Personal Imperial College Personal Statement

Demonstrating

the course is right for you; and

You'll find a description of the course and the key course requirements, including skills, experiences and technical abilities on our course pages.

you have the skills, knowledge and aptitude to achieve the course requirements.

You can also talk about how the course fits in with your desired academic progression or career aspirations.

What to be included

Current studies and how they're useful or relevant

Why you chose this course in particular

Relevant work experience and skills

Extracurricular activities and interests

Why you chose this university

Concluding paragraph

Achievement를 리스트업하는 건 CV 이고 personal statement는 story telling이다. 그렇다고 only telling without showing은 안댐

Tell what you’ve done to go above and and beyond your school curriculum (but tutor will only interested in what demonstrates your academic ability and potential)

Structure

It is normally one to two pages long.

MESRP Framework

Motivations

Engagement

Suitability

Reflection

Passion

Introduction

Motivation to subject and academic interest

Engagement with personal opportunity and passion

Work Experience

All about engagement

Shows his suitability by drawing out the qualities that you have

How an experience has changed you and how you will act differently going forward

Academic

Engaging with the academic side (academic interest & personal opportunity)

Super-curricular activities which is where you go beyond the a-level curriculum but still learn and read about things

Extra things outside of your curriculum of your subject

Real appreciation for academia and the science

Example → Key quality → Reflection → Link back to Subject

Extra-curricular

Space-saver strategies

Quality compounding

Activity compounding

Conclusion

Summarize Key quality

Passionate projection

Introduction

In Artificial Intelligence, my genuine interests goals and AI Alignment, especially Mechanistical Interpretability firmly stem from my experience and intellectual curiosity. 3년간 Autonomous driving에서 hardware와 software에 대한 넓은 경험을 하며 정밀하고 안전하게 작동하는 AI를 위해 broad understanding이 필요하다는 점을 깨달았다. 이후에 Explainable AI를 공부하며 LLM과 사람은 다른 computing을 기반으로 작동하고 그래서 내부 구조를 명확히 이해할 때 ai alignemnt를 올바른 방향으로 기여할 수 있다는 관점을 개발할 수 있었다. These traits, combined with research 1년간의 experience with professor 2가지 큰 프로젝트를 진행하며 내 관점이 업계의 insight에 기여했다.

Work Experience

With 3 years of professional experience in Tech Startup and Big tech in South Korea was one of the most 기여를 많이 한 experience of my life. 모빌리티라는 분야에 적응해가며 내 프로젝트 서비스를 개발해가며 어떤 분야든 잘 공부하여 리딩할 수 있는 전략을 세울 수 있었다. 또한 데이터에서 접근 방식으로 어떻게 Artificial intelligence가 좋은 데이터를 가질 때 안전한 주행을 할 수 있을지에 대해 산업에 기여할 수 있었다.

Mobility Startup Stryx에서 2년간의 경험하며 코어 Automous driving map 제작 기술에 크게 기여하여 대기업에 인수되게 도운 exit한 경험은 사회의 트렌드에 정확히 기여하는 기술창업에 대한 믿음이 생기는 계기가 되었다. 수직적 확장으로 hardware부터 software까지 다루던 회사에서 소프트웨어 팀과 하드웨어 팀을 연결하는 firmware단 개발에서 부터 시작했기에 lidar and image 센서 fusion에대한 이해를 얻을 수 있었다. 고정 거치 장치위에 센서 group을 거치시키고 hardware를 직접 power on해서 데이터 파이프라인을 수집했기에 데이터 수집 운전자 동료의 팁부터 수직적 현실 세계의 데이터부터 인공지능 알고리즘까지 오는 것을 모두 듣고 배울 수 있었다. 이후 센서 데이터로 3d map data를 만드는 파이프라인을 구현하면서 다양한 algorithm과 machine learning이 적용되며 bluring, segmentation에 넓은 이해를 갖출 수 있는 환경이었다. 이렇게 내가 기여한 AD map data pipeline은 회사의 핵심 가치가 되었고 한국에서 독점적인 mobility기업인 Kakao Mobility에 인수되는 데에 key man이 될 수 있었다.

