Diary - 2024 in Review

01 Jan 2025

The year-end review that I resolved to write every year ever since I became a developer in 2022. This is already my third one. In 2024, whenever I had to write a long piece I usually drafted it with GPT, but for the review I write it out character by character, by hand. Only then, it feels, can I properly sum up the year, reflect on it, and carefully set my goals for the year to come. I sketched out what happened in 2024 and how I want to live in 2025!! Brace yourself, this is a long one~~

(Along with the 2024 summary that Apple Intelligence picked out)

In 2024…

Getting to do the work I wanted

When I, who majored in autonomous driving, robotics, and reinforcement learning in grad school, decided to join SKT in 2023, one big reason was a reinforcement learning technique used in LLMs called “Alignment-tuning.” Ever since grad school I had been continuously interested in “the methodology of training a model to match human preferences,” also known as RLHF, and I thought this would be a chance to actually implement it using a company’s massive resources. Up through the early-to-mid part of 2023, there was still a lot of internal debate over whether this methodology had any real value, so it never turned into actual work, but as 2024 came around I was finally able to start working on it.

The work I did over the year can be broadly divided into four areas.

• Building and streamlining the alignment training pipeline
• Building alignment data
• Building the LLM evaluation pipeline
• Collaboration efficiency work

First, I started by laying out every alignment-tuning methodology that had come out so far and studying them. Over about two to three straight weeks I think I read something like 50 to 100 papers, and as you keep digging in, you slowly start to see which methodologies are the good ones. The moment you feel you’re learning, the moment you feel you’ve grown somehow. I found this process truly fun and exhilarating.

Reward model training, RLHF with PPO, offline DPO, online DPO — with five or six of us, we steadily built up the code, applied a variety of methodologies, and created our own repo. The parts I enjoyed most were the engineering tasks of streamlining the pipeline. We solved training bottlenecks using vLLM, and sped up training by incorporating things like batch packing. I can’t wait for our model, A.X LLM, to be open-sourced so that I can weave this whole process into a single technical report. (How long must it stay private…ㅠㅠ)

There was also something I accomplished by setting a goal for myself, not because anyone told me to. (Honestly this is the most rewarding and memorable of all.) As I trained models, the biggest bottleneck I felt was things like checking the parameters of models trained by other colleagues, or the difficulty of making fair comparisons because the evaluation process differed slightly from person to person. So I built an A.X LLM evaluation page where you could run LLM evaluations with just a click. I also built a web page where you could check a given model’s spec, backbone model, and what data it was trained on, all through a GUI.

I have always considered visualizable forms of communication channels, like these viewers, to be important. That’s partly because I believe a single image can persuade people of more than a page of text. The team members really used these tools well, and hearing feedback that it made collaboration easier was incredibly rewarding.

In any case, being able to do the work I most wanted to do when I joined the company was wonderful. People often say that at a big company you end up doing what you’re told rather than what you want, but I think I was really lucky. Personally, even though alignment-tuning is mainly used in LLMs, I have a hunch that someday it will create a major breakthrough in the field of robotics. When that day comes, this experience now will be a huge asset!

Big disasters

I remember causing three big disasters… just thinking about those moments makes me shudder.

The first was around mid-May. From February I had taken on building the RAG data, and that data had been going into training continuously since around April. But in mid-May my mentor looked at the data and said,

Jaekyung… it turns out you put the context into the “meta” field. That field was deprecated back in February… so all this time the context hasn’t been used in training.

For nearly a month, the wrong data had been going in. I should have asked my mentor to review it together when I first made the data… I should have put assertion code into the training code to check that the data was being used correctly… regrets came flooding in. In the end, the model that was supposed to be released in May got pushed to early June.

The second was at the end of June, a task where I trained the A.X model on data that another internal team had built for us and delivered the model. The problem was that, due to a GPU expansion project, I had to train the model within a week. I worked late every night for a week training the model, but the performance just kept not improving.

Then on the very last day, I discovered there was something wrong with the data I had been given. So absurd… but it’s partly my fault too, since I didn’t look closely at that data and just assumed it had been handed over properly and used it in training. I should have checked more carefully. In the end, the team lead got chewed out badly by the higher-ups and even delayed the GPU expansion project by a week for us. He told me it was fine, that I’d done it while trying to do well under a brutal schedule, and he seemed to take all the arrows coming from above himself, which made me feel so sorry and grateful at the same time.

The third happened in early November, when another SK group company sent an evaluation report by email that contained criticism of A.X. The problem was that at the time I was so swamped with the SK Summit event that I just skimmed the email and only verbally told the team lead, “I hear our model’s performance is a bit lacking.” The problem was… that report was written in a very aggressive tone. A week later at an executive meeting, our executive heard the news for the first time from an executive at the other group company and got absolutely furious — asking why nobody had relayed this content, and who had first received it. It was me.

