I recently gave a talk at Google Developer Groups on Campus (GDGoC), Google’s official program that empowers students to grow as developers and leaders. In my talk on "The Sovereign AI Strategy of Korea", I also introduced my research, Toward Sustainable AI: Holistic Co-Optimization Across Training, Inference, and Deployment, and shared perspectives on the direction Korea should pursue and the contributions we can make moving forward
I attended the 2025 AI Graduate School Symposium, hosted by the Ministry of Science and ICT under the theme “AI with Us: Future Innovation with AI Graduate Schools.” The event brought together 19 graduate schools (10 AI Graduate Schools and 9 AI Convergence & Innovation Graduate Schools), along with leading companies and institutions such as Samsung Electronics, LG AI Research, HD Hyundai, NC AI, and the Korea Electronics Technology Institute. The program featured keynote lectures from AI experts, graduate school and corporate exhibitions, student poster sessions, and presentations of outstanding graduate research achievements. Among the sessions, the keynote that left the deepest impression on me was “Challenges in AI Inference Serving and Direction for Future System Architectures for Cost-efficient AI Inference Serving” delivered by Professor Nam Sung Kim from the University of Illinois.
I participated in the 2025 Tech-Day hosted by the Human-Inspired AI Research Institute (HIAI) at Korea University. The first part of the event focused on the latest research trends in large language models (LLMs) and their connection to business applications, along with an introduction to HIAI’s industry-academia collaboration programs. The second part highlighted HIAI’s core technologies and featured success stories from companies that have effectively leveraged these innovations. It was a meaningful opportunity to see how cutting-edge AI research translates into real-world impact and collaboration.
I had the opportunity to attend the 2024 SK hynix Tech Day, an event designed for graduate students specializing in semiconductor design, systems, devices, and processes. I joined the System session, where I attended a keynote and discussion led by Vice President Youngpyo Joo on SK Hynix’s core technological strengths and future vision. It was a valuable experience not only to gain insights into the company’s technology roadmap but also to better understand its corporate culture.
I had the privilege of attending a seminar hosted by Professor Sung Woo Chung, my undergraduate research advisor, where the distinguished speaker was Professor Nam Sung Kim from the University of Illinois at Urbana-Champaign, an IEEE Fellow and renowned expert in power-efficient computing systems!
Professor Kim delivered insightful presentations on one ISCA 2025 paper: LIA: A single-GPU LLM inference acceleration with cooperative AMX-enabled CPU-GPU computation and CXL offloading, two MICRO 2025 papers: NetZIP: Algorithm/Hardware Co-Design of In-Network Lossless Compression for Distributed Large Model and Re-Architecting End-Host Networking with CXL: Coherence, Memory, and Offloading. With an illustrious career spanning academia and industry—including early tenure at the University of Wisconsin–Madison, impactful research at Intel, and over 8,000 citations—Professor Kim’s work bridges device, circuit, architecture, and software, offering a holistic approach to advancing high-performance, energy-efficient computing. The seminar was both technically enriching and inspiring, highlighting the forefront of innovation in large-scale model processing and next-generation interconnect technologies. After attending, I can confidently say that I have become a big fan of Professor Kim.
As a senior member of the AI research community at Korea University, I had the pleasure of giving a talk at AIKU, the Deep Learning Club, where I introduced the paper vLLM: Efficient Memory Management for Large Language Model Serving with PagedAttention and shared insights from my ongoing research on optimizing LLM serving engines.
The audience was a group of highly motivated undergraduate students with a deep passion for AI and a genuine curiosity about what it’s like to pursue graduate studies in this field. I walked them through how PagedAttention enables efficient memory usage in LLM inference, making large-scale serving more scalable.
It was inspiring to see the undergraduate students’ enthusiasm, thoughtful questions, and eagerness to dive deeper into both the technical and research aspects of AI!
I had the incredible opportunity to attend the Agentic AI Summit 2025 at UC Berkeley, hosted by the Berkeley Center for Responsible Decentralized Intelligence!
I was honored to receive a student fee waiver, which made it even more meaningful to be part of this landmark gathering. The summit brought together a powerful mix of academic leaders, industry pioneers, and policy voices to explore the rapidly evolving landscape of Agentic AI. From insightful keynotes by Dawn Song and Bill Dally to thought-provoking panels featuring experts from NVIDIA, Google DeepMind, OpenAI, ML Commons, PyTorch Foundation, SAMSUNG NEXT, and more, the event highlighted cutting-edge research, emerging frameworks, and the future of intelligent agents across industries.