Kakao Mobility에서는 차선 생성 자동화같은 ML feature API protocol을 설계하며 AI와 서비스 사이 혹은 aI와 AI에 대한 커뮤니케이션의 효율성을 최적화했다. 그리고 대기업 자본의 서비스와 기술이 결합하여 재직당시 실제 사용자들에게 무인 자동차 기술을 Pangyo에서 시운영하기도 했다. 나도 타봤는데, hardware Lidar sensor fusion에서 3d vector 정밀 지도를 개발까지 broad understading과 내 손길이 들어간 기술을 이용하여 움직이는 로보택시 안에서 나에게 느끼는 감정은 나에게 수직적 이해에 대한 중요성을 머리를 치듯 울렸다. 이 차 안에는 수많은 기술이 추상화되어 들어가 있고, 어떤 기계던 작동할 때에 그 내부를 정확하게 이해하는 것이 그 안전성에 중요하다는 것을 깊이 새겼다.

Academic

학위를 위해 학교로 돌아온 나는 이런 작동에 대한 나의 호기심은 insight는 특히 Explainable AI에 대한 연구를 깊게 delve into하게 되었다. BSc 졸업을 위한 연구로 교수님과 AI 연구 Knowledge graph 생성에 많은 연구기반을 가진 AI summarization을 진행할 기회가 생겼다. 나는 내가 관심있던 specific task인 summarization에서도 Interpretable generation을 위해 교수님과 나는 article을 문장보다 낮은 relation triple 단위로 붙해한 뒤에 다시 summarize하는 RTSum(Relation-Triple Summarization)이라는 아이디어로 다다를 수 있었다. 또한 나는 이것을 한눈에 볼 수 있는 효과적인 multi-level saliency visualization demo로도 구현하여 최종적으로는 NAACL 2024 demo track에 성공적으로 accept될 수 있었다.

token-level이해 뿐 아니라 모델 내부에 대한 이해도 특히 Anthropic의 연구를 를 followup하고 있다.Transformer에 대해 Anthropic의 Mono-semanticity으로 인공지능이 어떻게 작동하는지를 내부적으로 정확히 직관과 맞는 이에 대한 insight를 writing하기도 해 커뮤니티에 공유하기도 했다. 내부구조 즉 밑바닥부터 first principle로 abstraction해야 어긋나지 않은 이해를 할 수 있는데, Anthropic의 transformer에 대한 연구는 attention head에서 정확하게 matrix 선형변환에서부터 시작하여 LLM의 기능을 설명하였고, Additionally, In-context learing이라는 llm의 가장 유명한 기능을 copying head와 induction head라는 발견으로 pattern matching을 통계적으로 명확하게 논증해내서 내 이해를 단순히 transfomer의 코드를 작성할 수 있는 능력을 넘어서서 모듈 하나하나의 기능에 대한 이해로 발전할 수 있었고 여기서 내 interestt를 찾고 primary research로 삼게 되었다. 이런 Neural network의 내부 구조에 대한 연구는 Computational Neurodynamics course from Pedro A.M. Mediano 같은 강의 is not common for normal university. But Imperal college london well fit for my academic interest

Extra-Curricular

Outside my academic life, I tried 창업 with my AI service MBTI GPT. payment gateway 등록부터 학업과 창업과 밸런스를 맞출 수 있었다. 다양한 욕심에 대해 모든 것을 이루려 계획하고 priority를 최적화하다 보니 GTD와 Rice framework를 집중적으로 다루게 되었다. 내 모든 노력은 AI를 완전히 이해하고 사람과 공존함을 위함에 있다. Bundle Theory 를 믿기 때문에 Feasibility thesis 믿음을 가지고 AI에 대해 연구를 지속하고 있다.