That said, was I just sitting on my hands? Not exactly. My mentor and I had set up a plan and were in the middle of building additional data and training. The regret that lingered was that if I had just forwarded that email to the team lead the moment I received it, there would have been no problem at all, and that was so frustrating. Communication skills are just as important as development skills.

Honestly, all three incidents were things I worried about terribly at the time, but writing this year-end review now and looking back on them, it feels like they all worked out well. I felt that if you don’t hide your mistakes, don’t stop, and try to fulfill your responsibilities to the end, many people around you will help you. Of course you should minimize mistakes as much as possible, but when you do make an unexpected one, I resolved to always apologize to everyone involved and restore their trust by working even harder. I feel I’ve grown a step through these mistakes.

Work beyond development

From mid-September to mid-November I worked frantically. Not on development, but because all the other tasks were so busy. Making slide decks, becoming a hiring interviewer, even running an SK Summit booth… I pulled an enormous amount of overtime that I normally never do.

Around the Chuseok holiday I was suddenly drafted into the “slide deck factory.” That is, I ended up on a temporary task force making executive briefing materials and external lecture materials — which, of course, is not development work at all. To make a 15-slide deck, three of us pile on and crank out nearly 50 slides. Out of these, only 15 survive. We repeated this for two to three straight weeks with no weekends off. I guess I finally feel like a proper big-company employee?

Still, I didn’t want to think of it as meaningless repetitive work, so I tried to find meaning in it. In the end, the most important skill in making slide decks seemed to be detail. Back when I gave seminars in my lab, my professor would occasionally give this kind of feedback.

When you write a month as a three-letter abbreviation, you must put a period after it. Not “Feb” but “Feb.” The moment someone sees this on the first page, that’s all they see and they won’t even listen to the rest — there’s always at least one such person. That person might be someone important to me. When you make a PPT, always pay attention to these little details.

At the time I thought it was just the professor’s personal opinion, but I heard the exact same feedback at the company. I had to pay attention to every detail — text position, font size, alignment, symmetry. If you absolutely must persuade an unspecified crowd of people you don’t know well, detail is essential. This was the skill I most lacked, being the type who jumps into anything quickly and puts out results quickly, and this slide-deck-factory lesson was a big help.

After wrapping up the slide deck factory, the story I heard over dinner with the executive I’d worked with was interesting too. He’d come from Samsung, and he said that during his first year he made slide decks the whole time and only a single slide ever got used.

“Being good at making slide decks itself means you have high metacognition. That training of doing nothing but make slide decks for a year after joining Samsung ended up helping me internalize methods for persuading someone.”

Did my metacognition grow during this period? I hope that the next time I’m summoned to the slide deck factory, I can pull it off a little more smoothly.

In the middle of the slide deck factory, I also served as an SKT hiring interviewer. Hiring was harder than I expected. What stuck with me was the interviewer training point that “if you don’t set clear criteria, you end up hiring people similar to yourself.” I really could see myself trying to give very generous evaluations to people similar to me.

They say that if you only gather similar people, an organization’s diversity breaks down. As the saying goes, the company stagnates. I tried to set up a clear checklist for every task and evaluation in the interview process, and I was able to look at applicants more objectively.

For now, my evaluations as an interviewer had only a negligible effect on the company’s hiring results, but I believe that someday, when I reach a higher position or build my own company, this hiring experience will be a great strength.

Finally, over November 4th–5th, I was put in charge of the A.X LLM booth exhibit at the SK Summit. I came up with the idea of Attention visualization, a fun way to showcase A.X LLM’s performance, myself, and I think I wrapped up the booth operation successfully and to great acclaim. While running the booth, many people came by to ask questions, and I met a lot of acquaintances too. The booth operation also got good reviews, which made me happy.

Honestly, my colleagues around me might think I get dragged around doing useless tasks, but I think it’s a good sign that I get called on so much for work beyond development. I don’t think I’ll be able to do nothing but development my whole life, and I see it as getting opportunities to accumulate diverse experiences. In fact, looking back at the personal statement I wrote when joining SKT, I wrote that “I want to be the first person colleagues can comfortably turn to when they face a problem.” Isn’t this part of the process of becoming that person?

Starting to write again

This year I began participating in Devocean, SK Group’s developer community. My initial goal was to post one blog article per month. In the end I wrote 10 blog articles and even did a Devocean Young interview once. Excluding the busy September–November stretch, that works out to about one a month, so I’m satisfied.

