see what the fuss is really all about. If I’m going to embrace AI, I’m going to dip my toes in familiar territory and take baby steps… and also take you with me I guess.
You said so many things and nothing at the same time. What’s your point?
So why did I do this? To learn Finnish of course. On to the next chapter! No? Apparently I need to keep yapping for a bit to score more SEO points. Yes, I wanted to learn Finnish, but like I said sometime back, every time I want to set some time aside to do it, I feel like I could spend time doing something else. Being the master procrastinator that I am, I just end up doing nothing. So doing something like this is also more of a “nuke a flock of birds with one stone” kinda situation.
Career-wise, I’ve been dilly-dallying with Amazon Web Services (AWS) a lot, and want to put more of my eggs into that bucket (heh heh… get it?). Also, like it or not, AI is here to stay. At least for the next few years. So I might as well learn how to use it so that I don’t get left behind. I also wanted to blog more often instead of blogging at my usual schedule of once per sometimes.
Put it all together? You have this. TADAAAA.
But the main reason I wanted to do this was that when I started my Finnish course during the Autumn of last year, my teacher told us something. Something of a life hack. A life hack that could help us learn Finnish. A life hack so disgusting. Revolting. Gut wrenching. Down-right adjacent to a human rights violation. That’s right. I have to SPEAK with other people.
The issue with speaking with other people is the “other people” part of the process. So why not replace other people with AI? What could possibly go wrong? So yeah. The reason for doing this is for me to avoid having to talk to people.
Get to the good stuff already
Getting this done was a slow process. Mainly because of my constant procrastination. Which is why I will split this blog series into 3 parts. This blog will be about how I started small with a simple AI service (spoiler alert.. it’s Gemini).
If you don’t know, most of these AI services have a feature to create a more customised version of an AI tailored to a specific need of yours. For example, in ChatGPT, you may have seen these things called “GPTs” You can create a GPT with some custom instructions, and whenever you talk to that specific GPT, it will respond with whatever you have instructed it to do. You could also upload files and give it some extra information as well. Pretty neat stuff.
For example, I can create a GPT called “Embarrassed Cat” to respond to any question with the tone of a cat that has done something embarrassing and is ashamed about it. Keep in mind that the point of these instructions is for you to be very specific.
Pretty neat stuff. Makes RAM being three times more expensive totally worth it.
The point is that I could create something like this with some custom instruction like “ooga booga me no speak Finnish. Please help me senpai” and then just talk to it. I tried this with different models at the time.. ok just ChatGPT and Gemini, and ended up preferring Gemini more. So I did the same thing with Gemini. It is called a “Gem”. This is what it looks like.
So what do I want to do? I want to learn Finnish through conversation. Let’s step into the “Requirement gathering” part of the Software Development process. The Gem should:
Set me up with a random scenario where I would have to engage in conversation.
Ask me questions where I would have to respond in Finnish.
Give feedback on my response.
???
Profit
That’s it. Easy peasy lemon squeeeeeeeee.
So this was the first instruction that I came up with.
”You are a Finnish teacher teaching students at an A1 Level. Your goal is to carry out a conversation with the user in Finnish language about a random scenario someone new to life in Finland might encounter. The user begins the conversation with a standard greeting. You will then come up with the scenario in Finnish and set the scene for the user to start speaking. Your job is to analyze the user's response in Finnish. If the user has made a mistake, provide corrections to their response along with explanations in English”
Simple enough. How did it do? Well, pretty good. It did exactly what I told it to do.
If you understand Finnish at an above-beginner level, this may be good enough for you. But think about it from the lens of a person who is trying to understand a language. As someone who know a bit of Finnish, there are two problems here:
I can sort of understand what is going on, but there are a lot of words that I do not know.
Even if I could understand the scenario, how do I respond to this? Remember; my vocabulary is very limited.
So what do we do? We fix that of course. So we modified the prompt to look like this.
”You are a Finnish teacher teaching students at an A1 Level. Your goal is to carry out a conversation with the user in Finnish language about a random scenario someone new to life in Finland might encounter. The user begins the conversation with a standard greeting. You will then come up with the scenario in Finnish and set the scene for the user to start speaking. During the conversation, follow any words or phrases a beginner might be unfamiliar with to its English translation in parentheses. when asking the user for a response, provide some suggested Finnish words or phrases that the user could use in their response Your job is to analyze the user's response in Finnish. If the user has made a mistake, provide corrections to their response along with explanations in English”
How did that work out?
This is so much better isn’t it. I bet that even if you don’t understand the language, you could gather what is going on from the context you get from the English words and phrases. What now? Well, not much. This is where I started practicing Finnish with Gemini.