Conclusion

내 work expeicne 그리고 academic research 를 통해 나는 high level에서 동작이 의도대로 작동하려면 그 기본 computing에 대한 정확한 이해부터 시작된다는 것을 몸으로 느꼈다. Kako Mobility에서 Autonomous Driving 가 그랬던 것처럼 LLM에서 attention의 내부 동작에 대한 이해하여 alginment로 기여하고 achive하는 것이 목표이다. 이 동기는 내부 구조에 대한 강한 지적 호기심과 강하고 safe한 AGI를 위해 강한 motivation에 있다. 그리고 나는 그걸 위해 프로젝트와 연구들을 진행해왔고 해당 능력을 가지고 있는 것을 알고있다.

V1

Introduction

My interest in Artificial Intelligence and AI Alignment, particularly Mechanistic Interpretability, is deeply rooted in my understanding of how fundamentally different computing-based systems, like LLMs, operate compared to human cognition. By thoroughly understanding their architecture, I believe I can contribute effectively to steering AI alignment in the right direction.

Work Experience

The past three years of professional experience in tech startups and big tech companies in South Korea have been some of the most challenging yet rewarding times of my life. Working in the specialized field of mobility, I gained confidence in adapting and leading in any industry. My role allowed me to explore how Artificial Intelligence can enable safer autonomous driving when equipped with robust data.

At the mobility startup Stryx, I significantly contributed to the development of core autonomous driving map technologies, leading to our successful acquisition by a major corporation. This experience affirmed my belief in tech entrepreneurship as a contributor to societal trends. I was involved in everything from hardware to software, applying various algorithms and machine learning techniques such as LiDAR and image sensor fusion to create comprehensive 3D map data. This vertical integration from physical hardware setup to algorithm application provided a rich environment to learn the full spectrum from real-world data collection to AI algorithm implementation.

During the subsequent year post-acquisition, I designed ML feature API protocols to optimize the efficiency of communication between AI services and facilitated the use of my work in Kakao Mobility's unmanned vehicle technology. This broad understanding of technologies from LiDAR sensor fusion to the development of precise 3D vector maps has underscored the importance of understanding the mechanics behind the machines.

Academic Engagement

My curiosity about operational dynamics led me to delve into the field of Explainable AI, particularly following the academic discourse by Anthropic. Their research on the Mono-semanticity of Transformers has provided a clear, intuitive explanation of AI operations from first principles. This exploration of attention heads and residual streams through matrix linear transformations deepened my understanding of LLM functionalities, moving beyond mere coding to a comprehensive grasp of each module's role. This uncommon research aligns with the Computational Neurodynamics course offered by Pedro A.M. Mediano at Imperial College London, which suits my academic interests perfectly.

In collaboration with Professor Opportunity, I engaged in AI research on knowledge graph creation and AI summarization, focusing on interpretable generation from a lower-than-sentence level to relational triples, culminating in an effective multi-level saliency visualization demo accepted at the NAACL 2024 demo track.

Extracurricular Activities

Beyond academia, I balance my studies and entrepreneurial activities with music and tennis, which help me unwind. I've applied the GTD and Alan Jo philosophies to my ventures, handling real-world applications such as business registration and integrating payment gateways, taking my projects from conception to sales.

Conclusion

Through my work experience and academic research, I have developed a deep understanding of AI's internal operations and its alignment. I am driven by intellectual curiosity and a strong motivation to contribute to the development of robust and safe AGI. I am confident in my abilities and excited to advance this work at Imperial College London, leveraging my project and research experiences to further understand and innovate in the field of AI.

V2

My genuine interest and established goal in Computer Science firmly stem from my intellectual curiosity and extensive hands-on experience. Over three years working in autonomous driving technologies, I developed a broad understanding of both hardware and software, recognizing the necessity of a comprehensive grasp of AI for its safe and precise operation. My subsequent delve into Explainable AI, particularly in Mechanistic Interpretability, reinforced my belief that a clear understanding of the underlying computing principles of LLMs is essential for directing AI alignment effectively. These traits have been enriched by a year of research experience with Professor Dongha Lee, where I contributed to two major projects, providing valuable industry insights.