I had the pleasure of giving a talk titled "AI × Computer Architecture" at AIKU, the Deep Learning Club at Korea University. The audience was a group of highly motivated undergraduate students who are passionate about AI and curious about pursuing graduate studies in the field.
I spoke about the need to design AI systems that are not only powerful, but also efficient, sustainable, and accessible.
My key takeaway message was this:
To scale AI responsibly, we need:
– A shift from raw performance to system-level efficiency
– End-to-end optimization across the machine learning pipeline
– HW/SW Co-design
– A long-term, sustainability-driven mindset in AI system development
It was a great opportunity to share insights and spark conversations around building the future of AI responsibly!
Honored to Present at the K-Club World Conference: Next Intelligence for Planetary Solution
I’m thrilled to share that I was invited to present my poster, “Toward Sustainable AI: Resource-Efficient Model/System Co-Optimization for Large-Scale Generative AI,” at the K-Club World Conference: Next Intelligence for Planetary Solution.
Among numerous submissions, my poster was selected as an Outstanding Poster — with a competitive acceptance rate of just 13% — and was further invited for an Oral Presentation, which had an even more selective acceptance rate of 5%! I had the honor of presenting during Theme 4: Digital Revolution and Next Intelligence, specifically in Session 2 – AI in Today's Industry and Education.
One of the most memorable parts of the event was the opportunity to receive advice from distinguished professors. I was fortunate to speak with Professor Alok Sharma and Professor Vladimir Simic, both of whom shared thoughtful and encouraging insights on my research and career direction. Their perspectives were deeply motivating and have helped shape how I think about the long-term impact and direction of my work.
A heartfelt thank you to the organizers, mentors, and everyone who made this event unforgettable. I'm excited to keep moving forward in pursuit of sustainable, impactful AI.
The Rise of Local AI: Reflections from WERT TECH Summit 25
I recently had the opportunity to attend WERT TECH Summit 25, where I was particularly struck by a talk titled "The Return of Local AI" by Choi Hoon, Head of Business Development and Sales at Upstage. Among the many sessions, this one stood out—not only because of its timely insights but also because it resonated with my own growing interest in the concept of Sovereign AI, a theme that is gaining traction in political and industrial discourse alike.
Choi Hoon's talk painted a compelling picture of where we are headed. According to him, 2025 marks a pivotal moment—a year when AI begins to surpass human capabilities across a wide range of domains. However, the way general users and enterprise customers engage with generative AI is diverging significantly.
He outlined a key challenge in the current AI landscape: as competition to improve general-purpose foundation models intensifies, training costs are skyrocketing into the billions of dollars, yet marginal gains in output quality are diminishing. This is pushing AI model providers to shift focus from performance alone to non-performance factors such as cost efficiency, service reliability, and domain-specific features. In other words, the race is no longer just about who has the smartest model—but who can deliver the most practical, sustainable, and customized value.
On the flip side, inference costs are dropping rapidly, making powerful AI services accessible to more users at lower prices. This democratization of generative AI is enabling businesses to adopt multi-model strategies, choosing from a variety of models based on use case, budget, and performance needs. We're already seeing the rise of differentiated pricing and training strategies depending on model size and specialization. For smaller LLMs, their limitations can often be overcome through fine-tuning or Retrieval-Augmented Generation (RAG).
One of the most fascinating points was the shifting trend in B2B AI architecture. Traditional methods like prompt engineering and fine-tuning are starting to decline, while RAG and agent-based systems are quickly taking center stage. This reflects a broader transformation: enterprises are no longer just consuming AI—they are integrating it deeply into their internal systems and workflows.
As I listened, I couldn’t help but feel that we are witnessing the true beginning of the AI era—a time when the focus shifts from just building smarter models to crafting better user experiences. With RAG emerging as a backbone for internal systems and agent technology becoming crucial for orchestrating interactions with external services, the next frontier of AI will likely be defined not by model parameters, but by how seamlessly these tools integrate into real-world usage.
