Seonglae Choas AI develops more advanced capabilities like large context windows and fast weights, it may foster human-like interactions, potentially reducing regulatory resistance among progressive groups.
Tell us about your most significant technical accomplishment.
This might be a project you built individually or as part of a team, your performance in competitive programming or academic competitions, contributions to research work, impactful experiences during internships, or any other technical achievement you consider particularly meaningful. If you have relevant work on GitHub or similar platforms, please feel free to include links.
One project I am particularly proud of is my research on RTSum, which I developed as part of my internship at Yonsei University’s Data & Language Intelligence Lab. The project focused on validating an Interpretable AI framework. This work led to a publication, and it was accepted at NAACL 2024, where I was the main developer. https://github.com/seonglae/RTSum
Recently, I won the first prize (£3,000) at the UCL x Holistic AI 2024 competition. Using a Sparse AutoEncoder(SAE) and a stereotyped dataset, I manipulated GPT-2 via steering vectors to prevent generating stereotyped text. Additionally, I introduced a novel method to extract interpretable features without relying on LLMs as explainers, utilizing Point-biserial correlation. https://github.com/seonglae/emgsd-hermes
Another project that shows my engineering ability is MBTI GPT, an AI personality analyzer that gained over 1,000 users. I implemented it using Redis, OpenAI API, and Faiss, optimizing prompts and reducing API costs by 30%. This service highlights my ability to develop real-world AI implementation. https://mbti.texonom.com
Additionally, I’ve written about Mechanistic Interpretability and LLM steering. In my article on the Superposition Hypothesis for steering LLM with Sparse AutoEncoder, I introduced and explained Anthropic's research in a more accessible way. https://seongland.medium.com/superposition-hypothesis-for-steering-llm-with-sparse-autoencoder-c07b74d23e96
Another work titled Reversing Transformer to understand In-context Learning examines phase change and feature dimensionality within transformers. https://seongland.medium.com/reversing-transformer-to-understand-in-context-learning-with-phase-change-feature-dimensionality-13cbf8a2f984
These writings reflect my commitment to understanding and advancing the interpretability and control of large language models.
When predicting the future, I prefer technical analysis over high-level speculation for greater accuracy. Changes typically occur continuously, driven by MVPs and market dynamics. In this scenario, where AI intellectual labor becomes cheaper than human labor, intellectual tasks would be replaced far faster than physical tasks, surpassing current robotics advancements driven by reinforcement learning. This shift would raise the value of physical labor, with simpler tasks being automated first, leaving only highly complex physical tasks in demand.
Over the next five years, AI regulation will be the primary bottleneck. The speed of labor replacement, starting with driving and expanding into other areas, will heavily depend on regulatory decisions. Job losses in labor-intensive sectors could spark political conflicts between conservative and progressive ideologies.
As AI progresses with large context windows, fast weights, and reinforcement learning, it could enable more autonomous AI agents and extensive AI networks. These developments may lead to risks such as internal AI conflicts, systemic failures, or malicious misuse through mechanistic interpretability-steered features. Such incidents could result in stricter AI regulations, reduced open-source AI, and higher taxation on large-scale AI training.
What are your long-term career aspirations? (~100-150 words)
We ask this to better understand your interests and motivations. We know that career plans often evolve and might be broad at this point. Your response won't be used to evaluate your application's success.
How will MARS help you to achieve your long-term career aspirations? What are you hoping to get out of the program (~100-150 words)
Perplexity inference projects
One project I am particularly proud of is improving large-scale inference performance in Open-Domain Question Answering (ODQA). I boosted LLM evaluation efficiency by 40% by deploying a multi-GPU local inference server with Huggingface TGI and asynchronous batch processing. Additionally, I enhanced ODQA accuracy by 20% through zero-shot LLM context manipulation and completed large-scale QA benchmarks by indexing 21 million Wikipedia passages into a Milvus vector database within 12 hours. https://github.com/seonglae/ReSRer
Another key achievement is my research on RTSum, an interpretable AI framework developed during my internship at Yonsei University’s Data & Language Intelligence Lab. This work led to a publication, accepted at NAACL 2024, where I was the main developer. https://github.com/seonglae/RTSum
Recently, I won first prize (£3,000) at the UCL x Holistic AI 2024 competition by steering GPT-2 during inference using Sparse AutoEncoders to mitigate stereotyping. I also introduced a novel method for extracting interpretable features via Point-biserial correlation, which shows potential for applications in LLM advertisement or intention injection. https://github.com/seonglae/emgsd-hermes
Another project that demonstrates my engineering expertise is MBTI GPT, an AI-powered personality analyzer with over 1,000 users. I built it using Redis, OpenAI API, and Faiss, optimizing prompts and reducing inference API costs by 30%. This project showcases my ability to deliver practical and scalable AI solutions. https://mbti.texonom.com