Skip to content

Generative AI

Generative AI

Generative Artificial Intelligence & LLM

1. Overview of Generative AI, an Artificial Intelligence That Creates and Communicates

    flowchart LR
    A["AI focused on classification/prediction"] -- "Shift to creative content generation and communication" --> B["Generative AI"]
  

Definition: An artificial intelligence technology that goes beyond simply classifying or predicting from learned data, and instead generates new content — text, images, audio, and code — on its own.

Characteristics: (LLM-centered) Built on large language models (LLMs) to generate a wide range of content, including text, code, and images. (Prompt engineering) Output quality varies greatly with input style, making prompt engineering a core skill. (Multimodal) Multimodal capability integrating text, image, voice, and video is rapidly expanding the range of applications.


2. Core Technology and Architecture of Generative AI

A. Evolution Built on the Transformer Architecture

    flowchart TD
    DATA["Large-scale structured/unstructured data"]

    subgraph PRE["Pre-training"]
        TRANS["Transformer Model<br/>(Self-Attention mechanism)"]
    end

    subgraph FINE["Fine-tuning"]
        INST["Instruction Tuning"]
        RLHF["Reinforcement Learning from<br/>Human Feedback (RLHF)"]
    end

    OUT["Generated content<br/>(Text, Image, Code)"]

    DATA --> PRE --> FINE --> OUT

    style PRE fill:#E3F2FD,stroke:#1976D2
    style FINE fill:#FFF3E0,stroke:#F57C00
  
Core TechnologyDescriptionRole
Self-AttentionCaptures the relational meaning between words in a sentenceUnderstands context, resolves long-range dependencies
RLHFTrains using human preference as a reward functionAligns AI responses with human values
Prompt EngineeringOptimizes the instructions given to the AIControls the quality and accuracy of generated output
RAGGenerates answers by referencing an external knowledge basePrevents hallucination, provides up-to-date information

B. Major Generative AI Models and Service Types

    flowchart TD
  GenerativeAI["Generative AI"] --> NLP["Text (NLP)"]
  NLP --> GPT4ClaudeGeminiLlama["GPT-4, Claude, Gemini, Llama"]
  GenerativeAI --> Vision["Image (Vision)"]
  Vision --> DALLEMidjourneyStableDiffusion["DALL-E, Midjourney, Stable Diffusion"]
  GenerativeAI --> Coding["Code (Coding)"]
  Coding --> GitHubCopilotCursor["GitHub Copilot, Cursor"]
  GenerativeAI --> Multimodal["Multimodal"]
  Multimodal --> Node6["Integrated understanding and generation across text, image, and audio"]
  
CategoryKey Technology ModelRepresentative Service
Closed SourceUndisclosed weights and architectureChatGPT, Claude, Gemini
Open SourceCommunity-shared, deployable locallyLlama 3, Mistral, Gemma
SLMSmall models specialized for a specific domainOn-device AI, proprietary enterprise models

3. Expected Benefits and Risk-Response Strategies for Generative AI Adoption

CategoryKey Expected BenefitsRisk and Response Strategy
Productivity innovationAutomates repetitive tasks and content creationPrevent copyright infringement, ensure transparency (labeling AI-generated content)
Personalized experienceDelivers tailored education, healthcare, and service consultationPrevent data privacy and personal-data leaks (apply DLP)
Business innovationEnables new service models and creative problem-solvingVerify hallucination, adopt RAG
Ethical useExpands and democratizes access to knowledgeRemove bias, establish AI ethics guidelines