Sunghee Yun (He/Him)

Co-Founder & CTO @ Erudio Bio, Inc., Co-Founder & CEO @ Erudio Bio Korea, Inc., Leader of Silicon Valley Privacy Preserving AI Forum (K-PAI), Director of AI Semiconductor @ K-BioX, AI-Korean Medicine Integration Initiative Task Force Member @ The Association of Korean Medicine, KFAS-Salzburg Global Leadership Initiative Fellow @ Salzburg Global Seminar, Visiting Professor @ Sogang University, Advisory Professor @ DGIST, Global Advisory Board Member @ Innovative Future Brain-Inspired Intelligence System Semiconductor of Sogang University, Network Expert @ Gerson Lehrman Group, Chief Business Development Officer @ WeStory.ai, and Advisor @ CryptoLab, Inc.

Abstract

This lecture provides a comprehensive exploration of Large Language Models (LLMs) and Generative AI, tracing the evolution from early language models to today’s transformer-based architectures. Beginning with the historical foundations—from bag-of-words models (1954) through word embeddings, RNNs, and LSTMs—the presentation culminates in the revolutionary Transformer architecture introduced by Vaswani et al. in 2017. The lecture demystifies the “secret sauce” behind modern LLMs: the attention mechanism, with detailed mathematical explanations of scaled dot-product attention, multi-head attention, self-attention, and encoder-decoder attention. Through visualization examples, attendees gain intuitive understanding of how these mechanisms enable models to evaluate dependencies between arbitrarily distant words while maintaining parallelizability, fundamentally solving the memory limitations that plagued earlier recurrent architectures.

The generative AI component examines the broader landscape of content generation systems, from early models like Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) to modern transformer-based approaches. The lecture provides mathematical foundations for understanding these models while contextualizing them within the rapid evolution from the deep learning revolution of the 2010s through today’s multimodal AI systems. Real-world applications span text generation (Claude, ChatGPT, Gemini, Grok), image creation (DALL-E, Midjourney, Stable Diffusion), and emerging co-pilot products that enhance human productivity. The presentation addresses both the transformative capabilities and critical challenges facing LLMs, including hallucination, bias, resource requirements, and ethical considerations, while emphasizing the “unreasonable effectiveness of data” that has shocked the AI community.

The lecture concludes with comprehensive market analysis revealing the explosive growth and economic impact of AI technologies. Data shows that AI could add $15.7 trillion to the global economy by 2030, with the AI stack attracting $29 billion in funding dominated by model companies like OpenAI and Anthropic (60% of total). While user adoption may be approaching initial saturation, enterprise adoption remains below 10% despite 60%+ planning implementation, indicating vast growth potential. Performance metrics demonstrate game-changing value delivery: GitHub Copilot saves developers 55% of their time, AI support reduces response times by 44 minutes while achieving 14% higher customer satisfaction, and AI capabilities across domains from image recognition to code generation continue their steep upward trajectory toward human parity, with AI advancing at unprecedented speed—projected to progress from basic content generation to superhuman reasoning in just 5 years compared to 15 years for autonomous vehicles.