This session made it clear to me: the age of Local AI and sovereign intelligence is not just coming—it’s already here. And in this new paradigm, it’s not the biggest model that wins, but the most adaptable one.
Everything Comes Back to Edge: Highlights from the 2025 Advantech Embedded Design-In Forum
I recently had the opportunity to attend the 2025 Advantech Embedded Design-In Forum, held under the powerful slogan: “Everything Comes Back to Edge.” The event brought together some of the most influential players in the embedded and edge computing ecosystem, including Microsoft, NVIDIA, Qualcomm, FURIOSA, Intel, 마음AI, and Superb AI, who all sponsored the event.
One of the key themes that resonated with me throughout the forum was the concept of Physical AI — the convergence of artificial intelligence with robotics and edge computing to enable real-world autonomy and interaction. It was inspiring to see how the frontier of AI is moving beyond virtual environments and into physical spaces, where embedded systems play a critical role.
Among the presentations, NVIDIA’s Isaac GR00T N1.5 particularly stood out. It’s an open foundation model designed for generalized humanoid robot reasoning and skill acquisition, capable of understanding and acting on multimodal input — including language, vision, and proprioception. GR00T N1.5 isn’t limited to a single robot or task; instead, it supports cross-embodiment generalization, allowing developers and researchers to fine-tune it with synthetic or real-world data tailored to their own humanoid platforms.
The version showcased, Isaac GR00T N1.5-3B, is a medium-sized model built using pre-trained vision and language encoders. What’s particularly exciting is its use of a flow-matching action transformer, which enables it to model sequences of actions in context. This opens up new possibilities for robotic manipulation tasks across diverse and unstructured environments — a leap forward in enabling more capable and adaptive robots at the edge.
As someone deeply interested in system design and optimization for intelligent machines, this forum was not only a source of knowledge but also a strong reminder that the future of AI is tangible, physical, and increasingly decentralized.
The forum made one thing very clear: the edge is not just a deployment target — it’s the launchpad for the next generation of intelligent systems. From embedded inference to real-time decision-making in robotics, “Everything Comes Back to Edge” is more than a slogan — it’s a glimpse into where innovation is headed. The edge isn’t just where data lives anymore — it’s where intelligence is truly taking root.
Accelerating Innovation with Alibaba Cloud — Insights from Alibaba Cloud Day Korea 2025
I recently had the opportunity to attend Alibaba Cloud Day Korea 2025, and it was an eye-opening experience. Among the many insightful sessions, I was especially impressed by Mr. Jungjoo Yang, Executive Director at Red Hat Korea. His talk on “Accelerating Innovation through Red Hat and Alibaba Cloud” offered a compelling look into how open-source technologies and cloud-native infrastructure are driving enterprise transformation at scale.
Another highlight for me was the session titled “The Future of Intelligent Search with Elastic and Alibaba Cloud AI Search”, delivered by Kuan Zhou, AI Search Product Lead at Alibaba Cloud, and Jerry Zhu, Principal Solutions Architect at Elastic.
What made this session particularly fascinating was how closely it aligned with a research theme I had been exploring while drafting a recent corporate research proposal. The convergence of ideas was almost uncanny — the same challenges, the same visions, and even similar technical directions.
It was both exciting and reassuring to see that the problems I’m trying to solve — and the future I’m envisioning — are not only relevant in academia, but also actively being explored and invested in by leading tech companies.
Attending Alibaba Cloud Day Korea reminded me once again of the value of connecting academic research with real-world industry trends. It’s clear that the future of intelligent systems lies at the intersection of cloud infrastructure, AI, and open collaboration — and I’m more motivated than ever to be part of that journey.
I had the opportunity to attend the 'Forbes Korea '30 Under 30 Summit', an inspiring event that celebrates young innovators who are shaping the future across a wide range of fields. Each year, Forbes Korea honors 30 trailblazing individuals under the age of 30 whose work has the potential to create meaningful impact in Korea and beyond.
This year’s summit featured a rich lineup of keynote speeches, panel discussions, and pitches spanning social innovation, deep tech, consumer experience, art, and science. While many speakers stood out, one talk in particular left a lasting impression on me—that of Professor Jangsaeng Kim from Sogang University.