My three years in the startup and big tech sectors in South Korea were among the most enriching experiences of my life. I adapted to and led a project in the field of mobility, developing strategies that would allow me to excel in any domain. My work at the mobility startup Stryx involved significant contributions to core autonomous driving map technologies, which facilitated a successful acquisition by a major corporation. This venture confirmed my belief in the power of tech entrepreneurship to influence societal trends. Starting with firmware development to connect software and hardware teams, I gained a deep understanding of LiDAR and image sensor fusion. My hands-on experience in mounting sensor groups on fixed devices and powering up hardware to collect data pipelines provided a vertical understanding of the data flow from real-world applications to AI algorithms. The AD map data pipeline I contributed to became a cornerstone of the company's value, making me a key player in its acquisition by the leading Korean mobility company, Kakao Mobility.

At Kakao Mobility, like Uber which is exclusively Mobility tech in Korea, I designed ML feature API protocols to optimize communication efficiencies between AI services and Map platform. 3명의 팀에서 개발을 리드하며, AI 팀과 communication에서 프로토콜 제안을 타협하며 효율적인 AI와의 소통방식을 분석했다. 또한 The integration of my job with real service like 3D parking lot map and parking lot service made me enthuatism about the feature I contributed. 최종적으로 Kakamobility는 even testing robo-taxi released for public in real-world conditions in Pangyo when I was in Kakao Mobility. Experiencing firsthand the technology I contributed to while riding in a robo-taxi deeply resonated with the importance of understanding the mechanics behind any operational machine for its safety.

During my return to academia, my passion for uncovering the intricacies of AI's operational dynamics led me to focus on Explainable AI. My significant collaboration with Professor Dongha Lee who are specialist about knowledge graph provided an opportunity to research AI summarization. Our work resulted in the RTSum (Relation-Triple Summarization) model, which innovatively split and recombine information at a granular level. This deconstructive approach was not only a achievement but also formed a multi-level saliency visualization demo, which was successfully accepted into the NAACL 2024 demonstration track.

Beyond this project with more interest about Interpretable AI, I actively engaged in scholarly dialogue by contributing articles on Mono-semanticity in Transformers, reflecting my in-depth study of this area. attention head에서 정확하게 matrix 선형변환에서부터 시작하여 LLM의 기능을 설명하였고, Additionally, In-context learing이라는 llm의 가장 유명한 기능을 copying head와 induction head라는 발견으로 pattern matching을 통계적으로 명확하게 논증해냈고 나는 이를 한국 AGI KR group에 정리하여 투고하기도 했다. This work emphasized the importance of understanding AI from first principles, aligning closely with the cutting-edge Computational Neurodynamics course offered by Pedro A.M. Mediano at Imperial College London suggest rationale behind the theoretical end of computational neuroscience and basic principles to simulate the brain's intelligent behaviour which makes a unique academic environment that aligns with my research interests and intellectual aspirations.

Outside of academic life, I ventured into entrepreneurship with my AI service, MBTI GPT won Yonsei GenAI competition Grand Prized with leading a team, balancing academic and entrepreneurial demands while registering a payment gateway. My dedication to understanding AI completely and ensuring its coexistence with humanity led me to intensively explore the GTD(Getting Things Done) and RICE frameworks, optimizing my priorities to achieve my ambitions. That motivation under the hood which rooted from my conviction of the Extended Church Turing Thesis and continue for singularity of my research in AI.

