Seonglae ChoEngineer
In my recent AI Research Engineer role at Holistic AI, I implemented evaluation pipelines for AI Agents and AI workflow APIs with Python backend frameworks.
At Stryx and Kakao Mobility, I led and contributed to the development of scalable, production-level services, including a national-scale 3D mapping pipeline and data platform that supported autonomous driving applications, where I was responsible for API and backend system design.
As the team lead for MBTI-GPT, which served over 1,000 users, I architected and deployed a full-stack SaaS product that leverages Retrieval-Augmented Generation (RAG) and vector database.
Overall, I have substantial hands-on experience with Python for both production and research settings.
Scientist
two polar competing accelrrating each other develop
I believe there are two poles in AI development: Acceleratists, who believe that intelligence grows exponentially and aim to expedite this growth within compute constraints, and Alignmentists, who focus on guiding the growth of intelligence to ensure AI is robust and interpretable. These two groups have competed throughout AI history, with conflicts dating back further than many realize, particularly within communities like LessWrong.
I used to identify as an Acceleratist before encountering LessWrong and Anthropic’s circuit threads about two years ago. Their writings on circuit analysis and mechanistic interpretability deeply inspired me and encouraged my transition to Alignmentist perspectives. These insights primarily emerged from London-based communities dedicated to AI safety and alignment, which is a major reason I chose London for my MSc, to actively join this community.
Since then, I have fully committed my efforts to mechanistic interpretability, especially Sparse Autoencoders (SAE). Eventually, my research, FaithfulSAE, was accepted to ACL 2025 SRW, and another project, RTSum, utilizing higher-level interpretability methods, was demonstrated at NAACL 2024. Currently, I'm working on my thesis about AI circuit discovery methods that integrate reinforcement learning with control models.
I am aware that scientists at Cohere have deep connections with the London AI safety academic community. Cohere’s research such as "Procedural Knowledge in Pretraining Drives LLM Reasoning" and "Reverse Engineering Human Preferences with Reinforcement Learning" has inspired my approach to understanding AI. Their work has also fueled my ideas on improving model architectures. This aligns closely with my view that these two poles are not only reciprocal but complementary, historically driving the development of AI forward. That’s why I’m eager to contribute and continue learning alongside the team at Cohere