# AI Agent 시대: 인간과 기술이 공존하는 미래 l 윤경아, KT Agentic AI 랩장 | Kyung-A Yoon | TEDxSeoul Source: https://www.youtube.com/watch?v=bxs9tuAMs3E Recap page: https://rapidrecap.app/video/bxs9tuAMs3E Generated: 2025-11-10T16:42:07.9+00:00 --- ## Quick Overview The future of AI development is moving beyond simple generative models toward complex AI Agents that can reason, plan, and utilize tools to achieve user goals autonomously, as illustrated by the comparison between historical industrial revolution challenges and the current AI paradigm shift. **Key Points:** - Generative AI is evolving from simple pre-training models to sophisticated AI Agents capable of complex reasoning and planning. - The speaker, Kyung-A Yoon from KT AI Research Lab, discusses the advancement of AI Agents, which incorporate memory, planning, and tooling capabilities. - The evolution of AI mirrors historical technological shifts (like the Industrial Revolution), but the AI paradigm change is occurring much faster, demanding proactive preparation. - Early AI agents were rule-based (like Microsoft's Clippy), limited by explicit coding; current agents are driven by LLMs (Large Language Models). - AI Agent development includes two main stages: Pre-Training (on vast, general data) and Fine-Tuning (on specific, task-oriented data like legal precedents). - The speaker highlighted the necessity of preparing for AI to potentially replace cognitive labor, contrasting the Industrial Revolution's impact on manual labor. - The key takeaway is to "Don't ignore AI. Watch it closely," emphasizing proactive engagement with the technology's rapid advancement. ![Screenshot at 07:19: The diagram illustrating the core architecture of an AI Agent, showing the interaction between Memory, Prompt processing, Action execution, Tooling, and the foundational Generative AI Model \(LLM, sLM\), provides a visual overview of the complex, iterative process these agents use to fulfill user objectives.](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-07-19.png) **Context:** This presentation, delivered by Kyung-A Yoon, Head of KT's Agentic AI Lab, at TEDxSeoul 2025: CROSSOVER, focuses on the current state and future direction of Artificial Intelligence, specifically the development and implications of AI Agents. The speaker grounds the discussion by contrasting the perceived threats and opportunities of AI with those seen during the Industrial Revolution, emphasizing the accelerated pace of change driven by Large Language Models (LLMs). ## Detailed Analysis The presentation details the maturation of AI from static generative models to dynamic AI Agents, which are defined as 'intelligent entities that understand and act upon goals' (07:21). The speaker contrasts the historical context of the Industrial Revolution (18th-19th centuries) with the current AI Revolution (21st century), noting that while the former primarily automated manual labor and created new industries, the AI revolution threatens cognitive labor, demanding proactive preparation (15:44-15:48). The evolution of AI assistance is traced from early, rule-based systems like Microsoft's Clippy (1997-2003) to modern, LLM-powered agents (1:19-2:01). The core AI development process involves Pre-Training on massive, general data (text, image, sound) from the Random Model stage, followed by Fine-Tuning (or RLAIF/RLHF) for specific tasks like summarization or legal document creation (04:44-05:25). The resulting AI Agent architecture comprises key loops: Memory interacts with Prompt interpretation, which feeds into Task Planning/Reasoning, leading to Action execution via Tooling, which connects to external tools like Web Search (07:37-07:57). The speaker concludes with a quote from Ilya Sutskever: 'Don't ignore AI. Watch it closely' (17:57), urging the audience to engage actively rather than fear replacement, as AI is poised to augment human capabilities, especially in complex cognitive tasks. ### Generative AI Evolution - From movie depictions like 'Her' (2014) showing early human-AI relationships to the explosive adoption of ChatGPT (01:25-01:36) - The trend moves from simple GenAI to complex, multi-tool-using AI Agents (07:22-07:24). ### AI Agent Architecture - Core components include Memory, Prompt Interpretation, Action execution, Tooling, and the foundational Generative AI Model (LLM, sLM) (07:37-07:57) - Agents use tools (RAG, Web Search, 3rd Party Tools) to execute plans derived from user commands (07:57-08:00). ### Two-Stage AI Development - Pre-Training involves massive data ingestion from diverse sources (text, image, sound) starting from a Random Model (04:44-05:10) - Fine-Tuning focuses on specific domains like legal work (05:21-05:37), producing specialized models like LLM/sLM. ### Societal Impact Comparison - Industrial Revolution (18th-19th C) automated manual labor, creating new jobs (15:38-15:41) - AI Revolution (21st C) automates cognitive tasks, demanding new skills and preparation, or risking widespread job displacement (15:44-15:48). ### Agentic Patterns - AI Agents are developing complex interaction structures beyond single agents, including Network, Supervisor, Hierarchical, and Custom patterns for collaborative problem-solving (10:39-10:43) - The goal is to enable agents to cooperate to solve complex problems, moving beyond simple Python scripting examples (12:25-12:40). ![Screenshot at 00:00: The opening slide for TEDxSeoul 2025: CROSSOVER, setting the context for the talk.](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-00-00.png) ![Screenshot at 01:19: Slide illustrating the rise of Generative AI, featuring the film 'Her' \(2014\), the ChatGPT logo, and a chart showing rapid user growth \(01:35\).](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-01-19.png) ![Screenshot at 02:55: Diagram contrasting older AI helpers \(like Microsoft's Clippy\) with modern concepts like Siri and the emerging AI Agent structure \(03:00-03:07\).](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-02-55.png) ![Screenshot at 04:44: Diagram outlining the AI evolution path: Random Model -\> Pre-Training \(using diverse data\) -\> Fine-Tuning, leading to specialized models \(05:05-05:10\).](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-04-44.png) ![Screenshot at 07:19: Detailed architectural diagram of an AI Agent, showcasing Memory, Prompt processing, Action, Tooling, and the core LLM \(07:37-07:44\).](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-07-19.png) ![Screenshot at 10:28: Slide detailing 'Agent 2 Agent' concepts, showing various 'Agentic Patterns' like Network, Supervisor, Hierarchical, and Custom collaboration models \(10:31-10:42\).](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-10-28.png) ![Screenshot at 11:55: Slide presenting 'AI Agent Examples,' contrasting Vibe Coding \(Software Engineering\) with Legal AI, shown with code snippet and a bar chart indicating market growth \(11:57-12:03\).](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-11-55.png) ![Screenshot at 15:35: Comparison table between the Industrial Revolution \(18th-19th C\) and the AI Revolution \(21st C\) across themes like challenges, real impact, policy changes, and key differentiators \(15:37-15:41\).](https://ss.rapidrecap.app/screens/bxs9tuAMs3E/00-15-35.png)