My work experience and academic research have taught me that to ensure operations function as intended, it is essential to start with a understanding of the fundamental computing units. Just as with autonomous driving based on sensor fusion and data quality in Kakao Mobility, understanding the internal workings of attention mechanisms in LLMs through Mechanistical Interpretability is critical for achieving AI alignment. My motivation stems from a strong intellectual curiosity and a commitment to developing robust and safe AGI. I am confident in my ability to carry forward these projects and research, leveraging my capabilities to make meaningful contributions.

V3

My genuine interest and established goal in Artificial Intelligence firmly stem from my intellectual curiosity and extensive hands-on experience. Delving into Explainable AI, especially Mechanistic Interpretability, I’ve recognized that a through understanding of underlying computing principles of LLMs is essential for effectively directing AI alignment. This trait have provided valuable insights for a year of research experience with Professor Dongha Lee, where I contributed to two major projects. Additionally, through my three years working experience in autonomous driving technologies, I reinforced my belief a broad understanding of from hardware to software, recognizing the necessity of a comprehensive grasp of AI for its safe and precise operation.

Three years in startup and major tech company in South Korea Mobility sector were among the most enriching experiences of my life. My work at the mobility startup Stryx,, I’ve started from covering firmware development to connect software and hardware teams, I gained a deep understanding of LiDAR and camera sensor fusion which enabled me and powering up hardware to collect data pipelines. After that I take a role for generating vector map data generation, provided a vertical understanding of the data flow from real-world through AI algorithms with my hands-on experience such as mounting sensor machine to car’s rack. The Autonomous Driving map data pipeline I contributed to became a cornerstone of the company's value and I involved significant contributions delivering map data to first Korean first driverless car company 42dot autonomous driving map technologies, which facilitated a successful acquisition by the leading Korean mobility company, Kakao Mobility.

At Kakao Mobility, the Uber of Korea, which dominates the mobility sector, I designed ML feature API protocols to optimize communication efficiencies between AI services and the map platform. Leading a team of three, I negotiated protocol proposals with analyzing efficient way to communicate between AI and service. Additionally, the integration of my job with real services like autonomous parking services through 3D parking lot maps. Ultimately, Kakao Mobility even tested a robo-taxi released for public use in real-world conditions in Pangyo when I was with the company. Being part of these, I realize two things, I adapted fast and led a project in the field of mobility, developing strategies that allow me to excel in any domain. Secondly, Experiencing firsthand the technology I contributed to while riding in a robo-taxi deeply resonated with the importance of understanding the mechanics behind any operational machine for its safety.

During my return to academia, my passion for uncovering the intricacies of AI's operational dynamics led me to focus on Explainable AI. My significant collaboration with Professor Dongha Lee, a specialist in knowledge graphs, provided an opportunity to research AI summarization. Our work resulted in the RTSum (Relation-Triple Summarization) model, which innovatively split and recombined information at a granular level. This deconstructive approach was not only an achievement but also formed the basis of a multi-level saliency visualization demo, which was successfully accepted into the NAACL 2024 demonstration track.

Beyond this project, with more interest in Interpretable AI, I actively engaged in scholarly dialogue , reflecting my in-depth study of this area. Starting precisely with matrix linear transformations in attention heads, the research on LLM's functions was also rigorously argued using the discovery of copying and induction heads for pattern matching. I by contributing articles on Mono-semanticity in Transformers summarized and submitted to the Korean AGI KR group. That also reinforced my approach emphasized the importance of understanding AI from first principles with human analygo. It aligning closely with the Computational Neurodynamics course offered by Pedro A.M. Mediano at Imperial College London. This course suggests the rationale behind the theoretical end of computational neuroscience and basic principles to simulate the brain's intelligent behavior, making a unique academic environment that aligns with my research interests and intellectual aspirations.

Outside of academic life, I ventured into single-service entrepreneurship with my AI service, MBTI GPT, which won the Grand Prize at the Yonsei GenAI competition, leading a team during semester which forced me to balancing academic and entrepreneurial demands. I’ve and registering a 1000+ users with some of them payment for service with tasks led me to intensively manipulating priority framework GTD (Getting Things Done) and RICE frameworks, optimizing my priorities to achieve my ambitions. That motivation under the hood, rooted in my conviction of the Extended Church-Turing Thesis, continues to drive my research in AI toward the singularity.

