The COVID-19 pandemic has largely come to an end. Two acquaintances contacted me to say they had contracted it, but both reported symptoms no different from a common cold and recovered quickly.

Deep Learning

When I first learned about the Turing Test as a middle school student who had just started programming, it filled me with excitement. However, I never imagined that an AI capable of passing it would suddenly become a hot topic, let alone be freely available to everyone within a few months.

As I wrote last year, generative language models like GPT work by outputting continuations based on preceding text.

For example, if you write “To kill two birds with one stone means,” it outputs candidates for the continuation of that text. In the case of language models tuned for Q&A, for instance, if you write “Q. What does ‘killing two birds with one stone’ mean? A.,” it can output an answer to the question. (The format varies by language model)

On the other hand, BERT, which was mainstream until last year, is primarily a model for text classification and inference. For example, if you write “Today is [MASK] weather,” it outputs candidates like “nice” for the [MASK] portion. By the way, [MASK] is not always necessary—it can also output candidates to replace “nice” in “Today is nice weather.” It’s versatile through fine-tuning and creative pre-training methods, and recently, a voice generation method called “Bert-VITS2” that reads text aloud has become popular.

While BERT remains a useful language model, the generative approach proved more versatile, and as research progressed in this direction, it began achieving remarkable results.

However, ChatGPT has extraordinary flexibility and accuracy that cannot be explained by this alone. I think it combines technologies beyond deep learning, but I have no idea how it works.

Honorific Language Business

I wrote last year that I would try out ChatGPT, and when the API was released in March, I promptly implemented text generation functionality in the honorific translation app.

The honorific translation app had been in beta for a long time.

While many people use it as an app to check honorific text they’ve written, it lacked features for those who struggle with writing honorifics. If this app had existed when I was a student with poor honorific skills, I don’t think I would have found it useful. So I believed I couldn’t release an official version until it met broader needs.

With this new generation feature, I felt the missing elements were finally in place. As a result, the honorific translation app was finally released as an official version 2 years and 4 months after its initial launch.

However, I keenly feel the need to find and build technical assets that won’t become obsolete even as AI advances. With this in mind, I’ve been expanding the honorific conversion dictionary since launch—something that’s difficult for current AI to create—and this proved to be the right decision.

I was originally considering creating a Q&A box app, but given the trends of the times, I decided to create a location-sharing SNS instead.

Creating a location-sharing app was more difficult than I imagined, and from December last year when I started until early March this year, it was an extremely stressful period.

Once the basic features were complete, development became enjoyable. However, the challenge is that I can’t focus much on it due to low business profitability. I think it will take another two years to achieve the app form I’m aiming for.

Unfortunately, Batty’s release was too late to catch the first wave, but maintenance costs are minimal, and there’s potential for popularity, so I’ll continue developing it.

Additionally, apps using smartphone GPS functionality have many unexpected behaviors that you can’t understand without building them, and I think accumulating such know-how has become a company asset.

Incidentally, though not where I expected, a powerful foreign app emerged, and the initial user acquisition competition settled with others following the most popular one in Japan. However, despite both apps ranking high in Japan’s overall App Store rankings, they would suddenly disappear even from category rankings on certain days, making it impossible to judge which was actually more popular.

App Research

Actually, since I’ve roughly figured out the trends of popular apps, there’s less reason to conduct annual app research.

Just browsing the list gives me a general idea of trends, so I don’t need to install everything.

That said, the work has become more efficient, and if I gradually install apps as a daily habit, it’s not much of a burden, so I plan to continue.

However, I haven’t made much progress on 2023’s research because this year was too busy.

I haven’t had a single year where I could work on this calmly.

Studying English

For over ten years, I’ve thought that AI advancement might make English study unnecessary.

Additionally, as a programmer, I frequently encounter English, and with machine translation, there were few situations where I struggled, so I never studied seriously.

However, as AI development became reality, I realized the possibility of reading becoming useless is minimal. This is because subtle nuances are lost when translating to Japanese. Even human translators have limitations, so to know accurate information, you must be able to read the original text.

On the other hand, I predict that for writing, AI will eventually be able to accurately translate plain, easy-to-understand Japanese. It will probably write politely or include slang depending on the situation, so the ability to express your thoughts clearly in Japanese might be more important.

Also, studying Japanese grammar for keigo.jp sparked my interest in English. Particularly, I wanted to experience the sensation of naturally reading English as a second language without converting it to Japanese in my head.

I’m studying by reading The Big Fat Cat's World's Easiest English Book.

I encountered this book as a student but gave up because it seemed I’d graduate before it would help with school English.

This book advocates not blindly memorizing English word meanings, but developing a habit of inferring meaning from context before checking the dictionary, and becoming able to understand through experience gained from reading large amounts of text. However, to get good scores on school tests every few months, you need to memorize direct translations of English words.

Reading it now—or perhaps because I’m reading it now—I think it’s a very clear and excellent book.

It will probably take 3-5 years before I can read naturally, but since there’s no deadline, I’ll take it easy.

Music

This year I listened to many songs by Hisaaki Chiku.

Also, BIGMAMA’s Mathematics | RULER was excellent.

Bonsai

This year was unprecedentedly hot, so about 30% of what I received this year died within the year.

I was too busy to photograph many that bloomed, so I’m skipping this year’s bonsai article.

My teacher said that if the same climate continues in future years, we’ll need to fundamentally change our care methods.

Ring Fit Adventure

After clearing the third playthrough last year and resetting, I completed another round (4th playthrough), and finished the 5th playthrough in July this year.

When I cleared three rounds, progressing carefully through stages made the latter half too easy, so this time I’m not taking detours and decided to do missions after finishing the 6th round.

Additionally, since taking the “Rush Attack” skill that allows occasional extra attacks makes daily exercise inconsistent, I avoided it, resulting in my health being one-quarter less than maximum, making for quite a challenging playthrough.

These enemies were the most difficult in the 5th playthrough:

I couldn’t use items, and receiving attacks from three of the four enemies meant defeat. I could only keep healing until a lucky turn with no damage occurred, taking about an hour to defeat them.

In the 6th playthrough, with more money available, enemies became stronger but I progressed smoothly, thinking I’d clear it right at year’s end.

Same place as the 4th playthrough, but stronger—receiving attacks from two of four enemies means defeat. Moreover, defeating one enemy with a three-enemy attack skill takes three turns.

It was genuinely tough, and even if I level up, I’m nearly 70 levels below the appropriate level, so I don’t know how much I need to increase.

So I gave up on clearing it this year, but without a goal and with work getting busier, I lost motivation and barely played Ring Fit from around October.

I think I’ll have cleared it by 2024.