The Translation Program (1)
The process spanning from model selection to model design and construction was the most critical part of this translation program.
It was, after all, the creation of the artificial intelligence that would serve as the brain of the system.
Consequently, it demanded a massive amount of time.
At least selecting the model didn't take long.
From this point forward, the work would fully utilize NeuralScript, the transcendent programming language I had unlocked.
Therefore, I only needed to choose one of NeuralScript's models, and the choice was practically set in stone.
Even in the world where this language originated, one model stood head and shoulders above the rest.
Its name was Synapsion.
Normally, when choosing a model, one had to consider the characteristics of the data, the complexity of the task, and the required performance. This model met every single one of those criteria.
It was also the most perfect in terms of versatility, scalability, speed, and weight, making it the final choice.
Designing the model afterward took a very long time, though.
It was the process of meticulously designing the artificial intelligence's neural network.
If something went wrong here, even a flawless construction would be meaningless. Since a tiny mistake could snowball into a massive problem, I proceeded with extreme caution, step by step.
First, defining the structure of the model.
A model consists of multiple layers: an input layer, hidden layers, and an output layer.
First, I had to choose the type and number of each layer.
Next, determining the components of each layer.
In the case of an artificial neural network, this meant setting the number of neurons, weights, and biases.
This was to adjust the performance to match the characteristics of the data and its intended application.
The subsequent steps served similar purposes.
Selecting and refining the activation functions.
Specifying the loss functions.
And choosing the optimization algorithms.
With that, the model design was temporarily complete.
But building the model after that inevitably took even longer.
It was the actual process of implementing everything piece by piece.
While I had designed the model relatively quickly thanks to having all the information in my head, building it was a different story.
This time, physical limitations slowed me down. After all, I only had two hands. Still, I built the model step by step, slowly but surely.
Exactly as designed.
I typed out the code line by line, building the program.
It was like laying bricks to construct a building.
In construction, tools like shovels and hammers are used. The same goes for programming, where these tools are called frameworks.
But the problem was...
I didn't have a framework for this language.
If it were a standard Earth language, I could have simply downloaded a framework from the internet.
But this was not a language of this world.
Therefore, I had to code and implement them myself.
Naturally, it took extra time, but making the tools first was faster in the long run than trying to build a skyscraper with my bare hands.
Fortunately, it wasn't difficult.
For starters, the source code was perfectly etched in my mind.
However, some frameworks were written in languages other than NeuralScript, so I had to acquire those languages through my skills.
Perhaps because I already possessed NeuralScript, or because these languages were lower-level, the amount of Karma required wasn't that high.
I managed to acquire them using the 20,000 Karma I had left over, along with the Karma I steadily accumulated.
Do you really wish to pay 30,000 Karma to access the knowledge of "CerebNet"?
Yes. And apply the Korean language patch this time too.
Thanks to that, my coding speed increased dramatically.
If I had used English, the code would have looked like this:
language_model = Model(input_dimension=input_dimension, hidden_dimension=hidden_dimension, output_dimension=output_dimension)
But with the Korean patch, I could write it like this:
EoneoModel = Model(Ibryeok_Chawon=Ibryeok_Chawon, Eunnik_Chawon=Eunnik_Chawon, Chulryeok_Chawon=Chulryeok_Chawon)
The latter was not only slightly simpler, but it was also a format much more familiar to me. After all, Korean was my mother tongue.
Thanks to that, my overall coding speed became incredibly fast.
From the moment I started building the frameworks...
...to the actual construction of the model, everything went smoothly.
And thus, the artificial intelligence was complete.
"Genie."
"Yes, Master."
Having already fed it language data, it was capable of basic communication.
I ran several translation tests.
The results showed that translating straightforward sentences worked perfectly.
[Test 1]
Input (Korean): "I exercise every morning."
Output (English): "I exercise every morning."
[Test 2]
Input (English): "I like to listen to music while I work."
Output (Korean): "I like to listen to music while I work."
"It works well."
But it was still lacking.
The limitations became apparent in the subsequent tests.
[Test 3]
Input (Korean): "I attended my friend's birthday party."
Output (English): "I attended my friend's birthday party."
[Test 4]
Input (English): "She has a heart of gold."
Output (Korean): "She has a heart made of gold." [Literal translation]
In the former case, Genie translated the Korean verb for "attended" to "attended," but the word is actually quite formal in English. If a native speaker heard that translation, they would think the friend's birthday party was some grand, official event. To sound natural, "went to" would be much more appropriate.
In the latter case, Genie translated "heart of gold" literally as "a heart made of gold," but it is actually an idiom meaning a kind and generous nature. Therefore, it should have been translated into Korean as "She has a kind heart."
There are plenty of other errors too.
This was a common issue shared by existing translation programs.
Failing to grasp context, failing to recognize homophones, and failing to understand cultural nuances.
While current translation programs were trying to overcome these hurdles, they had yet to succeed.
But Genie can overcome this.