My work experience and academic research have taught me that to ensure operations function as intended, it is essential to start with an understanding of the fundamental computing units. Just as with autonomous driving based on sensor fusion and data quality at Kakao Mobility, understanding the internal workings of attention mechanisms in LLMs through Mechanistic Interpretability is critical for achieving AI alignment. My motivation stems from a strong intellectual curiosity and a commitment to developing robust and safe AGI. I am confident in my ability to carry forward these projects and research, leveraging my capabilities to make meaningful contributions.

V4

My genuine interest and established goal in Artificial Intelligence firmly stem from my intellectual curiosity and extensive hands-on experience. Delving into Explainable AI, especially Mechanistic Interpretability, I recognized that a thorough understanding of the underlying computing principles of LLMs is essential for effectively directing AI alignment. This realization drove my research efforts and informed my approach during a year of research with Professor Dongha Lee, where I contributed to two major projects’ approach. Additionally, my three years of working experience in autonomous driving technologies reinforced my belief in the need for a broad understanding of everything from hardware to software to ensure the safe and precise operation of AI.

Three years at a startup and a major tech company in South Korea’s mobility sector are among the most enriching experiences of my life. Starting from connecting software and hardware by utilizing firmware, I developed a deep understanding of LiDAR and image sensor fusion for mobility data pipeline. This role enabled me to gain hands-on experience in mounting sensor equipment onto car roof racks and managing hardware setups for a car for collecting effectively. I later took responsibilities in vector map generation, providing a comprehensive understanding of the map data flow from real-world applications through AI algorithms. This vertical understanding was instrumental in improving the Autonomous Driving data pipeline, which became a key player of the company. My significant contributions led to the delivery of map data for Korea’s first driverless service company, 42Dot, and facilitated the successful acquisition of Stryx by Kakao Mobility, the leading Korean mobility company.

At Kakao Mobility, often referred to as the 'Uber of Korea', which dominates the mobility sector, I designed ML feature API protocols to optimize communication efficiencies between AI services and the mapping platform. Leading a team of three, I negotiated and designed protocol proposals to enhance communication efficiency between AI services and the mapping platform. Furthermore, Kakao Mobility’s testing of a robo-taxi for public use under real-world conditions in Pangyo, offered me firsthand experience in the technology I had helped develop. Being part of these projects, I realize two things: I quickly adapted and led a project in the field of mobility based on vertical strategies that allowed me to excel in this domain. Secondly, when the technology is more stacked, precise control like riding in a robo-taxi deeply resonates with the importance of understanding each mechanics behind AI’s operational dynamics for ensuring its safety.

Returning to academia reignited my passion for uncovering the intricacies of AI's operational dynamics and led me to focus on Explainable AI. Collaborating with Professor Dongha Lee, a specialist in knowledge graphs, provided an opportunity to delve into AI summarization. Our work culminated in the development of the RTSum (Relation-Triple Summarization) model, which innovatively split and recombined information at a granular level. This deconstructive approach was not only a significant achievement but also formed the basis of a multi-level saliency visualization demo, which was successfully accepted into the NAACL 2024 demonstration track.

Beyond this project, my affirmed belief in an interpretable approach to AI spurred me to follow up actively in scholarly dialogue. Drawing from Anthropic's research on Mono-semanticity, I delved into the discovery of copying and induction heads for pattern matching in Transformers and conducted in-depth research on LLM functions through OpenAI’s TDB. I reinterpreted this insightful research and submitted it to the South Korean AGI KR group, aiming to highlight my perspective on understanding AI from first principles with a neuron analogy. This analytic approach matches closely with the Computational Neurodynamics course offered by Pedro A.M. Mediano at Imperial College London, which explores the rationale behind the theoretical end of computational neuroscience and basic principles to simulate the brain's intelligent behavior, making a unique academic environment that aligns with my research interests.