Professor Kim, who holds joint appointments in System Semiconductor and Electronic Engineering, developed the world’s first 3D vertically stacked AI semiconductor. This breakthrough addresses long-standing challenges in computing—namely, excessive energy consumption and limited chip area—by enabling compact, ultra-low-power AI systems. One real-world application of his work is real-time health monitoring via wearable devices, a promising step toward next-generation personalized healthcare.
Overall, the '30 Under 30 Summit' was more than just an award ceremony—it was a powerful reminder of the innovation and ambition driving Korea’s next generation of leaders. I left feeling inspired, and hopeful that gatherings like this will continue to foster collaboration, spark ideas, and push the boundaries of what’s possible.
I recently had the chance to attend a Campus Outreach event hosted by Google Korea, where I met with engineers from Google’s silicon team. One of them was Younghyuk Kim, a Silicon & Machine Learning Architect currently working on Edge TPU. He began his career at Samsung and now collaborates with Professor Jaeha Jung’s lab at Korea University—so the session felt both global and familiar.
During the Q&A session, I asked questions about the kinds of models TPUs are optimized for, whether they primarily target inference or training workloads, and how Google balances precision, DVFS, and thermal issues—especially on Edge TPUs. I was also curious about the challenges of kernel fusion and how unsupported operations are handled in TPU-based systems.
Overall, the session reminded me how system-level thinking—balancing power, performance, flexibility, and scalability—is at the heart of hardware design. It also made me appreciate the hidden complexities behind the chips that power AI at every scale.
I’m thrilled to share that I was selected as one of the top 10 teams in the AI Convergence Applied Technology Idea Competition, hosted by the Korea Electronics Technology Institute (KETI). As part of this achievement, I had the opportunity to visit KETI’s Pangyo branch (Intelligent Information Research Division), located in one of Korea’s leading tech innovation districts.
During the visit, I was introduced to some of the institute’s ongoing research initiatives and had the chance to meet with experts working at the forefront of AI convergence. One of the most meaningful parts of the experience was being paired with a mentor—Dr. Seungwoo Lee, Principal Researcher at the Data Convergence Platform Research Center. His insights and guidance will be invaluable as I continue to develop my project.
Our paper, "Not All Adapters Matter: Selective Adapter Freezing for Memory-Efficient Fine-Tuning of Language Models", has been selected as a review paper for the undergraduate course COSE405: Computer Science Colloquium at Korea University!
It’s a huge honor to see our work included in the curriculum and introduced to students exploring the latest advances in computer science. I'm truly excited that our research will be read, discussed, and critically reviewed by the next generation of scholars.
I had the opportunity to attend AWS Summit Seoul 2025, where I explored how AWS services and generative AI are driving technological innovation across key industries. The event was filled with insightful sessions, real-world success stories, and exciting new tech trends.
Among the most impactful sessions were: Large Scale Migration, Application and Database Modernization, VMware Oracle Modernization
These topics highlighted how businesses are not just adopting the cloud but transforming themselves through strategic modernization efforts. What resonated with me most was the shared vision of innovation as an ongoing journey, not a one-time effort. It reminded me that meaningful change takes sustained commitment — and the right tools and mindset.
Looking forward to applying these insights in my own work!
Teachers' Day Reflections: Gratitude for My Mentors
Recently, we celebrated Teachers’ Day, a day dedicated to honoring the legacy of King Sejong the Great and expressing gratitude to educators who have shaped our paths. As I reflect on this occasion, I’m reminded of the many mentors who have supported, guided, and inspired me throughout my academic journey.
First and foremost, I would like to thank Professor Young Geun Kim, my advisor, whose unwavering support and insightful guidance have been instrumental in my growth as a researcher. I’m also deeply grateful to Professor Sung Woo Chung, Professor Kim’s own Ph.D. advisor, who first mentored me during my undergraduate research years. Professor Chung, a leading expert in thermal modeling and thermal management systems, marks his 20th year of dedication to education and research this year—a truly meaningful milestone.
One lesson from Professor Chung that has stayed with me is the importance of "working smart." This simple yet powerful idea has become a cornerstone of my approach to graduate school and research. It continually reminds me to stay focused, intentional, and efficient in everything I do.
I would also like to extend my appreciation to Professor Seungho Choe, who first inspired me to pursue computer engineering, and to Professor Changhoon Kim, whose teachings have also left a lasting impact.