There are two key factors that determine the accuracy of machine translation.
The performance of the translation program itself, and the quality of the base data.
And since Genie's performance is already top-notch...
...I just need to secure high-quality data.
To that end, I issued a command.
"Gather as much data as possible from the internet."
"Understood, Master."
During the gap between the acceptance announcement and the start of college, I kept myself incredibly busy.
During the day, I provided quotes for water purification facilities and oversaw their construction; during my free time and late into the night, I coded.
But even in the midst of all this, I had to take care of what was important.
First, I took my parents to the hospital run by KAIST for a comprehensive health checkup.
It was a top-tier VIP package costing over five million won.
Naturally, my parents waved their hands in refusal.
"Oh dear... I don't think we need to do something this expensive," my mother said.
"Exactly. It's not like we're old folks on our deathbeds," my father chimed in.
"Let's just get the basic package."
"Yeah, that'll be plenty. If they run too many unnecessary tests, all that radiation or whatever might actually make our health worse."
Well, that wasn't entirely wrong, but...
"Don't worry. The doctors check the results and consult with you throughout the process, so they won't run those tests if they aren't necessary."
"Still..."
"Just do it. I have plenty of money now. If I run out, I can always make more."
With that, I practically forced them inside.
I nudged them forward by their backs as they continued to express their reluctance.
"Come on, in you go."
"Sigh..."
"This is so overwhelming..."
Yet, when I met them again a few hours later, they looked immensely satisfied.
"Hohoho, money really is a wonderful thing," my mother beamed. "The nurse stayed right by my side and guided me so kindly."
"I know, right? It felt like the doctor was with us the entire time. The explanations were so clear and friendly, too."
"Thanks, son."
"Me too. Since it's already paid for, I might as well enjoy it to the fullest. Hehe."
I wasn't sure how one could "enjoy" a medical checkup...
But I supposed it just meant the process had been that comfortable.
And the results of the checkup...
...showed no major abnormalities.
"See? I told you there wouldn't be any problems."
"Mhm. The doctors said I'm healthy. Very healthy. So don't you worry."
The official report that arrived a few days later confirmed it.
Aside from minor issues like slightly elevated blood pressure, there were no serious diagnoses. Both of them were in good health.
What a relief.
Meanwhile, my father was absolutely thrilled to see his biological age listed as early forties, despite being in his early fifties.
"Hahaha! Look at this! I've still got it! Forties! I'm in my forties!"
"Hmph, same here," my mother retorted.
"Oh, really? But my overall grade is an A!"
"So is mine!"
In any case, I was relieved.
"Then let's go look at some houses tomorrow."
The next day, I toured several real estate agencies with my parents. I had set a budget of one billion won. Since money was steadily rolling in anyway, I didn't mind spending it all.
With one billion won, we could easily find a reasonably large, clean apartment in Daejeon, even if it wasn't the most luxurious penthouse.
However, my mother proved to be unexpectedly picky in this regard.
"How about this place?" the agent asked.
"Oh, it's nice. What do you think, Hyunwoo?"
"It's good."
"Right? Honey, what do you think?"
"Hmm... Let's look at a few more places," my mother said.
"How about here?"
"It's nice. Honey, what about you?"
"Show us another one, please."
We spent several days touring apartments like this.
Finally, my mother found an apartment she liked.
"Oh, this one is quite nice."
I said coolly,
"Then let's sign the contract right away."
My computer remained on for an entire week.
This was to allow the AI, Genie, to harvest language data from the internet.
The data included everything from dense academic literature to live internet broadcasts.
"Yo, guys! Send some donations!"
"Love you, oppas~ ♥"
"Today, we're going to review this vacuum cleaner that's super hot right now!"
It wasn't just Korean broadcasts; the same went for English ones.
"If you're new here, make sure to...!"
"Babes! Don't forget to hit that subscribe button...!"
Thanks to this, a significant amount of highly dynamic, real-world data was gathered.
Although it wasn't completely sufficient yet, there was bound to be some noticeable improvement.
"Shall we test it again?"
I typed a somewhat crude Korean sentence into Genie's prompt window.
Input: "Monkeys f**king love ripe bananas."
Normally, a standard translator would translate the slang for "f**king love" into something tame like "really love."
But the English sentence Genie produced was far more vivid.
Output: "Monkeys f**king love ripe bananas!"
"Perfect."
I could confidently call this a working prototype now.
However...
...I need to run a wider variety of tests.
To do that, I visited a website called Knowledge Market.
It was a freelance platform where individuals bought and sold their expertise, offering services like drawing illustrations or teaching musical instruments.
Among the listings, there was a category for buying and selling translation services.
I signed up as a seller.
Nickname: God Tech Genie
Type: Seller
Specialty: Translation
I filled in a few brief credentials.
• Completed doctoral coursework.
• Extensive experience translating academic papers.
Then, I registered my listing.
Offering to translate for free.
It didn't take long for a flood of inquiry messages to arrive.
And among them...
...was a MeTuber.