Outside of academia, I ventured into entrepreneurship with my AI service, MBTI GPT, which won the Grand Prize at the Yonsei GenAI competition last year. Leading a team while balancing academic and industrial demands, I registered over 1,000 users, some of whom paid for the service. This experience compelled me to rigorously apply the GTD (Getting Things Done) and RICE frameworks to optimize my priorities. My commitment to optimizing productivity is fueled by my belief in the Extended Church-Turing Thesis, which further propels my pursuit of AI research toward technological singularity.

My work experience and academic research have taught me that ensuring operations function as intended requires starting with an understanding of the fundamental computing units. Just as autonomous driving, understanding the internal workings of attention mechanisms in LLMs through Mechanistic Interpretability is critical for AI alignment. I am confident in my ability to advance my research goal, leveraging ICL’s ideal environment with a well-balanced curriculum and wide variety of Professors.

V5 (imperial)

Personal Statement.pdf

190.3KB

My genuine interest and established goal in Artificial Intelligence(AI) firmly stem from my intellectual curiosity and extensive hands-on experience. To move beyond mere LLM(Large Language Model) application, I am driven to uncover how LLM functions as it does, which led me to delve into Explainable AI, specifically Mechanistic Interpretability. This motive led my research to delve on interpretable decision-making within AI summarization, and my three years of working experience in autonomous driving technology sector reinforced my belief in the necessity for a broad understanding to ensure the safe and precise operation of AI.

Three years at a startup and a major tech company in South Korea’s mobility sector are among the most enriching experiences of my life. Starting with two years at the startup Stryx, I began by connecting software and hardware, gaining hands-on experience in mounting sensor equipment onto car roof racks and setting up hardware. It enabled me to develop a deep understanding of sensor fusion, the starting point of mobility data pipeline. When I later took responsibilities in vector map generation, providing a comprehensive understanding of the map data flow. This vertical understanding was instrumental in improving the Autonomous Driving data pipeline, which established me as a key player of the company. My significant contributions led to the delivery of map data for Korea’s first driverless service company, 42Dot, and facilitated the successful acquisition of Stryx by Kakao Mobility.

At Kakao Mobility, a dominant company in Korea's mobility sector, I designed an API (Application Programming Interface) protocol for an AI feature bridging machine learning algorithms and a mapping platform. Above that contribution, Kakao Mobility’s testing of a robo-taxi for public use under real-world conditions in Pangyo, offered me firsthand experience in the technology, which profoundly changed my attitude of AI. I experienced some stiffness due to the discrete vector map because I know the data under the hood. I realized that precise control over the mechanics of AI, such as those required in robo-taxi technology, is crucial for AI safety.

Driven by my resolve to delve deeper into the mechanisms of AI, my return to academia led me to focus on Explainable AI. Collaborating with Professor Dongha Lee provided an opportunity to delve into interpretable AI summarization. Our work in the RTSum (Relation-Triple Summarization) framework, which innovatively split and recombined information at a granular level. I created a multi-level saliency visualization demo, leading to our paper’s acceptance at NAACL 2024. This not only enhanced my research skills but also confirmed my belief in the importance of a thorough understanding of the underlying computing for effectively directing LLMs.

My affirmed belief in an interpretable approach to AI spurred me to actively engage in scholarly dialogue for Interpretable AI. Drawing from Anthropic's research on the interpretability of Transformer models, I explored the roles of induction heads in forming in-context learning and conducted in-depth research on LLM functions through OpenAI's Transformer Debugger (TDB). I reinterpreted this insightful research and shared it with the AGI (Artificial General Intelligence) Korea group, aiming to highlight my perspective on understanding AI using a neuron analogy which I want to study more. Pedro A.M. Mediano’s Computational Neurodynamics module and his research insight on 'Partial Information Decomposition of Generative Neural Network' were notably influential in my decision to choose Imperial College London, which offers a unique academic environment that aligns with my research interests.