There are two reasons I write blog posts like this. The first is the hope that it helps someone. Kim Jin, who’s on the team next to mine, writes a tremendous number of posts on Devocean, and I got so much help from his LLM-related articles. I wanted to be that kind of person too, which is why I started participating in Devocean.

The second is that by writing, I come to understand the content more deeply myself. Reading something, presenting on something, writing about something — personally, of the three, I think writing is the one that’s only possible if you understand the content the most deeply.

In the end, the number of articles I wrote put me 12th in the whole community. I started thinking I should write enough next year to aim for the top 10. Also, my article Building a Telegram bot that delivers the latest AI paper info using the OpenAI API was selected as the most-liked blog post across all of Devocean in 2024. Thinking that so many people liked it and got help from it makes me so proud.

In the second half of the year, I started being active on LinkedIn. The posts I write on LinkedIn are comparatively simpler in content than the ones on Devocean. The biggest advantage is feedback. I’d actually been running a personal blog before, but since there weren’t many comments or reactions, it felt less rewarding. Writing on Devocean or LinkedIn, being able to see people’s interest in numbers, is a big motivator.

I could also see what kinds of content people are most interested in. Among the posts I put up in 2024, the one that got the most views and likes was my post on Attention noise, which earned 25K views, 400 likes, and 50 reshares. With this kind of numerical feedback, I was strongly motivated to keep writing. In that sense it struck me as a really well-built platform, and I felt it was the SNS that suits me best.

Speaking experiences

This year, gratefully, there were many places that invited me.

• 2024 SKT Conversation with Junior Talent
• LIKELION Central Hackathon judge
• SK Summit A.X LLM booth operation
• Weights & Biases 2nd meetup speaker

The W&B meetup in particular was an external speaking gig I’d really wanted to do, and being able to present about the work I do was wonderful. I’d been a devoted W&B user since 2022, so being invited there made me so proud.

In any case, I plan to keep up my Devocean and LinkedIn activity hard next year too. So that the world becomes one where someone gets help through what I wrote, and can pass that help along to yet another person. Just like with writing, I hope to have more speaking experiences next year too. Whenever I stand in front of people and present, I always get a lot of energy and the dopamine flows hard.

Other odds and ends…

I got the bartender’s license (Joju-gineungsa), one of the goals I’d set last year. (Not metal-casting.) Strangely, once I got it, all desire to drink cocktails vanished. Maybe it’s because there are fewer tasty cocktails than I expected. I’ve only acquired a taste for highballs and whiskey, so I keep drinking those. Everyone asks what you can do with a bartender’s license, but it’s nothing special. It’s cool, right? Have a drink~

This year I was also fortunate not to get sick. The basketball skill training (Stay Focus) I started in February in particular was so much fun. My shots seem to be going in less and less, which is a problem, but my technique also seems to be getting a little better, maybe? I also played tennis faithfully once every two weeks and had a great time.

I tried to keep up with deep learning trends through DPST activities and an ML/DL study group with SKT peers, and through the StudyMate book club I read a few books I normally wouldn’t have.

Personally, I think 2024 was a life of just the right golden balance — working hard and enjoying life hard. Thanks to that, I think there weren’t many moments of extreme stress, and I didn’t get sick either.

In 2025…

What is it that I truly want? It’s important to precisely define when I feel exhilaration — from “Make the World Want What You Have” (by Choi In-a)

In my case, I feel exhilaration when I make the invisible visible, when I can explain something difficult in an easy way, when I can persuade someone with a very simple demo rather than a long explanation.

I also feel exhilaration not so much when I do well, but when someone wants me. Take basketball as an example: rather than scoring a lot, I prefer to hear someone say “I want to play with Jaekyung.”

Pay it forward: Rather than paying back the help you received from a specific person, giving the things you’ve received so far to an unspecified crowd of juniors

This was my favorite word that I learned this year. I’ve developed a life goal of sharing the countless pieces of advice and teaching I received on my way to becoming who I am now with many people. Although I’m still lacking and can only share very small things for now, until the day comes when I can share bigger things and through that pay-it-forward spreads more widely! I want to grow hard and share my experiences.

Lastly, I want to seek out wonderful mentors who already live this kind of life. I’ve always found it hard to get close to seniors. When I entered the first year of high school, at the freshman welcoming party and club auditions the seniors made us do a ton of weird things, and I think those bad memories play a part. Approaching seniors and people older and more experienced than me has always been hard, but this year I want to try to get much closer to them.

This year I also have a big event ahead of me: marriage. Practically speaking I’m already living as if married, but using the wedding as an excuse, I can get in touch with friends I haven’t seen often and see their faces on the wedding day.

It’s a chaotic time, not very year-end-like, and a lot of sad news has been reaching us, but I hope everyone can bravely overcome it well. May only good news reach us next year.