EXAONE, transcending language barriers into the world

Image 1. Sunkyoung Kim from EXAONE Lab at LG AI Research

LG AI Research's “Advancing AI for a better Life” is more than just a mission statement. It is the philosophy and special bond that unites researchers striving to make the proposition “Researching AI for a better life” a reality. Sunkyung Kim, an AI researcher at LG AI Research's EXAONE Lab, embodies both a deep understanding of data and a warm perspective focused on people. What exactly is the “technology that brings us closer together”?

 

AI, the special link

Image 2. Sunkyoung Kim having a relaxed conversation in the lounge

Q. What led you to begin researching AI after majoring in library and information science. What common ground did you discover between the two fields?

There was a professor in our department who had dual degrees in Library and Information Science and Computer Engineering. Influenced by them, I took a computer engineering class during my first year. In that class, I found it incredibly appealing to solve problems that had only been covered theoretically before, and to realize that the technology I created could bring convenience to others. I think that's when I started seriously considering a career shift toward engineering.

Developing a service that recommends weather-appropriate clothing for clubs was a major turning point where I encountered deep learning. I realized that through artificial intelligence, which can solve more complex and diverse problems through a training process, rather than simply performing calculations under fixed conditions, we could get closer to people's daily lives.  That's why I entered the Department of Computer Science in graduate school and began full-fledged AI research. 

Both library and information science and AI share a similar goal: they are disciplines that “research and utilize data to improve people's lives.” Both fields study how to gather information and how to build services that people can conveniently use. While library and information science focuses on libraries as the center of these processes, AI manifests itself through computers as the outcome known as a “model.”.

Q. Was there a specific reason you joined LG AI Research? Please tell us what you liked about LG's AI research.

The mission of LG AI Research, “Advancing AI for a Better Life,” was the decisive factor. While doing my master's, I researched a field related to pre-training Korean language models, and I once worked on a project to build a model to achieve first place on a target benchmark. Fortunately, we did achieve first place, but I was left with a strong sense of regret that this model wasn’t easy to use in practice. 

Then, I came across the mission of LG AI Research, and it struck me that LG AI Research isn't just about developing high-performance models, but about conducting research that can provide practical help to many people. I suppose you could say it aligned with my vibe? That's why I decided to join LG AI Research.

Q. How do you find time to fully focus on yourself amidst your busy research schedule? Tell us about hobbies or activities that inspire your life.

I'm not particularly athletic, but I try to enjoy a variety of workouts. I set aside time for exercise at least twice a week, and during those sessions, I strive to focus entirely on myself and the present moment. For example, I reflect on what I have and center my mind through moments like the impact of the ball hitting the tennis racket, my heart rate and my breath while running, or the focusing on my joints and muscles when holding a yoga pose.

Meeting people is also an important part of gaining inspiration in life. I especially enjoy talking with people from different fields. When I encounter perspectives and thoughts different from my own, I feel like my world expands. Since my undergraduate major wasn't in developing, I have many friends working in diverse fields. Thanks to that, I seem to be in an environment where I can encounter various viewpoints. At work, I participate in the AIRWX Crew, a productivity group, and run a writing community. I also actively seek out conversations with colleagues from diverse roles within the company.

 

EXAONE, the challenge toward global AI

 

Image 3. EXAONE Lab Pretraining & Adaptation Squad (From left) Soyeon Kim, Eunbi Choi, Sunkyoung Kim, Yireun Kim

 

Q. Please introduce your squad and members at EXAONE Lab, along with the work you’re currently handling.

I’m responsible for collecting, processing, and generating data for EXAONE's pre-training within the Pre-training & Adaptation Squad. My primary focus is handling Korean and multilingual data. Beyond simply preparing the data, I also participate in developing methodologies for effective learning and creating benchmarks tailored to the data's purpose^[1][2].

 

Q. Understanding and reflecting the subtle nuances of language and cultural context is not an easy task. How about the challenges involved in securing and refining high-quality data, and how you're addressing them.

Having been born and raised in Korea, I believe there are inherent limitations to how well I can reflect cultural differences. While I enjoy learning various languages and can read and understand Japanese, Chinese, Vietnamese, and Spanish, it's fundamentally difficult to ensure I've accurately reflected the cultures of those language regions in the data. To address this, we employ native-speaking reviewers and have established our own standards for cultural fluency in language to review the data. 

Furthermore, while the volume of training data is massive, amounting to at least tens of millions of examples, data for languages other than English remains scarce. Therefore, we’re also researching methods to enhance multilingual performance, including data structuring methodologies for knowledge transfer from English^[3].

 

Q. What is the most important consideration during the data construction process to achieve performance capable of competing with global models?

The most important thing is the license, and next is the quality of the data. This is because even the best data can become a risk factor for the model itself if there are licensing issues, potentially creating a situation where it cannot even be challenged. It’s also important to select data from the licensed data that can match the performance of global models. This is because models built efficiently with minimal data demonstrate higher growth potential.

 

Image 4. Team photo from hiking trip to celebrate EXAONE 3.0 model success

 

Q. While researching and developing EXAONE from 2.0 to 4.0 and EXAONE Deep, is there anything that stands out in your memory?