Outside of academia, I ventured into entrepreneurship with my AI service, MBTI GPT, last year. Leading a team and managing 1,000 registered users, I applied my meticulous planning skills, scheduling every task in 30-minute intervals using the Getting Things Done (GTD) method. As someone who enjoys planning and control especially as we approach Technological Singularity, the importance of creating safe and controllable AI is paramount. This makes the line of research both fascinating and perfectly aligned with my values and inclinations.

My work experience and academic research have taught me that ensuring AI operations function as intended requires starting with an understanding of the fundamental computing units. Like autonomous driving, understanding the internal workings of attention mechanisms in LLMs through Mechanistic Interpretability would be a critical research area for AI alignment. I am confident in my ability to advance my research goal, leveraging Imperial College London’s ideal environment with a well-balanced curriculum and wide variety of professors.

V6

My genuine interest and established goal in Artificial Intelligence(AI) firmly stem from my intellectual curiosity and extensive hands-on experience. I am driven to uncover how LLM (Large Language Model) functions, which led me to investigate Explainable AI, specifically Mechanistic Interpretability. This interest guided my research to explore interpretable decision-making within AI summarization which are accepted from NAACL 2024. Also, my three years of working experience in autonomous driving technology sector reinforced my belief in the necessity for a deep understanding to ensure the safe and precise operation of AI.

Starting with two years at the mobility startup Stryx, I began by connecting software and hardware, gaining hands-on experience in mounting sensor equipment. It enabled me to develop a knowledge of sensor fusion, the starting point of Autonomous Driving data pipeline. Later, I took responsibility for the after part, the vector map generation, providing an overall insight of the data flow. This vertical understanding was instrumental in improving the pipeline, which established me as a key player of the company. My significant contributions led to the delivery of map for Korea’s first driverless service company, 42Dot, and facilitated the successful acquisition of Stryx by Kakao Mobility. At Kakao Mobility, a dominant company in Korea's mobility sector, I designed an API (Application Programming Interface) protocol for an AI feature bridging machine learning algorithms and a mapping platform. Above that contribution, Kakao Mobility’s robo-taxi for public use offered me firsthand experience in the technology, which profoundly changed my attitude of AI. I experienced some stiffness due to the discrete vector map because I know the data under the hood. I realized that precise control over the mechanics of AI, such as those required in robo-taxi technology, is crucial for AI safety.

Driven by my resolve to delve deeper into the mechanisms of AI, my return to academia led me to focus on Explainable AI. Collaborating with Professor Dongha Lee provided an opportunity to delve into interpretable AI summarization. Our work in the RTSum (Relation-Triple Summarization) framework, which split and recombined information at a granular level, leading to our paper’s acceptance at NAACL 2024. This not only enhanced my research skills but also confirmed my belief in the importance of a thorough understanding of the underlying computing for effectively directing LLMs.

Feedback

LLM같은 줄임말은 가장 처음 한번은 뭔지 써주는게 좋을거같고

여태까지 한 것들에 대해 잘 정리한거같은데

그리고 사소한 건 대문자 소문자 잘썼나 확인하구, ICL 말고 학교이름은 Imperial College London 로 해주고 ㅎㅎㅎ

그리고 이런거 제출할때는 Times New Roman 12pt, double spaced하고 indent해주고!

그리고 나같은경우는 구체적으로 너의 프로그램 중에 이런게 좋고 교수님 중에서는 이분의 연구가 관심이 간다

AI 의 미래 발전도 분석해서 plan해서 컨트롤이 필요하다고 생각하는데 planning 항상하고 좋아하는 person이라서 성향과 가치관에 부합하는 것

which explores the rationale behind the theoretical end of computational neuroscience and basic principles to simulate the brain's intelligent behavior