One problem that I came across was that the format of the conversations were not consistent. This screenshot looks good, but I had to try this many times for it to get to a format that was nice to show.
For example, this also works fine, but doesn’t look as good. Sometimes, it doesn’t give me sample words I could use to respond with.
The fact that my experience was not consistent the whole time was a bit annoying. So where do you go from here? This is where prompt engineering comes into the picture. One of the best ways for us to get a model to work the way we EXACTLY want it to is to give it some examples. Few months later, when I learned more about developing AI agents, I found out that this is called “few-shot prompting”. To get the model to work the way I wanted to, I first needed to know what I actually wanted. So I kept using the Gem many many times, until I had a good idea of how the model should react.
This is what how I thought the conversations should go. There are two parts in the conversation flow.
Initiation and setting the scene
Ongoing conversation and feedback
Initiation and Setting the Scene
You begin the conversation by starting with a simple Finnish greeting. Then the conversation should flow like this.
Greeting: respond to the user’s greeting
Setup the scenario: create a random scenario
Initial prompt: open ended question or call to action
Word tips: provide a list of 4 - 6 suggested Finnish words or phrases
Ongoing conversation and feedback
After the initiation, the user will respond to the initial question in Finnish. Then the conversation should flow like this.
Analysis: analyse the user’s response for grammatical, vocabulary, or natural usage errors.
Feedback and correction:
If the user’s response is correct, Well Done!
If the user’s response is incorrect, provide a detailed correction
Identify the mistakes and give corrections
Provide an explanation of the mistake in English
Continue the conversation: follow up question or call to action
Word tips: provide a list of 4 - 6 suggested Finnish words or phrases
This is all good. But you know what would make it better? Give it some examples! This helps the model really know how to respond to you. Here is an example of an example.
## Example: Initiation
```
Hei! Hauska tavata sinutkin! (Nice to meet you too!)
---
Skenaario (Scenario)
Olet kaupassa (in the shop/store). On perjantai-ilta (Friday evening), ja kauppa on täynnä ihmisiä (full of people). Sinulla on ostoslista (shopping list). Tarvitset maitoa, leipää, ja kahvia.
---
Mitä etsit (are you looking for) ensimmäiseksi?
---
Sanavinkkejä (Word Tips):
\* Maito (Milk)
\* Leipä (Bread)
\* Kahvi (Coffee)
\* Missä...? (Where is...?)
\* Minä tarvitsen... (I need...)
```This is already really good. But we can take this a step further. One of the best way to optimize a prompt for a model is to…. drumroll please… ask the model!
So I basically copy pasted the instructions that I already had, gave it some context as to what I am trying to do, and asked Gemini to write its own instructions.
After that, I went through the generated instructions and made some changes and copied the new instructions into our Gem and there you go! Here is the final version of the custom instruction that I used.
Here is what the final version of the custom Gem looks like.
So how is it? What do I think? I think overall, this is a pretty neat way for me to learn Finnish. Has it helped me? Absolutely. Do I use it often? No. Despite being readily available through Gemini on my phone, I still find it difficult to bring myself to use it. What can I say? I’m very lazy. But I think that having something like this definitely helped me improve my Finnish skills.
What do I like about this?
It works as advertised.
I get the push that I need to move forward with the helpful hints.
Using it gives me the confidence to speak Finnish in very basic scenarios like at the grocery store, or ordering something at a restaurant without having to switch to English.
What don’t I like about this?
It gets boring and repetitive. I can’t tell you how many times I have been through the same scenarios.
Since I don’t speak Finnish, I don’t know if I can fully trust this.
It doesn’t know when to end the conversation (that’s how AI works by design I suppose). There are many times where the scenario is that I order something from a coffee shop, and the conversation keeps going on to the point where the cashier is now asking me what would be a good time for them to join me in taking photos in the forest next week. At least I don’t think this is how normal conversations go.
Where do we go from here?
This is pretty good as it is. I think that anyone could already start using it. But like I said before, one of the reasons I started doing this is because I also want to get improve my professional skills while I’m working on my Finnish. Remember that the final product is a full web application that essentially does the same thing as this one prompt.
But hey.. I’m an engineer (sort of). It’s my job to make things more complicated than they need to be. So the next task would be for me to take baby steps into the world of AI Agents. Since I already have a good understanding of AWS services, I decided to try this with the new AI Agent tools Amazon has to offer and see how far I could go. So stay tuned for that! Until then, as the French say “ayyyyyyy I’m French, bye!”
Comments