To all the mentors who have helped shape my journey—thank you. Your dedication has not only built knowledge but has also nurtured countless dreams, including mine.
I had the incredible honor of being invited as an oral presenter at NAACL 2025 in Albuquerque. It was a truly meaningful experience to connect with researchers from both industry and academia—across the globe.
Massive thanks to my advisor, YOUNG GEUN KIM; mentor and co-corresponding author, Changhoon Kim; and to my colleague, Yonglak Son. I’m also grateful to the members of the Intelligent Computer Architecture & Systems Research Lab (Korea University) for their valuable comments and insightful discussions. Lastly, special thanks to Jaewoo Jung from KAIST AI for his invaluable support with the Loss Landscape Analysis.
I had the opportunity to meet leading international researchers from big tech companies such as Microsoft, Meta, IBM, and Adobe, as well as distinguished academics in the NLP community. I also had insightful conversations with outstanding researchers from Korea, including those from NAVER, HD KSOE, and academic leaders like Prof. Heuiseok Lim from Korea University, a highly respected figure in NLP, and Prof. Bugeun Kim from Chung-Ang University.
I visited World IT SHOW 2025! I visited with my advisor, Professor YOUNG GEUN KIM, and CASL member Kyudong Kim. By chance, I met my high school friend SeongRyong Oh, who is now a Ph.D. student at KAIST, and had an interesting conversation about On-device AI.
As graduation season arrives, I want to take a moment to celebrate and congratulate the amazing people around me who are closing one chapter and stepping into the next.
To the incredible researchers at SMRL, where I had the chance to work as an undergraduate researcher—especially Seunghoon Choi, my mentor, and Hoseok Kim, thank you for your guidance, patience, and for being such inspiring role models. I’m truly grateful for the time I spent learning from you.
To my fellow undergraduate researchers at CASL and to the friends I’ve shared the Computer Science journey with since 2019—it’s been a long, challenging, and memorable journey. From all-nighters to last-minute submissions, to breakthroughs in labs and the laughter between classes, we’ve come so far together.
To the amazing Class of '19—Jiweon Kang, Suhyeok Kim, Jaeseo Lee, Sangwook Lee, Seongjin Lee, Hanyoung Yoo, Hongbin Park, Chihun Choi, Jinwoo Choi, Jonghyeok Han, Wonjun Han, Guijung Woo, Wooseong Jeong —thank you for the memories, the support, and the shared growth. You’ve all made this journey special in your own ways. And to those whose names I couldn’t list here, please know you are just as appreciated and celebrated.
Congratulations again to all of you. I hope the future holds everything you’ve worked so hard for—whether that’s in academia, industry, or a path uniquely your own.
Let’s keep in touch and keep cheering each other on. The best is yet to come.
I visited AI SEOUL 2025 Conference - AI for a Harmonious Society .
I attended the Korea University ICT Creative Consilience Program 2024 AI Tech Day. I participated in the poster session as a CASL member.
I attended the Korea University ICT Creative Consilience Program 2023 AI Tech Day. I participated in the poster session as a CASL member.
I visited SK TECH SUMMIT 2023. I attended the first unveiling event for SAPEON's Energy Efficient AI semiconductor and Next Generation Chip, X330.
I visited Samsung AI Forum 2023 (SAIF 2023). There was a keynote session by Professor Yoshua Bengio, a leading scholar in the AI field, on the topic 'Towards a safe AI scientist system'. I was also able to hear insights from speakers from Meta, Microsoft, and Samsung Advanced Institute of Technology.
I was invited to visit LG Sciencepark (LG R&D center), as an industry-academia scholarship student. LG Electronics' future technology strategy introduction and LG Innovation Gallery Tour were held.
I graduated from Korea University with Great Honor and achieved Early Graduation!
I was selected to the Dean's List and received an Academic Excellence Award!
I completed 11 months of mandatory military service!
I was admitted to the Army war wound/accident on duty examination by the Korean Army Headquarters.
The Ministry of Patriots and Veterans Affairs (MPVA) recognized me as a disabled veteran! I receive medical support in accordance with the FRAMEWORK ACT ON VETERANS AFFAIRS. [Army headquarters decision] [MPVA decision]
I got a drone as a gift — it’s now my second greatest treasure, right after my beloved science kit. A new adventure begins, this time from the sky!