Before the EXAONE 3.0 model was released, we went hiking together as a team to wish for success. The day before, heavy snow suddenly fell, and we debated canceling, but we decided to challenge ourselves with our shared determination. It turned out to be a special experience, enjoying the snowy mountain scenery as we reached the summit. We learned firsthand that while unexpected hardships can make the journey arduous, reaching the summit reveals a brilliant view. This experience helped us endure the challenging path through model development and launch.

 

Q. What do you think is EXAONE's unique strength in the technological competition with global big tech companies?

In one word, I want to emphasize efficiency. The field of LLM development places great importance on vast data and abundant infrastructure that it can be called a “battle of resources.” According to scaling laws, achieving high performance is known to require more data and larger model sizes as success conditions. Despite this, EXAONE competes on equal footing with global models even under the difficult constraints of having less data and fewer resources compared to global big tech companies. I believe this demonstrates EXAONE's tremendous strength.

 

Q. What are the most notable current global AI trends, and how is EXAONE reflecting them?

As we enter the era of agents, Large Language Models (LLMs) are expanding their capabilities to function as a single computer by leveraging existing applications. OpenAI and Anthropic are also releasing several features this year with computer usage in mind, and among these, benchmarks and related techniques enabling LLMs to act like browsers are being extensively researched. 

LG AI Research is also developing models aimed at enhancing the agentic capabilities of the next-generation EXAONE to support multiple tools, starting with support for the Tool Calling feature in EXAONE 4.0.

 

Image 5. LG AI Research's booth at ACL 2024 conference

 

Q. What has been the most rewarding or happiest moment for you while working at LG AI Research?

Operating the LG AI Research booth at ACL 2024 and personally introducing EXAONE 3.0 to people of diverse nationalities was truly joyful and rewarding. Before that, while conducting various thought experiments and development with colleagues in the office and achieving good performance, I always harbored the question: “Can the models I create actually have an impact on people?” Being able to introduce it verbally to people worldwide on-site and hear their reactions was incredibly encouraging. This experience gave me the confidence that what I was doing could make a tangible contribution to the world. It quenched the thirst I felt during my master's program and allowed me to truly embody the mission of “Advancing AI for a Better Life,” making it the most memorable part of the experience. (YouTube : LG AI Research Unveils EXAONE 3.0 at ACL 2024!)

 

Next-level goals

 

Image 6. Sunkyoung Kim discussing what makes good AI

 

Q. As an AI researcher, please tell us about your future goals

I'd like to share two major goals. The first goal is for EXAONE to become synonymous with “Foundation LLM.” Just as Koreans naturally grew up thinking LG = home appliances, I hope EXAONE will permeate people's perceptions so that Foundation LLM = EXAONE becomes a widely accepted notion. Of course, I hope to contribute to making that happen.

The second goal may seem unrealistic, but it’s to reach a point where there are no more goals left to achieve through AI. AI technology still has many areas needing improvement, and the world is full of pressing challenges. I want to solve all these problems and create AI technology so perfect that there are no more problems left to solve.

 

Q. What do you think constitutes "Good AI"? What role do you think EXAONE should play in achieving that kind of  AI?

I believe that “Good AI” is one that’s widely used. Recently, AI has demonstrated a level of knowledge sufficient to win gold medals at the IMO Olympiad. Now, a lot of AI is showing human-like performance at a high level of knowledge across various benchmarks. EXAONE also consistently demonstrates excellent performance on multiple benchmarks with each new model release. 

However, for people to truly perceive something as “Good,” it has to convey not just “intelligence,” but also the value of “usability.” Since many people still feel overwhelmed about how to use AI, I think EXAONE should serve as a bridgehead. Just as many people communicate using “language,” we at LG AI Research are constantly thinking and striving to ensure EXAONE helps people easily communicate with AI.

 

Q. For students dreaming of becoming future AI researchers and wishing to work at EXAONE Lab, as a senior, please give them some advice.

Do not limit yourself; dare to challenge the problems you face. Despite starting with a humanities background outside of engineering, I made it here by tackling small, achievable tasks and persistently working to break through my limits. Don't let the thought "I can't" confine you—just start trying. Doing nothing yields nothing, but challenging yourself will teach you valuable lessons, even from setbacks. Keep pushing forward against those challenges.

Q. Finally, what does EXAONE mean to you? 

To me, EXAONE is like my own child. Though I've never had children myself, that feeling grows stronger with each model release. Like a child left by the water's edge, it's a precious thing I worry about, and I want it to be loved through positive evaluations from others. I’m making EXAONE with the love of a parent, willing to sacrifice a bit of my daily life to nurture and develop it well, hoping it will achieve high performance.

 

As Sunkyung recounted climbing a snow-covered mountain and gaining the insight that “beyond unexpected hardship lies brilliant rewards,” EXAONE's challenges can be seen as a journey of perseverance and teamwork that overcomes uncertainty, going beyond mere technological development. EXAONE Lab is now leading the new trend of AI Agents that understand human intent and leverage diverse applications, focusing on becoming the center of the future AI ecosystem. We look forward to your continued interest.